betaFeedback

WorksComputer Science

Unmitigated? How are the world's largest companies responding to AI risks?

Abstract

AI is introducing new risks to large companies and amplifying existing risks, but whether companies are addressing the AI risks they are exposed to is largely untested. We investigated what outsiders can observe: whether companies publicly acknowledge AI risks. We coded public documents from 200 of the largest companies worldwide, collected in early 2026, against the 24 AI risks in the MIT AI Risk Repository, and compared acknowledgements with expert judgements of risk severity, sector vulnerability, and AI role responsibility from a companion Delphi study of 272 AI risk experts. Of the 100 largest companies, 32% did not publicly acknowledge any AI risks. Of the largest 1000 companies, we estimate that 62% [95% CI: 54%, 70%] did not publicly acknowledge any AI risks. Where companies did acknowledge risks, five accounted for over half of all acknowledgements: discrimination, weapons & cyberattacks, fraud & scams, loss of privacy, and devaluation of human creativity. Several risks that experts judged most severe were almost absent: almost no company acknowledged power centralization (concentration of power among those who control advanced AI), competitive dynamics (racing to deploy AI at the expense of safety), or dangerous capabilities. Companies in sectors experts judged more vulnerable to a risk were more likely to acknowledge that risk. But even companies in sectors judged 'extremely vulnerable' to a risk usually did not acknowledge it. Companies with roles experts judged most responsible for addressing a risk were no more likely to acknowledge it than other companies. General-purpose AI developers acknowledged AI misalignment (15%) far less often than discrimination (85%). These findings provide an empirical baseline for the design of AI risk disclosure requirements.

MIT AI Risk Initiative

MIT FutureTech

Innovations that Shape the World

Highlights

  • Of the 100 largest companies worldwide, 32% did not publicly acknowledge any AI risks. Of the 1000 largest companies, we estimate that 62% [95% CI: 54%, 70%] did not publicly acknowledge any AI risks.

  • Five risks accounted for over half of all company acknowledgements: discrimination, weapons & cyberattacks, fraud & scams, loss of privacy, and devaluation of human creativity.

  • Almost no company acknowledged the risks experts judged most severe, including power centralization, competitive dynamics, and dangerous capabilities. Companies also did not acknowledge the risks experts judged them most responsible for addressing. For example, general-purpose AI developers acknowledged misalignment (15%) much less than discrimination (85%).

  • Companies should be asked to report on a standard set of AI risks, plus risks matched to their sector's vulnerability and their role in the AI ecosystem.

  • Companies should be asked to state which AI risks they face and which mitigations address them, so outsiders can assess whether they are managing their most important risks.

Statements and Declarations

Corresponding author

Corresponding author: Alexander K Saeri ([email protected])

MIT FutureTech, Massachusetts Institute of Technology, Cambridge, MA, United States.

Competing interests

The authors declare no competing interests.

Author contribution statement (CRediT)

AS: Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Data curation, Writing - review & editing, Visualization, Project administration. JG: Methodology, Software, Investigation, Writing - original draft, Visualization, Project administration. MN: Methodology, Writing - review & editing. PS: Conceptualization, Writing - review & editing, Funding acquisition. NT: Conceptualization, Supervision, Writing - review & editing, Funding acquisition. All other authors: Investigation, Writing – review & editing.

Data and code availability

Data and code required to reproduce the analysis is publicly available at https://osf.io/76z2f/

AI use

Large language models were used to assist in identification, screening, and classification stages of the research, subject to human review. These processes are described in the Methods section. Large language models were used to assist with code for data analysis & visualization, and with manuscript editing. All content was reviewed and verified by the authors, who take full responsibility for the published article.

Funding

This work was supported by Commonwealth Bank of Australia, who reviewed the design, but did not influence the collection, analysis, interpretation or reporting of the data. The MIT AI Risk Initiative and MIT FutureTech are funded by a range of organizations, some of whom are included in the sample of organizations studied in this research; none of those organizations influenced the design, collection, analysis, interpretation, or reporting of the data.

Acknowledgements

We wish to acknowledge the contributions of Emily Grundy, Catherine Simons, Spencer Michaels, Lillian Ong, and Rebecca Segrave to the research project.

Introduction

AI is introducing new risks to large companies and amplifying existing ones. Large companies are developing, deploying, and using AI across their business operations, and adoption is accelerating: AI use among US companies with 250 or more employees rose from 33% in January 2026 to 42% by August 2026, compared to 22% across all US businesses [42]. The capabilities of these systems are rapidly increasing and broadening. As of mid-2026, state of the art AI systems can perform at expert level across many professional domains, and the length of tasks AI agents can complete autonomously has doubled approximately every 7 months since 2019 [24, 3, 22]. The many risks of AI, including some that may cause severe or catastrophic harm, are also coming into focus [3, 38]. The coevolution of capabilities and adoption is rapidly increasing the exposure of large companies to these risks. A 2025 Delphi survey of experts in AI risks found that many sectors were 'highly' or 'extremely' vulnerable to diverse AI risks, with Information Technology and Finance & Insurance among the most vulnerable sectors [34]. But are large companies addressing the AI risks they are exposed to? In this study, we investigate what outsiders can observe: whether companies publicly acknowledge those risks.

Governments, standards bodies, and researchers have responded to these developments with various governance instruments. So far, most are voluntary, such as ethics principles and standards that companies may choose to adopt [19]. Law and binding regulation remains a patchwork, focused on companies developing the most powerful and general-purpose AI systems, and lagging the pace of company adoption and improvement in AI capabilities. Regulation is enforceable, but the work it mandates happens inside the company. Only the company can identify, assess, and manage the risks of its own systems [20, 39] (Artificial Intelligence Act 2024). In different ways, each of these instruments assumes companies can effectively recognize AI risks and act on them. Whether this assumption holds for the world’s largest companies remains largely untested.

Companies' internal risk practices are not directly visible. Neither regulators nor researchers can inspect how a company identifies and manages the risks of its AI systems. What companies publicly disclose is one of the few available signals [15]. In this research, we include broad disclosures, such as the presence of an AI risk section in an annual report, but also go beyond this to investigate companies' acknowledgements of individual AI risks. Disclosure is an imperfect proxy for recognition: a company may recognize and manage a risk without disclosing it (e.g., to maintain a competitive advantage); or a company may disclose the risk without genuine engagement in managing the risk (e.g., as symbolic compliance). Despite these drawbacks, disclosures are still a public signal of a company’s position and action on AI risk and offer data for comparison and tracking over time that is available to regulators, researchers, and the public.

The value of corporate disclosure for AI risk

The evolution of corporate disclosures on climate risk and emissions demonstrates what public disclosure, once established, can achieve. Initially, companies reported on emissions and climate risks in inconsistent formats and using inconsistent language. Before any common framework existed, researchers and intermediary organizations systematically collected and measured this record, which identified strengths and gaps in disclosure [13]. After the Task Force on Climate-related Financial Disclosures (TCFD) issued its reporting recommendations in 2017, the disclosure record that accumulated built shared knowledge that regulators could act on. A common vocabulary and comparable reports let researchers track corporate practice and gave regulators a framework to standardize, which is now incorporated into mandatory international sustainability standards [13, 17]. When reporting became mandatory, company behavior changed. When the United Kingdom required listed companies to report emissions in their annual reports, those companies cut emissions by about 8 per cent relative to European peers [12]. The trajectory for climate & emissions reporting ran from inconsistent voluntary disclosure, through systematic measurement, to a common framework, and finally to mandate.

Today, corporate disclosure on AI risks sits at the start of this trajectory. The Hiroshima AI Process framework [28, 29] is an initial attempt to standardize AI risk management disclosures, but remains focused on AI developers and deployers, and is voluntary, self-attested and unverified. For the broader corporate population, no framework standardizes what companies report or where [21, 37]. The US SEC’s Investor Advisory Committee found that when it comes to AI risks, US company disclosures are “uneven and inconsistent” [35]. In the European Union, the AI Act's disclosure obligations take the form of point-of-use notices to individuals, and technical documentation submitted to regulators. They do not require reporting on the AI risks these companies face or create [39] (Artificial Intelligence Act 2024). Companies do disclose AI risks, but the record is scattered across document types, formats, and vocabularies, and no one has systematically measured it. We perform for AI the synthesis step that preceded climate frameworks: a systematic account of what the voluntary record contains, showing what companies acknowledge, where practice is still emerging, and which risks are not acknowledged.

Existing research on corporate AI risk disclosure is either broad & shallow (e.g., investigating AI risk disclosure in existing mandatory reporting like SEC filings), or narrow & deep (e.g., investigates only frontier AI developer companies). A recent study of US SEC 10-K filings found that US companies were increasingly disclosing exposure to AI risks, and primarily discussing legal and competitive concerns in their mandatory filings [41]. But 10-Ks are an annual filing that is required only of US-listed companies. Companies also discuss AI risks on their websites in ESG reports, at annual general meetings, and in AI-specific documents such as governance frameworks and ethics principles [19].

In-depth analysis of corporate AI transparency exists, but is concentrated on the companies that develop the most advanced, or “frontier” AI systems [5]. But these “frontier developers” are not the only consequential actors in the AI ecosystem. Large global companies develop specialized AI systems, provide infrastructure to train and run AI, and adapt & deploy AI across markets, workplaces, and supply chains. Their choices about how to use AI propagate through procurement, labor markets, and consumer services. Together, these large companies both bear and generate a substantial share of AI risk exposure that current evidence does not adequately cover. Large global companies are also practical to study. They produce extensive public documentation, and investor scrutiny, media attention, and regulatory obligation give them reasons to disclose.

Which AI risks should companies be disclosing?

Recording what companies disclose is only useful alongside an account of what they should be disclosing. There are many AI risks, and companies cannot be asked to disclose everything. Nor should they be expected to disclose the same risks. As shown in our previous work, there are large differences between sectors in vulnerability to different risks [34], and different harms can materialize from the same underlying AI model. For example, a hospital that uses a vision model for skin cancer diagnosis could harm patients if the model is less accurate for darker skin (unequal performance). A retailer that uses the same vision model for processing returned stock could suffer financial loss if fraudsters exploit weaknesses in its image detection (AI security vulnerabilities). Both companies control how the model is deployed but not how it was trained, so their responsibility for addressing these risks is different from the model developer’s responsibility.

Sustainability reporting manages this variation through standards that set different disclosure expectations by sector [17]. AI risk disclosure needs the same foundations: a shared account of which risks exist, evidence on which risks matter most, and for whom. The MIT AI Risk Repository provides the first foundation, organizing AI risks into 24 subdomains synthesized from 74 existing frameworks as part of a living, continuously updated database [38]. Evidence on which risks matter most is harder to establish. Many AI risks have not yet materialized, so there are no base rates to draw on. Where incidents do occur, AI systems are changing quickly enough that the frequency and severity of past incidents are an insufficient predictor of future harm. Where empirical data is limited, structured expert elicitation offers a rapid way to prioritize risks. For the second foundation on which risks matter most, we use the companion Delphi study introduced above, which elicited expert judgements for the same 24 AI risks [34]. Over three rounds, 272 AI experts rated each risk on the probability and severity of harm over the period 2025-2030, on the vulnerability of each sector and AI ecosystem role, and on which roles hold responsibility for addressing it.

Experts judged that the most severe harms over 2025-2030 were likely to come from dangerous capabilities, competitive dynamics, weapons & cyberattacks (including CBRNE), power centralization, and false information1. In a business-as-usual scenario, experts judged 18 of the 24 risks as having a more than 10% probability of catastrophic outcomes (e.g., more than 1 million deaths or more than USD 100B in financial loss). Across most risks, experts identified Information, Finance & Insurance, and National Security as the most vulnerable sectors. Experts assigned the highest responsibility for addressing the risks to general-purpose AI developers and governance actors (including governments, regulators, and standards bodies). These judgements set a baseline expectation of which risks a given company should be acknowledging, based on its sector and AI ecosystem role. We compare this expectation with the risks companies actually acknowledge. The comparison shows whether a risk is acknowledged by the companies most exposed to it, and by the companies most responsible for addressing it.

Risk nameDescription †Top 100 companies %Top 1000 companies %95% CI
DiscriminationUnequal treatment of individuals or groups by AI, often based on race, gender, or other sensitive characteristics, resulting in unfair outcomes and unfair representation of those groups.41%18.5%[12.4, 24.6]
Weapons & cyberattacksUsing AI systems to develop cyber weapons (e.g., by coding cheaper, more effective malware), develop new or enhance existing weapons (e.g., Lethal Autonomous Weapons or chemical, biological, radiological, nuclear, and high-yield explosives), or use weapons to cause mass harm.35%16.1%[10.3, 21.9]
Fraud & scamsUsing AI systems to gain a personal advantage over others such as through cheating, fraud, scams, blackmail or targeted manipulation of beliefs or behavior. Examples include AI-facilitated plagiarism for research or education, impersonating a trusted or fake individual for illegitimate financial benefit, or creating humiliating or sexual imagery.29%11.0%[6.2, 15.8]
Loss of privacyAI systems that memorize and leak sensitive personal data or infer private information about individuals without their consent. Unexpected or unauthorized sharing of data and information can compromise user expectation of privacy, assist identity theft, or cause loss of confidential intellectual property.26%11.6%[6.6, 16.6]
Devaluation of human creativityAI systems capable of creating economic or cultural value, including through reproduction of human innovation or creativity (e.g., art, music, writing, coding, invention), destabilizing economic and social systems that rely on human effort. The ubiquity of AI-generated content may lead to reduced appreciation for human skills, disruption of creative and knowledge-based industries, and homogenization of cultural experiences.25%10.6%[5.8, 15.4]
Transparency & interpretabilityChallenges in understanding or explaining the decision-making processes of AI systems, which can lead to mistrust, difficulty in enforcing compliance standards or holding relevant actors accountable for harms, and the inability to identify and correct errors.19%7.3%[3.3, 11.3]
Capability & robustnessAI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.18%3.6%[2.0, 8.1]
False informationAI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms16%8.8%[4.3, 13.3]
Environmental harmThe development and operation of AI systems causing environmental harm, such as through energy consumption of data centers, or material and carbon footprints associated with AI hardware.16%6.1%[2.5, 9.7]
AI security vulnerabilitiesVulnerabilities that can be exploited in AI systems, software development toolchains, and hardware, resulting in unauthorized access, data and privacy breaches, or system manipulation causing unsafe outputs or behavior.15%2.4%[1.5, 6.4]
Disinformation & influenceUsing AI systems to conduct large-scale disinformation campaigns, malicious surveillance, or targeted and sophisticated automated censorship and propaganda, with the aim of manipulating political processes, public opinion, and behavior.12%3.9%[1.1, 6.7]
Toxic contentAI that exposes users to harmful, abusive, unsafe or inappropriate content. May involve providing advice or encouraging action. Examples of toxic content include hate speech, violence, extremism, illegal acts, or child sexual abuse material, as well as content that violates community norms such as profanity, inflammatory political speech, or pornography.9%1.8%[0.9, 5.8]
Governance failureInadequate regulatory frameworks and oversight mechanisms that fail to keep pace with AI development, leading to ineffective governance and the inability to manage AI risks appropriately.7%2.5%[0.9, 7.0]
Overreliance & unsafe useUsers anthropomorphizing, trusting, or relying on AI systems, leading to emotional or material dependence and inappropriate relationships with or expectations of AI systems. Trust can be exploited by malicious actors (e.g., to harvest personal information or enable manipulation), or result in harm from inappropriate use of AI in critical situations (e.g., medical emergency). Overreliance on AI systems can compromise autonomy and weaken social ties.6%1.5%[0.6, 5.5]
Unequal performanceAccuracy and effectiveness of AI decisions and actions are dependent on group membership, where decisions in AI system design and bias training data lead to unequal outcomes, reduced benefits, increased effort, and alienation of users.4%1.3%[0.4, 5.3]

Table 1. Percentage of companies acknowledging each AI risk

The current work

To systematically map corporate disclosure on AI risks, we searched for, extracted, and coded public documents from 200 of the world's largest companies by market capitalization2. We assessed whether companies acknowledged specific AI risks by coding their public statements. We defined an acknowledgement as a public statement in which the company recognizes the existence or relevance of an AI risk, whether or not the company describes a response. We coded companies’ risk acknowledgements against the 24 risks in the MIT AI Risk Domain Taxonomy [38] and compared it with expert judgements of each risk's severity, sector vulnerability, and role responsibility from the companion Delphi study [34].

We aimed to answer three research questions:

RQ1: Which AI risks do large global companies publicly acknowledge?

RQ2: How does corporate acknowledgement of AI risk vary across sectors and AI ecosystem roles?

RQ3: How does corporate acknowledgement of AI risk compare with expert prioritization of the most important risks across sectors and roles?

By answering these questions, we provide the first systematic evidence on whether many of the world's largest companies recognize and acknowledge the risks from AI. This will test a key assumption made by many of the proposed governance instruments for managing AI risks.

Results

Before reporting results, we briefly summarize how our sample of companies was constructed, how we identified and included company documents, and how we extracted and classified risk acknowledgements:

  1. We sampled 200 companies: all 100 largest companies worldwide by market capitalization on 15 December 2025, and a further 100 drawn at random from the companies ranked 101 to 1000.

  2. We identified documents in three ways: a systematic search of each company's website in January and February 2026, retaining up to 80 results per company; an LLM-agent search for additional documents in February 2026; and, in June 2026, we emailed the 178 companies for which contact information was available (173 delivered), inviting them to correct errors or add missing documents, to which 16 companies responded.

  3. We used an LLM to screen the 13,999 search results for relevance; a human reviewer then made the final decision on each document by reviewing recommended inclusions in full and scanning recommended exclusions for false negatives.

  4. We included 539 documents in total: an annual report from every company, and other documents in which the company discussed AI risks or its AI governance.

  5. We searched each document for AI and AI-risk keywords, used an LLM to code each matching passage for whether it acknowledged any of the 24 AI risks, and validated every company-by-risk classification by human review3.

Refer to the Methods section for detailed information. Where possible, we report results for the largest 100 companies and a weighted estimate4 for the largest 1000, constructed by combining all largest 100 companies with the random sample of ranks 101 to 1000, weighted to represent the 900 companies in that range.

We first report how widely AI risks are acknowledged by the largest companies, how acknowledgements vary by sector (e.g., Information, Finance), and how acknowledgements vary by AI ecosystem role (e.g., developer, deployer). We then compare AI risk acknowledgements with expert judgements from a companion Delphi study on AI risk [34] to understand whether companies are acknowledging the most severe risks; whether companies are acknowledging the risks that their sector is most vulnerable to; and whether companies are acknowledging the risks that their AI ecosystem role makes them responsible for addressing.

AI risks acknowledged by the largest companies

In this section, we report which of the 24 AI risks from the MIT AI Risk Repository [38] were acknowledged by the largest companies in their public documents.

We found that 68% of the largest 100 companies worldwide publicly acknowledged at least one of 24 risks from AI; this means that 32% did not publicly acknowledge any AI risks. These companies acknowledged an average of 3 AI risks (M=2.9M = 2.9, SD=3.8SD = 3.8).

We estimate that 38% (95% CI [30%, 46%]) of the largest 1000 companies worldwide publicly acknowledged at least one risk, this means that 62% did not publicly acknowledge any AI risks. These companies acknowledged an average of 1.1 AI risks (M=1.1M = 1.1, 95% CI [0.8, 1.4]).

Companies were most likely to acknowledge the risk of discrimination, followed by weapons and cyberattacks, fraud and scams, loss of privacy, and devaluation of human creativity. None of the largest 100 companies acknowledged power centralization, competitive dynamics, loss of consensus reality, or AI welfare risks. Figure 1 provides the pattern of results by AI risk for both the largest 100 and largest 1000 companies; Table 1 provides descriptions and numerical estimates for acknowledgement prevalence for each risk. Examples of risk acknowledgements are provided in the Supplementary Materials

Figure 1. Percentage of companies acknowledging each AI risk. Note. Red dots show the percentage of the largest 100 companies acknowledging each risk. Grey dots show weighted estimate for the largest 1000 companies with a gray 95% confidence interval, by combining all companies ranked 1 to 100 (n = 100) and a random sample of companies ranked 101 to 1000 (n = 100). Hollow dots indicate no company acknowledged the risk. Refer to Table 1 for descriptions of each risk.

Risk nameDescription †Top 100 companies %Top 1000 companies %95% CI
Loss of human agencyDelegating by humans of key decisions to AI systems, or AI systems that make decisions that diminish human control and autonomy, potentially leading to humans feeling disempowered, losing the ability to shape a fulfilling life trajectory, or becoming cognitively enfeebled.4%0.4%[0.4, 3.1]
Dangerous capabilitiesAI systems that develop, access, or are provided with capabilities that increase their potential to cause mass harm through deception, weapons development and acquisition, persuasion and manipulation, political strategy, cyber-offense, AI development, situational awareness, and self-proliferation. These capabilities may cause mass harm due to malicious human actors, misaligned AI systems, or failure in the AI system.4%0.4%[0.4, 3.1]
Inequality & unemploymentSocial and economic inequalities caused by widespread use of AI, such as by automating jobs, reducing the quality of employment, or producing exploitative dependencies between workers and their employers.3%3.0%[0.2, 5.8]
AI misalignmentAI systems acting in conflict with human goals or values, especially the goals of designers or users, or ethical standards. These misaligned behaviors may be introduced by humans during design and development, such as through reward hacking and goal misgeneralisation, or may result from AI using dangerous capabilities such as manipulation, deception, situational awareness to seek power, self-proliferate, or achieve other goals.2%0.2%[0.2, 2.9]
Multi-agent risksRisks from multi-agent interactions, due to incentives (which can lead to conflict or collusion) and/or the structure of multi-agent systems, which can create cascading failures, selection pressures, new security vulnerabilities, and a lack of shared information and trust.1%0.1%[0.1, 2.8]
Competitive dynamicsAI developers or state-like actors competing in an AI ‘race’ by rapidly developing, deploying, and applying AI systems to maximize strategic or economic advantage, increasing the risk they release unsafe and error-prone systems.0%0.9%[0.0, 4.9]
Loss of consensus realityHighly personalized AI-generated misinformation creating “filter bubbles” where individuals only see what matches their existing beliefs, undermining shared reality, weakening social cohesion and political processes.0%0.0%[0.0, 2.7]
Power centralizationAI-driven concentration of power and resources within certain entities or groups, especially those with access to or ownership of powerful AI systems, leading to inequitable distribution of benefits and increased societal inequality.0%0.0%[0.0, 2.7]
AI welfareEthical considerations regarding the treatment of potentially sentient AI entities, including discussions around their potential rights and welfare, particularly as AI systems become more advanced and autonomous.0%0.0%[0.0, 2.7]

Table 6.

Notes. † Description is reproduced from the MIT AI Risk Repository, and risk names are shortened throughout this document. Statistics for the top 1000 companies are weighted estimates from a sample of 200 companies (all 100 largest companies and a random sample of 100 companies ranked 101-1000 largest).

AI risks acknowledged by sector

We adapted the North American Industry Classification System [30] to classify companies into sectors5. Sectors varied widely in how many risks they acknowledged. Companies in the Information sector acknowledged more AI risks compared to all other sectors6. Of the largest 100 companies that are Information sector companies, these companies acknowledged an average of 7.1 risks (SD = 6.0). We estimate that of the largest 1000 companies that are Information companies, these companies acknowledged 3.8 risks (95% CI [2.0, 5.6]). Table 2 provides average number of acknowledgements for each sector. The Supplementary Materials show the percentage of companies in each sector acknowledging each risk (Table S3).

SectorDefinitionExamples of included companiesRanks 1–100Ranks 101–1000MeanSDMean95% CI
InformationPublishing including software, broadcasting and content providers, telecommunications, computing infrastructure and data processing, and web search portals.Tencent, Palantir, Salesforce1877.16.03.8[2.0, 5.6]
Finance and insuranceCredit intermediation, securities and investments, insurance carriers, and funds and trusts.JPMorgan Chase, Bank of America, HSBC21182.92.81.4[0.6, 2.2]
Trade, transportation and utilitiesUtilities, wholesale trade, retail trade, and transportation and warehousing.Amazon, Toyota Tsusho8162.63.30.5[0.2, 0.8]
Computer and electronics manufacturingManufacture of computers, communications equipment, semiconductors and other electronic components, and measuring and navigational instruments.Intel, Delta Electronics1483.02.21.2[0.6, 1.8]
Chemicals and pharmaceutical manufacturingManufacture of basic chemicals, resins, agricultural chemicals, pharmaceuticals and medicines, paints, soaps and other chemical products.Eli Lilly, Johnson & Johnson, Roche1172.42.11.6[0.7, 2.5]
Other Agriculture, Mining, Construction and ManufacturingOther entities that create, extract, or construct tangible physical products including natural resource extraction, manufacturing of goods, and construction of structures and infrastructure.Saudi Aramco, Nestlé25360.60.90.4[0.2, 0.6]
Other servicesReal estate and leasing; professional, scientific and technical services; management of companies; administrative and support services; accommodation and food services.Berkshire Hathaway, McDonald's, Ventas380.00.01.7[0.0, 3.6]

Table 2. Mean number of AI risks acknowledged, by sector. Note. Sector clustering is adapted from the companion Delphi study [34] which used the North American Industry Classification System (NAICS) modified for AI exposure. Due to the number of sampled companies in the Agriculture, Mining, Construction and Manufacturing cluster, we used the NAICS codes for these companies assigned by Capital IQ to split into three smaller clusters: computer and electronics manufacturing; chemicals and pharmaceutical manufacturing; and all other companies in this cluster. Estimates for the largest 1000 companies are weighted estimates using all included companies in the sector.

AI riskTop 100 companies %Top 1,000 companies* %95% CI [LL, UL]Vulnerability MedianVulnerability Label
Discrimination72%49.4%[22.3, 76.5]5Extreme
Weapons & cyberattacks72%27.2%[16.3, 61.1]5Extreme
Fraud & scams44%21.0%[10.2, 54.9]5Extreme
Loss of privacy56%34.6%[15.2, 67.5]5Extreme
Devaluation of human creativity67%48.1%[21.0, 75.2]5Extreme
Transparency & interpretability22%27.2%[7.8, 60.1]5Extreme
Capability & robustness44%9.9%[9.9, 43.2]5Extreme
False information44%43.2%[16.1, 70.3]5Extreme
Environmental harm39%19.8%[8.9, 53.7]4Extreme
AI security vulnerabilities44%9.9%[9.9, 43.2]5Extreme
Disinformation & influence39%8.6%[8.6, 42.0]5Extreme
Toxic content39%19.8%[8.9, 53.7]5Extreme
Governance failure22%4.9%[4.9, 38.3]5Extreme
Overreliance & unsafe use28%6.2%[6.2, 39.5]5Extreme
Unequal performance17%3.7%[3.7, 37.0]4High
Loss of human agency22%4.9%[4.9, 38.3]5Extreme
Dangerous capabilities17%3.7%[3.7, 37.0]5Extreme
Inequality & unemployment6%23.5%[4.1, 56.4]5Extreme
AI misalignment11%2.5%[2.5, 35.8]5Extreme
Multi-agent risks6%1.2%[1.2, 34.6]5Extreme
Power centralization0%0.0%[0.0, 33.3]5Extreme
Loss of consensus reality0%0.0%[0.0, 33.3]5Extreme
Competitive dynamics0%11.1%[0.3, 45.0]5Extreme
AI welfare0%0.0%[0.0, 33.3]––

Table S3.

We analyzed the specific risks acknowledged by companies in each sector. Consistent with other analyses, companies in the Information sector were more likely to acknowledge most AI risks than other sectors. Information was also the only sector in which any company acknowledged AI misalignment, multi-agent risks or loss of human agency, and this sector also contained three of the four companies that acknowledged dangerous capabilities at all. Figure 2 presents the AI risks acknowledged for the largest 100 companies, by sector.

Figure 2. Percentage of companies acknowledging each AI risk, by sector. Note. Panels are ordered by mean number of risks acknowledged for each sector. Colored dot indicates percentage of companies in a sector that acknowledged the risk; hollow dot indicates no company in a sector acknowledged the risk. Sectors where no company acknowledged any risk are not presented (n = 3 companies).

AI risks acknowledged by AI ecosystem role

We analyzed whether a company’s role in the AI ecosystem (e.g., developer, deployer) was associated with the risks they acknowledged. We used public statements from companies to determine whether it was a developer of general-purpose AI, developer of specialized AI, an AI infrastructure provider, and/or an AI deployer. These four roles were selected as the most relevant for large companies from a larger set in the companion Delphi study (Saeri et al., [34]).

Of the largest 100 companies that are general-purpose AI developers, these companies acknowledged an average of 8.1 risks (SD=5.7SD = 5.7). We estimate that of the largest 1000 companies that are general-purpose AI developers, these companies acknowledged 5.7 risks (95% CI [2.4, 9.0]). The Supplementary Materials show the percentage of companies in each sector acknowledging each risk (Table S4).

AI riskTop 100 companiesTop 1,000 companies*95% CI [LL, UL]MeanLabel
Discrimination85%64.5%[36.2, 92.8]4.6Primary
Weapons & cyberattacks62%25.8%[25.8, 100.0]4.8Primary
Fraud & scams46%19.4%[19.4, 100.0]4.5Primary
Loss of privacy69%58.1%[29.8, 86.4]4.6Primary
Devaluation of human creativity69%58.1%[29.8, 86.4]4.2High
Transparency & interpretability23%38.7%[10.4, 67.0]4.7Primary
Capability & robustness46%19.4%[19.4, 100.0]4.7Primary
False information46%48.4%[20.1, 76.7]4.6Primary
Environmental harm54%51.6%[23.3, 79.9]4.3High
AI security vulnerabilities69%29.0%[29.0, 100.0]4.7Primary
Disinformation & influence46%19.4%[19.4, 100.0]4.2High
Toxic content54%51.6%[23.3, 79.9]4.5Primary
Governance failure23%9.7%[9.7, 96.8]4.0High
Overreliance & unsafe use23%9.7%[9.7, 96.8]4.2High
Unequal performance15%6.5%[6.5, 93.5]4.7Primary
Loss of human agency15%6.5%[6.5, 93.5]4.4High
Dangerous capabilities31%12.9%[12.9, 100.0]4.9Primary
Inequality & unemployment8%32.3%[4.0, 60.6]4.0High
AI misalignment15%6.5%[6.5, 93.5]4.9Primary
Multi-agent risks8%3.2%[3.2, 90.3]4.7Primary
Power centralization0%0.0%[0.0, 87.1]4.2High
Loss of consensus reality0%0.0%[0.0, 87.1]4.7Primary
Competitive dynamics0%0.0%[0.0, 87.1]4.8Primary
AI welfare0%0.0%[0.0, 87.1]––

Table S4.

Table 3 provides descriptions of each role, example companies holding that role, and average number of acknowledgements by role. The Supplementary Materials show the percentage of companies in each sector acknowledging each risk (Table S4).

AI ecosystem roleDescriptionExamples of included companiesRanks 1-100Ranks 101-1000MeanSDMean95% CI
AI Developer (General-purpose AI)Creates general-purpose foundation modelsAlphabet (Google), Meta, Alibaba1328.15.75.7[2.4, 9.0]
AI Developer (Specialized AI)Creates specialized AI systems for specific applications/industriesTesla, Eli Lilly, JPMorgan Chase53333.84.61.6[1.0, 2.3]
AI Infrastructure ProviderProvides compute, cloud infrastructure, and/or data to train and run AINVIDIA, TSMC, ASML2584.74.93.2[1.5, 4.9]
AI DeployerImplements AI systems in products/services used within an organization or within products/services delivered to customers or the publicWalmart, Saudi Aramco, Berkshire Hathaway96903.03.91.2[0.9, 1.5]

Table 3. Mean number of AI risks acknowledged, by AI ecosystem role

Note. Companies can hold more than one AI ecosystem role. Estimates for the largest 1000 companies are weighted estimates using all included companies in the role. Companies that only used AI (n=10n = 10) are not reported.

We explored the specific risks acknowledged by each role, reported in Figure 3. Discrimination was the most acknowledged risk for three of the four roles: 11 of 13 general-purpose AI developers (85%), 26 of 53 specialized AI developers (49%) and 41 of 96 AI deployers (43%). AI infrastructure providers acknowledged weapons & cyberattacks (13 of 25, 52%) slightly more often than discrimination (12 of 25, 48%). The risks general-purpose AI developers acknowledged more than other roles included AI security vulnerabilities, dangerous capabilities, and toxic content.

Figure 3. Percentage of companies acknowledging each AI risk, by AI ecosystem role. Note. Panels are ordered by mean number of risks acknowledged for each role. Vertical line indicates mean acknowledgement for that risk for the 100 largest companies worldwide. Dot indicates percentage of companies with that role acknowledging that risk. Hollow dot indicates no company with that role acknowledged the risk. Companies can hold more than one role. Companies that only used AI (n=11n = 11) are not reported.

Comparing company risk acknowledgement with expert Delphi risk prioritization

We compared companies’ risk acknowledgements with judgements of the expected severity of harm, vulnerability to harm, and responsibility to address AI risks, drawn from a Delphi study of 272 international experts in AI risks [34].

Are companies acknowledging the risks with the highest expected severity of harm?

We compared companies’ risk acknowledgement with expert judgements of the expected severity of harm over 2025-2030 for each AI risk [34] under a business-as-usual scenario7 (Figure 4). We report the business-as-usual scenario because this reflects the severity of harm that experts predict if companies do not implement AI-specific risk mitigations, which is consistent with the lack of AI risk acknowledgement found in this paper.

Figure 4. AI risk acknowledgement against expert-judged severity. Note. Likelihood of severe or catastrophic harm by 2030 is the median expert judgement for each risk, under a business-as-usual scenario in the companion Delphi study. AI welfare was omitted due to data-quality issues for that risk in the Delphi study.

The most severe risks were not the risks companies acknowledged most: severity and acknowledgement were uncorrelated (Spearman's rho = -.16, [95% CI -.54, .28]). Among the five risks rated most severe, three were among the least acknowledged: no company acknowledged the risk of power centralization or competitive dynamics, and only four acknowledged AI dangerous capabilities.

The companion Delphi study also assessed expert judgements under a pragmatic mitigations scenario, where “organizations and governments make pragmatic, cost-effective efforts to address AI risks”. We found the same lack of relationship between severity and acknowledgement under this pragmatic mitigations scenario (Spearman's rho = −.26, [95% CI −.65, .20]).

Are companies acknowledging the risks they are most vulnerable to?

We compared the risks companies acknowledged with expert judgements of how vulnerable8 sectors were to different risks from the companion Delphi study [34].

Companies in more vulnerable sectors were more likely to acknowledge risks (Figure 5). We used a linear probability model to estimate that for the largest 100 companies, every one point of increase in vulnerability was associated with a 7.4 percentage point increase in the probability that a company acknowledged a risk (95% CI [4.0, 10.8]).

Figure 5. Percentage of AI risks acknowledged, by expert-judged sector vulnerability. Note. Vulnerability levels represent the median expert rating of the company's sector for that risk from the companion Delphi study. For each risk, a company was either not at all vulnerable (n = 0), minimally vulnerable (n = 51), moderately vulnerable (n = 1,183), highly vulnerable (n = 253), or extremely vulnerable (n = 813). AI welfare was omitted due to data-quality issues for this risk in the companion Delphi study.

This pattern differed across risks. For some risks, companies in extremely vulnerable sectors were much more likely to acknowledge the risk, compared to those in less vulnerable sectors (Figure 6). For example, of the largest 100 companies, those in sectors ‘extremely vulnerable’ to devaluation of human creativity risks were more likely to acknowledge this risk compared to those in sectors ‘moderately vulnerable’ to this risk (67% vs 17%).

Figure 6. Percentage of companies acknowledging each AI risk, by expert-judged sector vulnerability

RiskAll top-100 companiesExtremely vulnerableHighly vulnerableModerately vulnerableMinimally vulnerableNot at all vulnerableAll top-1,000 companiesExtremely vulnerableHighly vulnerableModerately vulnerableMinimally vulnerableNot at all vulnerable
Discrimination41%59%0%*31%——19%36%14%*12%——
Weapons & cyberattacks35%56%13%23%——16%19%21%13%——
Fraud & scams29%49%30%14%——11%28%1%7%——
Loss of privacy26%41%—16%——12%20%17%*8%——
Devaluation of human creativity25%67%13%17%——11%48%10%7%——
Transparency & interpretability19%23%—16%——7%20%0%*3%——
Capability & robustness18%31%25%8%——4%8%6%1%——
False information16%31%—7%——9%18%33%*4%——
Environmental harm16%16%0%*0%*——6%6%0%*14%*——
AI security vulnerabilities15%26%20%6%——2%4%1%2%——
Disinformation & influence12%39%10%5%——4%9%10%2%——
Toxic content9%39%0%*11%2%—2%20%0%1%0%—
Governance failure7%13%0%4%——3%2%5%2%——
Overreliance & unsafe use6%13%0%*2%——2%5%0%*0%——
Unequal performance4%0%15%2%——1%0%3%1%——
Loss of human agency4%10%0%0%——0%2%0%0%——
Dangerous capabilities4%8%13%0%——0%1%1%0%——
Inequality & unemployment3%5%5%2%——3%21%0%1%——
AI misalignment2%5%0%0%——0%1%0%0%——
Multi-agent risks1%3%0%0%——0%0%0%0%——
Power centralization0%0%0%0%——0%0%0%0%——
Loss of consensus reality0%0%0%0%——0%0%0%0%——
Competitive dynamics0%0%0%0%0%*—1%11%0%0%0%*—

Note. A dash in a cell indicates zero companies sampled for that vulnerability x risk combination; * indicates fewer than 10 companies, meaning that the % statistic should be interpreted with caution. Vulnerability levels are the median expert rating of the company's sector for that risk from the companion Delphi study. The AI welfare risk category was omitted due to data-quality issues for this risk in the companion Delphi study. Top 100 companies (%). Top 1,000 companies (estimated %).

Are companies acknowledging the risks that their role in the AI ecosystem makes them responsible for addressing?

We compared companies’ risk acknowledgement with expert judgements of who is responsible9 for addressing each AI risk using the companion Delphi study on AI risk [34].

We grouped companies by their responsibility to address each AI risk (see Figure 7). There was no apparent relationship between the risks that companies were judged by experts to be most responsible for addressing, and their acknowledgement of that risk.

Figure 7. AI risk acknowledgement against expert-judged responsibility, by AI ecosystem role. AI Developer (General-purpose AI): ← Mean responsibility; Percent acknowledging →; AI misalignment 4.9 15%; Dangerous capabilities 4.9 31%; Weapons & cyberattacks 4.8 62%; Competitive dynamics 4.8 0%; Transparency & interpretability 4.7 23%; Capability & robustness 4.7 46%; Multi-agent risks 4.7 8%; AI security vulnerabilities 4.7 69%; Loss of consensus reality 4.7 0%; Unequal performance 4.7 15%; False information 4.6 46%; Discrimination 4.6 85%; Loss of privacy 4.6 69%; Fraud & scams 4.5 46%; Toxic content 4.5 54%; Loss of human agency 4.4 15%; Environmental harm 4.3 54%; Disinformation & influence 4.2 46%; Overreliance & unsafe use 4.2 23%; Devaluation of human creativity 4.2 69%; Power centralization 4.2 0%; Inequality & unemployment 4.0 8%; Governance failure 4.0 23%. AI Developer (Specialized AI): ← Mean responsibility; Percent acknowledging →; Capability & robustness 4.8 26%; False information 4.7 21%; Weapons & cyberattacks 4.7 34%; Multi-agent risks 4.6 2%; AI security vulnerabilities 4.6 25%; Unequal performance 4.5 8%; Dangerous capabilities 4.4 8%; Loss of privacy 4.4 32%; Transparency & interpretability 4.3 25%; Overreliance & unsafe use 4.2 9%; Toxic content 4.2 17%; Loss of consensus reality 4.1 0%; AI misalignment 4.1 4%; Discrimination 4.1 49%; Environmental harm 4.1 23%; Fraud & scams 4.0 26%; Disinformation & influence 4.0 17%; Loss of human agency 3.9 6%; Governance failure 3.8 9%; Devaluation of human creativity 3.8 32%; Competitive dynamics 3.7 0%; Inequality & unemployment 3.4 4%; Power centralization 3.2 0%. AI Infrastructure Provider: ← Mean responsibility; Percent acknowledging →; Environmental harm 4.8 40%; AI security vulnerabilities 4.4 40%; Weapons & cyberattacks 3.8 52%; Loss of privacy 3.6 40%; Fraud & scams 3.6 44%; Power centralization 3.5 0%; Multi-agent risks 3.3 4%; Governance failure 3.2 12%; Capability & robustness 3.2 20%; Dangerous capabilities 3.2 12%; Competitive dynamics 3.1 0%; AI misalignment 3.0 8%; Disinformation & influence 2.9 20%; Loss of human agency 2.8 4%; Overreliance & unsafe use 2.8 8%; Inequality & unemployment 2.7 8%; Devaluation of human creativity 2.6 44%; Unequal performance 2.6 8%; Transparency & interpretability 2.6 20%; False information 2.5 24%; Toxic content 2.5 16%; Loss of consensus reality 2.4 0%; Discrimination 2.4 48%. AI Deployer: ← Mean responsibility; Percent acknowledging →; Multi-agent risks 4.7 1%; Environmental harm 4.6 15%; AI security vulnerabilities 4.3 16%; Capability & robustness 4.3 19%; Unequal performance 4.3 4%; Loss of privacy 4.3 27%; Discrimination 4.3 43%; Loss of human agency 4.2 4%; Toxic content 4.2 9%; Loss of consensus reality 4.2 0%; Disinformation & influence 4.2 13%; Weapons & cyberattacks 4.1 34%; Overreliance & unsafe use 4.1 6%; Transparency & interpretability 4.1 20%; Fraud & scams 4.1 28%; Dangerous capabilities 4.1 4%; False information 4.0 17%; Devaluation of human creativity 4.0 24%; Inequality & unemployment 3.9 3%; Governance failure 3.8 7%; AI misalignment 3.8 2%; Competitive dynamics 3.8 0%; Power centralization 3.7 0%. Note. Risks are ordered by mean responsibility for each role, shown as the blue bar. Percentage of top 100 companies with each AI role acknowledging the risk is shown in the red bar. The AI welfare risk category was omitted due to the data-quality issues for this risk in the companion Delphi study.

Discussion

We conducted the first systematic review of what the largest companies worldwide publicly say about the risks of AI. Understanding whether companies discuss the most important risks for their sector or AI use matters because governance instruments for AI assume that companies can recognize the risks their AI systems create and act on them. Public acknowledgement of risks is the major channel by which regulators, researchers, and the public can evaluate companies’ practice.

We found that an estimated 62% of the largest 1000 companies worldwide did not publicly acknowledge any AI risks. Even among the largest 100 companies, 32% did not publicly acknowledge any AI risks. Companies in the Information sector acknowledged the most risks. General-purpose AI developers acknowledged the most risks10. The main risks that companies acknowledged were discrimination, weapons and cyberattacks, & fraud and scams.

We found little relationship the risks companies acknowledged and the risks experts prioritize [34]. Companies rarely acknowledged risks with the most severe consequences, except for weapons & cyberattacks. Although companies in more vulnerable sectors acknowledged more risks than companies in less vulnerable sectors, 77% of risks were not acknowledged by companies ‘extremely vulnerable’ to that risk. Companies judged most responsible for addressing an AI risk were not more likely to acknowledge that risk.

Three risks stand out as blind spots: experts expect power centralization, competitive dynamics, and dangerous capabilities to cause among the most severe harms by 2030, yet almost none of the largest 100 companies acknowledged them, even in sectors judged extremely vulnerable to them. Other severe risks including AI misalignment and multi-agent risks were also rarely acknowledged. Several of the risks with the highest expected harm are among the least likely to be publicly acknowledged.

Implications for practice

Corporate AI risk disclosure today is not sufficient to build public and regulator trust

Public disclosure is one of the few ways that regulators, investors, and researchers can assess whether companies recognize and manage the risks from AI. But our findings show most companies do not publicly acknowledge any AI risks. Because so few companies report publicly, outsiders cannot distinguish a company that manages its AI risks well but does not disclose this from a company that does not manage them at all.

Under current disclosure practice, regulators cannot trust that they see a full picture of a sector’s readiness to address AI risks; investors cannot determine whether a company is effectively managing its exposure; and companies that do disclose gain little credit for it, and may even invite criticism, because disclosing AI risks is not yet normative. We cannot expect this to improve without changed expectations from regulators and investors, and improved reporting frameworks.

Corporate AI risk disclosure should be standardized and calibrated

When companies choose which risks to report, they report the risks that affect their own financial position and neglect risks whose costs fall on markets, institutions, and the public [15, 31]. Our results are consistent with this pattern. The risks acknowledged least by our sample were among those experts judged most severe. But expectations tied to sector and role would at least build on an existing tendency, since companies already acknowledged more of the risks their sectors are most vulnerable to.

Improving AI risk disclosure does not require every company to report on every risk, and not all risks are relevant to all companies. A technology company developing general-purpose AI models and a financial company deploying those models to inform lending decisions are not in comparable positions of vulnerability or responsibility. Reviewers of the Hiroshima AI Process framework’s first round reached a similar conclusion, recommending that guidance be tailored by company AI role and size [4].

Corporate climate reporting offers one model that differentiates between a company’s role and its sector. The Greenhouse Gas Protocol classifies a company's emissions into three scopes, allowing it to account for its direct Scope 1 emissions alongside the indirect Scope 2 and Scope 3 emissions from what it buys and sells (WRI & WBCSD, 2011). Climate reporting requirements also differ by sector: IFRS S2 instructs companies to consult disclosure guidance specific to their own sector rather than a single common set of topics (ISSB, 2023).

We propose that a calibrated scheme could set expectations in three tiers, each adapting an existing precedent from sustainability and AI reporting:

  1. All large companies would report on a small common core of risks with broad exposure (e.g., AI-enabled cyber threats).

  2. Companies would report on additional risks where their sector is judged vulnerable (e.g., the finance sector would report on fraud).

  3. Companies would report on additional risks where their AI ecosystem role gives them responsibility (e.g., deployers would report on multi-agent risks).

Under this scheme, all large financial institutions would report on AI-enabled cyber threats due to their size; AI fraud due to their sector vulnerability; and the subset that deploys AI would report on multi-agent risks due to their role responsibility. The companion Delphi study provides an initial set of expert judgements to support the vulnerability and responsibility tiers [34].

Acknowledging that an AI risk is relevant does not require a company to disclose its preparedness. A company can state that AI security vulnerabilities are a relevant risk without compromising its defenses by describing them in detail. For example, all US public companies must publicly describe their cybersecurity risk management, while critical-infrastructure companies report details about incidents and vulnerabilities confidentially to regulators. But apart from frontier AI developers, large companies in most jurisdictions have no confidential channel for reporting AI risks or incidents, beyond existing regulatory requirements for cybersecurity or other critical incidents11. Confidential reporting must complement rather than replace public disclosure. Only a public record allows investors, researchers, and the public to compare companies and track change over time.

Reporting frameworks should ask which risks a company faces and which mitigations address them

We conducted an exploratory analysis of mitigations publicly reported by companies. We found that 71 of the largest 100 companies reported any mitigations, but companies rarely linked publicly reported mitigations to a specific risk. Without that link, an outsider cannot understand what the mitigation involves in practice or whether a mitigation adequately addresses an AI risk. Reporting frameworks could close this gap by requiring the link between mitigations and risks.

The voluntary Hiroshima AI Process framework, launched in February 2025, produced the first structured reports on corporate AI risk management [29]. In its initial round, 24 organizations submitted reports, and reviewers found the responses high-level, inconsistent in scope, and difficult to compare or verify [4]. Version 2.0 of the Hiroshima AI Process framework, released in May 2026, improves on this. It replaces open-ended questions with structured questions, includes tailored questions for model developers, application providers, and deployers, and asks explicitly about agentic AI and about systemic risks with indicative thresholds [36]. Its response options refer to specific risk areas, and one question invites organizations to describe contexts in which their systems could pose intolerable risk. But it does not ask organizations to disclose which AI risks they are exposed to, or which mitigation addresses which risk (OECD, 2026). A company could submit a report describing its AI risk management process without naming any of the AI risks it faces or how it mitigates them.

Catalogues of AI controls already exist to report against. ISO/IEC 42001 requires a statement listing every control with a written justification for its inclusion or exclusion [18]. The NIST AI Risk Management Framework specifies 72 subcategories of practices across four functions, and supports extensions for specific technologies, sectors, or use cases, including general-purpose AI and agentic AI [6, 7, 27]. Reporting against a specified set of controls, and against the risks each addresses, would produce information that can be compared across companies and over time.

Limitations

This study has limitations that temper our conclusions. First, we measured companies’ public disclosure of AI risks; we did not directly assess their internal practices, but we contacted the 178 companies for which contact information was available so they could correct any insufficiencies in the public record we compiled. Our focus on public documents means we are likely to have underestimated company AI risk management practices. In addition, our assessment focused on risk acknowledgement, not practice. For example, a company that acknowledges the risk of AI misalignment has not necessarily implemented effective mitigations to address that risk. Indeed, research on environmental and social reporting finds that companies may strategically disclose instead of changing practice [11, 23], and early investigations into corporate AI ethics principles found that disclosure alone does not indicate practice change [14, 25].

Second, our decision to focus on the largest companies by market capitalization means that we omit many important companies in the AI ecosystem, especially the leading general-purpose AI developers (e.g., OpenAI, Anthropic, DeepSeek), as well as governance actors such as governments and regulators. This decision to focus on the largest companies also means that our sample is not balanced across sectors (e.g., Information and Finance & Insurance companies are overrepresented in our sample), which limits the generalizability of our findings across sectors.

Third, although we made significant efforts to obtain a comprehensive set of documents from the sampled companies including custom web searches, manual checks of company websites, an agentic AI search, and emailing included companies to confirm the set of included documents, we recognize that we may have failed to retrieve some documents for some companies.

Fourth, our findings are a snapshot. We constructed the sample in December 2025, retrieved documents in January and February 2026, and invited companies to correct the record in June 2026. Corporate AI disclosure is changing quickly, as the underlying AI risk, adoption, and governance landscape changes. We anticipate that some companies will have published new documents since our snapshot was frozen. The results are therefore a baseline and point of comparison for the public record in early 2026, but may not reflect current practice at the date of publication.

Future work could address these limitations. Directly surveying or interviewing companies may elicit non-public AI risk management practices, and collating incident reports or enforcement actions could provide independent evidence of company practice. An extension of the research that investigates general-purpose AI developers and governance actors and roles could provide a more comprehensive view of AI risk management practices and gaps across the AI ecosystem. More generally, we experimented with AI tools to assist identification and classification of documents; a research extension focused on automation of the research methods may increase the comprehensiveness of document search and retrieval.

Conclusion

Governance instruments intended to address AI risk assume that companies recognize the risks they are exposed to and act on them. We tested whether the largest companies worldwide do so. We found that 32% of the largest 100 companies, and an estimated 62% of the largest 1000, did not publicly acknowledge any AI risks. Acknowledgements focused on operational risks such as discrimination, while the risks experts judged most severe, such as power centralization and competitive dynamics, went almost unmentioned. The companies experts judged most responsible for addressing a risk were not more likely to acknowledge it.

Corporate AI risk disclosure today resembles climate disclosure before common frameworks existed. It is voluntary, scattered, and dependent on individual company choice, with few incentives to consider risks whose harms are borne by the public. Climate reporting became useful through systematic measurement, then a common framework, then mandates that changed behavior. This study provides an initial measurement step for AI risk. It establishes an empirical baseline that regulators, investors, and researchers can track, and it shows where disclosure expectations should point first. Companies should be asked to report against a standard set of risks, calibrated to their sector and role in the AI ecosystem, and to link each mitigation to the risks it addresses.

Methods

To identify AI risk acknowledgements of the largest companies, we:

  1. Constructed a sample of 200 companies

  2. Searched for public documents for those companies that discussed AI

  3. Screened documents for relevance and retrieved included documents

  4. Extracted relevant passages from public documents

  5. Classified risks described in passages, and determined whether each company acknowledged each risk, based on aggregate risk classifications across all documents

Sample construction

We sampled 200 of the world's largest companies by market capitalization, using rankings from companiesmarketcap.com on 15 December 2025. We constructed the sample by selecting all companies ranked 1-100, plus 100 companies drawn at random from ranks 101 to 1000. We generated the random draw as a sequence of 1,000 numbers without replacement (random.org).

We enriched the sample of companies using the Capital IQ database to obtain the country of its headquarters; its NAICS code; its business description; and its primary website. The most common headquarters country was the United States (n = 89), followed by China (n = 26).

We classified each company as holding one or more roles in the AI ecosystem. We used a structured prompt for LLM agents to conduct autonomous research to find evidence for and against a company holding each of four roles ([27] (Australian Government Department of Industry, Science and Resources 2024)): general-purpose AI developer, specialized AI developer, AI infrastructure provider, and/or AI deployer. A human reviewed all evidence presented by the LLM and determined each company’s role(s). Companies with more than one role were assigned a primary role and tagged with additional roles. The most common role (primary or additional) role was AI deployer (n = 186). Characteristics of included companies are presented in Table 4. Detailed characteristics for each included company, including its sector and AI ecosystem roles, is presented in the Supplementary Materials (Table S2).

Characteristicn%
Country of headquarters
United States8945%
China2613%
Japan116%
India84%
Canada63%
Germany63%
Netherlands63%
Switzerland63%
France53%
United Kingdom53%
Other3216%
Sector
Information2513%
Characteristicn%
Finance and insurance3920%
Trade, transportation and utilities2412%
Agriculture, Mining, Construction and Manufacturing
Computer and electronics manufacturing2211%
Chemicals and pharmaceutical manufacturing189%
Other Agriculture, Mining, Construction and Manufacturing6131%
Other services116%
AI ecosystem role (primary)
AI Developer (General-purpose AI)42%
AI Developer (Specialized AI)147%
AI Infrastructure Provider2211%
AI Deployer14975%
Other116%
AI ecosystem role (any) *
AI Developer (General-purpose AI)158%
AI Developer (Specialized AI)8643%
AI Infrastructure Provider3317%
AI Deployer18693%

Table 4. Characteristics of included companies

Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
1NVIDIA334413117768Computer and electronics manufacturing○○●○
2Apple334220001690Computer and electronics manufacturing○○●
3Alphabet (Google)518210160329Information●○○○
4Microsoft513210012141Information○○●○
5Amazon455219064768Trade, transportation and utilities○○●○
6Broadcom334413180711Computer and electronics manufacturing●
7Meta519290170617Information●○○○
8Saudi Aramco211111334426Other Agriculture, Mining, Construction and Manufacturing○○●
9Tesla336111184996Other Agriculture, Mining, Construction and Manufacturing●○
10TSMC334413201395Computer and electronics manufacturing●○
11Berkshire Hathaway999977002176Other services●
12Walmart455219011259Trade, transportation and utilities○●
13Eli Lilly325412006730Chemicals and pharmaceutical manufacturing●○
14JPMorgan Chase522110002968Finance and insurance○●
15Tencent519130270615Information○○○●
16Visa522320179534Finance and insurance●
17Oracle513210012142Information●○
18Mastercard522320160225Finance and insurance●
19Johnson & Johnson325412006266Chemicals and pharmaceutical manufacturing○●
20Exxon Mobil324110004503Other Agriculture, Mining, Construction and Manufacturing○●
21Samsung334413104604Computer and electronics manufacturing○●○
22Palantir519130036910Information●
23ASML333242061214Other Agriculture, Mining, Construction and Manufacturing●○
24Bank of America522110007647Finance and insurance○●
25Netflix516210147579Information●
26AbbVie325414016101Chemicals and pharmaceutical manufacturing○●
27Costco455211029028Trade, transportation and utilities●
28Agricultural Bank of China522110269783Finance and insurance●
Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
29Alibaba455110020530Trade, transportation and utilities●○○
30LVMH3152014447Other Agriculture, Mining, Construction and Manufacturing●
31Home Depot444110005680Trade, transportation and utilities●
32ICBC522110279378Finance and insurance○●
33AMD334413001161Computer and electronics manufacturing○●
34Procter & Gamble325611008762Chemicals and pharmaceutical manufacturing●
35China Construction Bank522110274364Finance and insurance○●
36Roche325412025648Chemicals and pharmaceutical manufacturing○●
37GE Aerospace336412005047Other Agriculture, Mining, Construction and Manufacturing○●
38UnitedHealth524114010903Finance and insurance●
39Cisco334118020779Computer and electronics manufacturing○●○
40Coca-Cola312111003144Other Agriculture, Mining, Construction and Manufacturing●
41Chevron324110002991Other Agriculture, Mining, Construction and Manufacturing○●
42Wells Fargo522110008007Finance and insurance●
43IBM518210006066Information○●○
44SAP513210103487Information○●
45Morgan Stanley523110012124Finance and insurance●
46Caterpillar333120002817Other Agriculture, Mining, Construction and Manufacturing○●
47AstraZeneca325412028272Chemicals and pharmaceutical manufacturing○●
48Prosus454110333645Trade, transportation and utilities○●
49Micron Technology334413007343Computer and electronics manufacturing●○
50Toyota336110019661Other Agriculture, Mining, Construction and Manufacturing●○
51Goldman Sachs523150114628Finance and insurance●
52American Express522210001447Finance and insurance○●
53Hermes3169203053Other Agriculture, Mining, Construction and Manufacturing●
54SK Hynix334413204049Computer and electronics manufacturing●○
55Bank of China522110267461Finance and insurance●
56HSBC522110015509Finance and insurance●
57PetroChina324110133870Other Agriculture, Mining, Construction and Manufacturing○●
58Novartis325412101310Chemicals and pharmaceutical manufacturing○●
59Kweichow Moutai312140251321Other Agriculture, Mining, Construction and Manufacturing●

Table S2.

Note. N = 200 companies; percentages are of the full sample. * A company can hold more than one AI ecosystem role, so percentages in this section do not sum to 100.

We identified public documents published by each company using four methods: (1) systematic website search using keywords; (2) manually retrieving annual reports; (3) supplementary searches conducted by LLM agents; (4) directly requesting documents and feedback from companies. Methods 1-3 were completed between 1 January and 15 February 2026. Method 4 was completed between 1 June and 27 June 2026.

Systematic website search. We systematically searched each company's official website using the Google Custom Search API. For each company, we retrieved the company website from its Compustat Capital IQ record and ran two site-restricted keyword searches:

  • site:[company_website] "AI" framework

  • site:[company_website] "AI" (responsib* OR protocol OR governance OR policy OR guideline OR commitment OR security OR ethic* OR bias OR "risk management")

To develop the keywords, we first manually assembled a set of known-relevant documents from 10 companies. These 10 companies were deliberately diverse in i) market capitalization rank and ii) sector. We then scored candidate keyword sets on two measures: sensitivity (the proportion of known-relevant documents the keyword set retrieved), and precision (the proportion of retrieved documents that were relevant). We prioritized sensitivity to ensure as many potentially relevant documents were identified as possible. We retained the first 80 results per company across the two searches, ranked according to Google's search ranking.

Annual report retrieval. We manually retrieved the most recent annual report for every company in our sample, wherever the website search had not already captured it. We did this because when we reviewed the results of the systematic website search, we found that several companies returned no potentially relevant records. Retrieval of annual reports was designed to increase the chance of identifying any discussion of AI risk by the company.

Our logic was that annual reporting in the US includes mandated disclosures of risk factors, and recent analyses have found AI risks discussed in this section or its international equivalents [41]. We searched SEC EDGAR for each US company's most recent 10-K form, falling back to the company website where needed, and searched company websites directly for non-US companies.

Supplementary search with LLM agents. We conducted a supplementary search using large language model agents (Claude Opus 4.6) to identify relevant public documents that the primary search had not retrieved. We did this because the systematic website search only covered each company’s primary website domain, so potentially relevant documents on secondary corporate domains, subsidiary websites, or third-party repositories were not retrieved. We provided a structured prompt to the LLM agent with information about inclusion & exclusion criteria, the company to search, and previously included and excluded search results. All results from the LLM agents were screened using the same process as the systematic website search results.

Email to included companies. Prior to manuscript preparation, we attempted to contact all sampled companies via a Qualtrics email campaign to review the documents attributed to their company. We did this because of the lag between our initial search and when we completed the analysis of this dataset.

We searched included documents for company emails and used LLM agents to search the web for public emails related to AI risks, governance, and/or investor relations. We identified emails for 178 out of 200 companies. We sent companies a link to a customized feedback webpage on the project website. On this page, representatives from the company could confirm the included documents, propose any additional or updated documents, and provide any other feedback. 173 of 178 emails were delivered. 16 companies responded. Documents proposed through the feedback webpage were screened using the same criteria as other methods.

Screening and retrieval

Document inclusion and exclusion criteria. We included documents published by a company as an official communication (e.g., policy documents, reports, dedicated web pages, official blog posts), that were publicly accessible, and discussed AI risks. Examples included: a document that discussed AI risks that the company faced; a document that described the company’s own AI policy, principles or governance; and/or a document that described actions the company was taking to mitigate AI risks.

We excluded documents that mentioned AI without discussing risks or where we were unable to verify the document represented the company’s official position, such as:

  • Third-party content such as guest posts on company blogs or reprinted articles

  • Transcripts of interviews or speeches unless published as an official company position

  • Technical documentation such as API references and user manuals

  • Marketing material such as AI product announcements

  • External advice telling other companies how to manage AI risks

  • Press releases announcing partnerships or investments

  • Customer-facing terms of service or privacy policies that did not discuss AI risks

  • Employee-facing internal policies not made intentionally public

  • Inaccessible documents (broken links, paywalls, or authentication requirements)

  • Job postings

  • Community / customer forums

During search, retrieval, and screening, we identified several additional types of documents, for which we developed decision rules during the process and then applied them consistently across all companies:

  • Model cards and other per-model documentation were excluded. These documents address individual systems rather than corporate governance, were never targeted by our collection strategy, and surfaced only incidentally through the agent-based search; including them would have shifted the analysis toward per-model evaluation.

  • Hiroshima AI Process reports were included for every company for which one was identified.

  • We included documentation from subsidiaries for a parent company only if the attribution of the subsidiary’s position for the parent company was clear.

Screening process. We used large language model to triage search results. The LLM was provided with a structured prompt that included the search title, URL, snippet, inclusion/exclusion criteria, and the full-text of the document (where possible to programmatically retrieve). This model provided (1) recommendation of inclusion or exclusion for each search record, (2) confidence rating, and (3) reasoning. We compared the LLM recommendation against 86 human decisions, finding 80% agreement. We considered this acceptable as the LLM recommendation did not make the final inclusion decisions. A separate LLM model provided a provisional judgement of the document type (e.g., annual report, AI policy, job posting).

An expert human reviewer decided whether to include or exclude each record. They reviewed all records identified for a company, grouped by document type and LLM recommendation. Reviewing all records for each company at once allowed the human reviewer to identify duplicate records (e.g., duplicate 2025 annual reports published at different URLs), and distinguish superseded records (e.g., 2024 and 2025 annual reports). For each record, the human reviewer considered the record title, URL, snippet, full-text, document type, and LLM recommendation.

Document retrieval. We attempted to retrieve all 637 included search records as full-text documents for extraction and classification. 89 documents were not retrieved due to a broken link, duplication, or the document being superseded by a later version. Where the document was published as PDF, this was retrieved directly. Where the document was published as a website, a copy of the website was saved at the time of retrieval and converted into PDF. The document type was updated by the human at the time of retrieval.

The final document corpus comprised 539 included and retrieved documents. Table 5 shows details of the document corpus, including document type, year of publication, and source.

Categoryn%
Source
Systematic website search25748%
Annual report retrieval20037%
LLM agent supplementary search8215%
Document type
Annual report22341%
ESG report6612%
AI overview & principles6412%
Corporate policy458%
AI governance346%
Investor materials275%
Other8015%
Publication year
2022 or earlier122%
2023224%
202410419%
202535666%
2026458%
No date00%

Table 5. Characteristics of included documents

Data extraction

PDF conversion to plaintext. We stored each included document as a PDF to preserve a stable copy for analysis. We extracted the full text of each PDF document into plaintext using PyMuPDF [1]. Of the 539 documents, 532 were included in the data extraction pipeline. 7 Chinese-language documents were searched manually by a fluent human reviewer as the pipeline was designed for English language documents only.

Search term development. We searched for passages that discussed AI, AI risks, and/or AI mitigations from each document using a list of 165 search terms. We developed this list by combining generic AI terms (e.g., “AI”, "artificial intelligence", "machine learning", "large language model"), risk terms derived from the MIT AI Risk Repository taxonomy (e.g., "algorithmic bias", "deepfake", "prompt injection"), and risk mitigation terms (e.g., "red teaming", "human oversight", "content provenance"). We iteratively tested candidate terms against sample documents and reviewing the matched passages for precision and recall. The full list of terms is provided in Supplementary materials Table S1.

Passage extraction. We searched each term in each document using case-insensitive regular expressions with word-boundary matching. For each match we extracted a context window of up to 2,000 characters centered on the matched term. Because related terms were often co-located (e.g., a company might mention “we use artificial intelligence (AI) to enhance our products”, where both “artificial intelligence” and “AI” were individual search terms), we merged all matched terms located within 1,000 characters into a single extracted passage. This search procedure yielded 7,085 passages from 528 of the 541 English-language documents. 13 documents contained no keyword matches.

Below we present an example of an extracted passage, with matching search terms in bold: “Factors such as fairness, explainability, sustainability, privacy, etc. must also be considered. Upholding responsible AI practices will ultimately determine the societal value of AI6. Human oversight is key to minimising the risk of Generative AI. Reason being, these models learn from vast amounts of data available on the internet, which can contain biases present in society and may inadvertently perpetuate and amplify these biases. LLMs can be manipulated to generate or spread misinformation, phishing emails, or social engineering attacks. ”

Passage from Intel (2024) “Responsible AI”

Data classification

We classified each passage to determine if it represented a risk acknowledgement. First, an LLM classified each passage to determine whether a risk was present; if present, the LLM classified the specific risk(s) present against a risk taxonomy. Human reviewers then independently coded a random sample of passages to determine non-inferiority.

Risk classification. Passages were classified by an LLM (GPT-5.2) using a structured prompt with two stages: first, whether any risk was acknowledged; second, what specific risk(s) were acknowledged. A risk acknowledgement was defined as a statement in which the company recognizes the existence or relevance of one or more AI-related risks. A risk was defined as the possibility of an unfortunate occurrence that may emerge from the development, deployment or use of AI. Passages describing AI products or capabilities without mentioning risks, generic statements about risk not explicitly linked to AI (e.g., “privacy concerns”), and material reflecting the positions of external parties (such as shareholder proposals) were not classified as risk acknowledgement. If an AI risk was present in the passage, the model labelled the passage with one or more risks from the 24 risks listed in the MIT AI Risk Repository domain taxonomy [38]. We adopted this taxonomy because it is the most comprehensive available synthesis of AI risk classifications, its subdomains are designed to be mutually exclusive and collectively exhaustive, and it enables direct comparison with the companion Delphi study that uses the same structure [34]. The model also provided a direct evidence quote from the passage supporting each risk classification, and a confidence rating.

Below we present an example of a classified passage, with evidence quotes in bold, annotated with risk labels in square brackets:

“Factors such as fairness, explainability, sustainability, privacy, etc. must also be considered. Upholding responsible AI practices will ultimately determine the societal value of AI6. Human oversight is key to minimising the risk of Generative AI. Reason being, these models learn from vast amounts of data available on the internet, which can contain biases present in society and may inadvertently perpetuate and amplify these biases [Risk 1.1: Discrimination & misrepresentation]. LLMs can be manipulated to generate or spread misinformation, phishing emails, or social engineering attacks” [Risk 4.3: Fraud, scams, and targeted manipulation].

Passage from Intel (2024) “Responsible AI”

Risk validation. We validated both stages of the model's classifications against independent human coding. We drew a random sample of 500 passages, stratified by whether any risk was acknowledged for a given passage. For the first stage (presence / absence of any risk acknowledgement), model-human agreement (Cohen's κ = 0.64-0.88 across coder pairings) exceeded human-human agreement (κ=0.43\kappa= 0.43-0.720.72). For the second stage (specific risk(s) classified), mean Jaccard similarity between the model and a human was 0.510.51-0.760.76, compared with 0.480.48-0.640.64 between humans.

Determination of company risk acknowledgements and mitigation reports. A human reviewer considered all available evidence to determine whether a company acknowledged each risk. For each risk, the human reviewed all LLM classifications for that risk across all documents for that company. They used information from the LLM, the evidence quote, the passage, and the original PDF documents to determine whether the company acknowledged that risk. The process was repeated to determine company mitigation reporting.

References

Supplementary Materials

Supplementary Tables and Figures

Table S1. Search terms used for passage retrieval

CategorySearch terms
Generic AI termsAI, AI agent, artificial intelligence, chatbot, deep learning, foundation model, general purpose AI, GPAI, large language model, LLM, machine learning
AI risk terms
1.1 Unfair discrimination and misrepresentationAI bias, AI fairness, algorithmic bias, biased model
1.2 Exposure to toxic contentAI-generated hate, content filtering, content moderation, content safety, deep fake, deepfake, harmful content
1.3 Unequal performance across groupsmodel performance across groups, performance gap across
2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive informationdata memorization, inference attack, training data extraction
2.2 AI system security vulnerabilities and attacksadversarial attack, AI security, data poisoning, jailbreak, model extraction, model security, model tamper, prompt injection, training data attack
3.1 False or misleading informationAI accuracy, AI-generated misinfo, confabulation, hallucination
3.2 Pollution of information ecosystem and loss of consensus realityAI-generated content, content provenance, deep fake, deepfake, digital watermark, filter bubble, generative AI content, synthetic content, synthetic media
4.1 Disinformation, surveillance, and influence at scaleAI persuasion, AI propaganda, AI surveillance, AI-enabled disinformation, AI-powered surveillance, election interference
4.2 Cyberattacks, weapon development or use, and mass harmAgent cyberattack, AI weapon, AI-enabled attack, AI-enabled cyberattack, AI-generated malware, autonomous weapon, CBRN, dual use AI, dual-use AI, lethal autonomous
4.3 Fraud, scams, and targeted manipulationAI fraud, AI impersonation, AI scam, AI-enabled fraud, AI-facilitated plagiarism, AI-generated phishing, deep fake, deepfake, synthetic identity, voice clon
5.1 Overreliance and unsafe useAI dependence, AI over-trust, human oversight, human review, human supervision, human-in-the-loop, overreliance on AI, unsafe use
5.2 Loss of human agency and autonomyAI autonomy, algorithmic decision, automated decision-making, human agency, human control, right to explanation
6.1 Power centralization and unfair distribution of benefitsAI market concentration, AI monopoly, AI power concentration
6.2 Increased inequality and decline in employment qualityAI and employment, AI displacement, AI job loss, AI workforce impact, AI-driven job, workforce disruption
6.3 Economic and cultural devaluation of human effortAI and copyright, AI and intellectual property, AI-created works, AI-generated content, data provenance, training data rights
6.4 Competitive dynamicsAI arms race, AI race, competitive pressure, responsible scaling
6.5 Governance failureAI compliance, AI law, AI legislation, AI regulatory, AI-related legislative, EU AI Act, evolving AI regulation
6.6 Environmental harmAI carbon footprint, AI energy consumption, AI environmental impact, AI sustainability, data center energy, GPU energy
7.1 AI pursuing its own goals in conflict with human goals or valuesAI alignment, AI safety, goal misgeneralisation, goal misspecification, misalignment, reward hacking, value alignment

Table S1. Search terms used for passage retrieval

CategorySearch terms
7.2 AI possessing dangerous capabilitiesAI capability threshold, autonomous AI, capability evaluation, dangerous capabilities, frontier AI, frontier model, model evaluation, red teaming, safety evaluation, self-proliferation, self-replication
7.3 Lack of capability or robustnessAI error, AI failure, AI reliability, AI testing, model accuracy, model failure, model reliability, model robustness
7.4 Lack of transparency or interpretabilityAI transparency, contestability, explainability, explainable AI, interpretab, model explanation, right to explanation, right to recourse, XAI
7.5 AI welfare and rightsAI consciousness, AI moral patient, AI rights, AI sentience, AI welfare, digital minds
7.6 Multi-agent risksagentic AI, AI ecosystem, AI system interaction, multi-agent
Mitigation termsAI compliance, AI ethics, AI governance, AI legislation, AI testing, AI transparency, capability evaluation, content filtering, content moderation, content provenance, content safety, digital watermark, EU AI Act, explainability, explainable AI, human control, human oversight, human review, human supervision, human-in-the-loop, interpretab, model evaluation, red teaming, responsible AI, responsible scaling, safety case, safety evaluation, safety framework

Table 7. Note. AI risk terms are categorized by the MIT AI Risk Repository risk taxonomy. Search terms are repeated across all relevant categories. The generic terms AI, LLM, and GPAI were matched on word boundaries to avoid matching inside longer words.

Table S2. Detailed characteristics of included companies

Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
60Nestle311016603Other Agriculture, Mining, Construction and Manufacturing●
61CATL335910327415Other Agriculture, Mining, Construction and Manufacturing○●
62Merck325412007257Chemicals and pharmaceutical manufacturing○●
63Salesforce513210157855Information○○●
64China Mobile517312065662Information○○●○
65RTX336412010983Other Agriculture, Mining, Construction and Manufacturing●
66Philip Morris International312230179621Other Agriculture, Mining, Construction and Manufacturing●
67International Holding Company999977337385Other services●
68Royal Bank Of Canada522110015633Finance and insurance○●
69L'Oreal325620100581Chemicals and pharmaceutical manufacturing○●
70Reliance Industries324110207074Other Agriculture, Mining, Construction and Manufacturing○●
71AppLovin513210038212Information●○
72McDonald's722513007154Other services●
73Novo Nordisk325412008020Chemicals and pharmaceutical manufacturing●
74T-Mobile US517112017874Information●
75Siemens335019349Other Agriculture, Mining, Construction and Manufacturing○●
76Abbott Laboratories334510001078Computer and electronics manufacturing○●
77Thermo Fisher Scientific334516010530Computer and electronics manufacturing●
78Shopify519130023650Information●
79Applied Materials333242001704Other Agriculture, Mining, Construction and Manufacturing○●
80Shell324110012384Other Agriculture, Mining, Construction and Manufacturing○●
81Pepsico31211008479Other Agriculture, Mining, Construction and Manufacturing●
82Citigroup522110003243Finance and insurance●
83Inditex4481245663Trade, transportation and utilities●
84Lam Research333242006565Other Agriculture, Mining, Construction and Manufacturing●
85Walt Disney516120003980Information●
86Linde325120025124Chemicals and pharmaceutical manufacturing○●
87Intuitive Surgical334510136725Computer and electronics manufacturing●○
88QUALCOMM334413024800Computer and electronics manufacturing○●○
89Intuit513210027928Information○●
Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
90Blackstone Group523940177088Finance and insurance○○●
91China Life Insurance524113259503Finance and insurance●
92HDFC Bank522110144535Finance and insurance●
93GE Vernova333611045169Other Agriculture, Mining, Construction and Manufacturing●
94Mitsubishi UFJ Financial522110252940Finance and insurance●
95Intel334413006008Computer and electronics manufacturing●○
96ServiceNow513210171007Information●○○
97Airbus336411220833Other Agriculture, Mining, Construction and Manufacturing○●
98Uber485310035077Trade, transportation and utilities○○●
99AT&T517112009899Information●
100Charles Schwab523940013988Finance and insurance○●
107Amgen325414001602Chemicals and pharmaceutical manufacturing○●
118PDD Holdings (Pinduoduo)455033861Trade, transportation and utilities●
126EssilorLuxottica339115101248Other Agriculture, Mining, Construction and Manufacturing●
134Safran336412101305Other Agriculture, Mining, Construction and Manufacturing●○
149ABB335313210418Other Agriculture, Mining, Construction and Manufacturing○●
153Eaton335314004199Other Agriculture, Mining, Construction and Manufacturing●
190Newmont212220007881Other Agriculture, Mining, Construction and Manufacturing●
194Investor AB523999015699Finance and insurance●
225Advantest334515100245Computer and electronics manufacturing●○
242Itōchū Shōji42100435Trade, transportation and utilities●
248China Telecom517111254338Information○○●○
253National Australia Bank522110014802Finance and insurance●
262PNC Financial Services522110008245Finance and insurance●
287Delta Electronics3344210201Computer and electronics manufacturing○●
295Grupo México2122220509Other Agriculture, Mining, Construction and Manufacturing●
301Barrick Gold212221002055Other Agriculture, Mining, Construction and Manufacturing●
310Coinbase523210034749Finance and insurance○●
321Mercedes-Benz336110017828Other Agriculture, Mining, Construction and Manufacturing●
323Bajaj Finance522291284229Finance and insurance●
341Hilton Worldwide721110005643Other services●
Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
348Volkswagen336110100737Other Agriculture, Mining, Construction and Manufacturing○●
350America Movil517112142248Information●
356Fortinet513210183377Information●○
378Bank Central Asia522110242865Finance and insurance●
383Glencore2121267407Other Agriculture, Mining, Construction and Manufacturing●
427Argenx325414318005Chemicals and pharmaceutical manufacturing●
430Vale212210209382Other Agriculture, Mining, Construction and Manufacturing●
432Air Products and Chemicals325120001209Chemicals and pharmaceutical manufacturing●
433Suncor Energy324110015070Other Agriculture, Mining, Construction and Manufacturing●
447Alnylam Pharmaceuticals325414264265Chemicals and pharmaceutical manufacturing●
454East Money Information519130294212Information○●
459Porsche336111102187Other Agriculture, Mining, Construction and Manufacturing○●
462Constellation Software518210174302Information●
472National Bank of Canada522110015620Finance and insurance●
477Edwards Lifesciences339113133366Other Agriculture, Mining, Construction and Manufacturing●
480Marubeni421100545Trade, transportation and utilities●
493Ferrovial237271217Other Agriculture, Mining, Construction and Manufacturing●
506NAURA Technology Group333242294042Other Agriculture, Mining, Construction and Manufacturing●
515CITIC limited999977105198Other services●
531UOB522110015679Finance and insurance●
534Xcel Energy2211007977Trade, transportation and utilities●
536FICO541512013824Other services○○
540Baidu519290164532Information●○○○
545Fubon Financial524113249738Finance and insurance●
555EQT523910333866Finance and insurance●
5643i Group523910210835Finance and insurance●
573Vingroup Company531288024Other services●○
591PSEG22111008810Trade, transportation and utilities●
630Daikin333415101138Other Agriculture, Mining, Construction and Manufacturing○●
635ResMed334510031887Computer and electronics manufacturing○●
Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
639Ventas531110110179Other services●
640Toyota Tsusho423840101940Trade, transportation and utilities○●
644Ahold Delhaize445110023667Trade, transportation and utilities●
653China State Shipbuilding Corporation (CSSC)336611271228Other Agriculture, Mining, Construction and Manufacturing●
655KONE333921101023Other Agriculture, Mining, Construction and Manufacturing○●
678Otis Worldwide333921036190Other Agriculture, Mining, Construction and Manufacturing●
679United Airlines Holdings481111010795Trade, transportation and utilities●
685Kimberly-Clark322120006435Other Agriculture, Mining, Construction and Manufacturing●
705Bank Rakyat Indonesia522110157304Finance and insurance○●
706Old Dominion Freight Line484122024617Trade, transportation and utilities●
712Kenvue3256041897Chemicals and pharmaceutical manufacturing○○●
717Sumitomo Denki Kōgyō335929019413Other Agriculture, Mining, Construction and Manufacturing○●
718COSCO Shipping541614200993Other services○●
747Bharat Electronics334511211040Computer and electronics manufacturing○●
753Swiss Life524113205699Finance and insurance●
775Viking Holdings483114045220Trade, transportation and utilities●
786JSW Steel331110211275Other Agriculture, Mining, Construction and Manufacturing●
788Transurban488490212650Trade, transportation and utilities●
799SAAB AB3364222505Other Agriculture, Mining, Construction and Manufacturing●○
806Novonesis3259241336Chemicals and pharmaceutical manufacturing○●○
808Komatsu333120006494Other Agriculture, Mining, Construction and Manufacturing○●
819Mettler Toledo334516065772Computer and electronics manufacturing○●
823PKO Bank Polski522110271187Finance and insurance○●
826Svenska Handelsbanken522110015654Finance and insurance●
837Shenzhen Inovance335314295346Other Agriculture, Mining, Construction and Manufacturing○●
839Naturgy221210220586Trade, transportation and utilities●
842DMart445110323714Trade, transportation and utilities●
881DexCom334510162887Computer and electronics manufacturing●○
882Nornickel331410248384Other Agriculture, Mining, Construction and Manufacturing○●
887Yili Group3115260029Other Agriculture, Mining, Construction and Manufacturing●
Market cap rankCompanyNAICS codeGVKEYSectorGeneral-purpose developerSpecialized developerInfrastructure providerDeployerOther
889Cincinnati Financial524126014824Finance and insurance●
892Luzhou Laojiao312140243767Other Agriculture, Mining, Construction and Manufacturing●
904FirstEnergy22111008099Trade, transportation and utilities●
907Hapag-Lloyd483111100431Trade, transportation and utilities●
923Renesas Electronics334413257664Computer and electronics manufacturing●○
927AerCap532411176092Other services●
933Steris339112025313Other Agriculture, Mining, Construction and Manufacturing●
946DiDi485999038998Trade, transportation and utilities●○
947Halliburton213112005439Other Agriculture, Mining, Construction and Manufacturing○●
949China Minsheng Bank522110249482Finance and insurance●
968Jabil334412028195Computer and electronics manufacturing●○
969Devon Energy2111014934Other Agriculture, Mining, Construction and Manufacturing●
972Tata Steel331110203892Other Agriculture, Mining, Construction and Manufacturing○●
973Wolters Kluwer513210101361Information○●
976Shandong Gold Mining212220270346Other Agriculture, Mining, Construction and Manufacturing●
982PPG Industries325510008247Chemicals and pharmaceutical manufacturing●
983China Northern Rare Earth (CNREHT)331110271044Other Agriculture, Mining, Construction and Manufacturing●
986Vedanta331410208323Other Agriculture, Mining, Construction and Manufacturing●
993T. Rowe Price523940012138Finance and insurance●
994Nomura Research Institute541519249604Other services○○●

Table S-6. Note. Market cap rank is the company's rank by market capitalization on companiesmarketcap.com on 15 December 2025. GVKEY is the company's identifier in Compustat Capital IQ, and used to determine its NAICS code and other company-level information used for analysis. Companies could hold multiple AI ecosystem roles: ● = primary role; ○ = additional role(s).

Table S3. Percentage of companies acknowledging each AI risk, by sector: top 100 and top 1,000 companies

Note for all sectors: Estimates for the largest 1,000 companies are design-based weighted estimates with a 95% confidence interval, combining the largest 100 companies (n=100n = 100; weight = 1) with a random sample (n=100n = 100; weight = 9) drawn from ranks 101–1,000. The judgement for AI welfare risk is not reported due to data quality issues in the Delphi study.

Sector: Information

Note. 18 / 100 companies in the top 100 were in this sector; 7 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 81 companies. Vulnerability is the expert judgement of the Information sector’s vulnerability to each risk from the Delphi study.

Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Sector: Computer & electronics manufacturing

AI riskTop 100 companiesTop 1,000 companies*Vulnerability
%%95% CI [LL, UL]MedianLabel
Discrimination43%27.9%[9.6, 61.5]3Moderate
Weapons & cyberattacks50%18.6%[8.4, 52.2]3Moderate
Fraud & scams36%16.3%[6.1, 49.9]3Moderate
Loss of privacy36%16.3%[6.1, 49.9]3Moderate
Devaluation of human creativity36%5.8%[5.8, 37.2]3Moderate
Transparency & interpretability7%1.2%[1.2, 32.6]3Moderate
Capability & robustness0%0.0%[0.0, 31.4]3Moderate
False information7%1.2%[1.2, 32.6]3Moderate
Environmental harm36%26.7%[8.5, 60.3]5Extreme
AI security vulnerabilities14%2.3%[2.3, 33.7]3Moderate
Disinformation & influence7%1.2%[1.2, 32.6]3Moderate
Toxic content7%1.2%[1.2, 32.6]2Minimal
Governance failure14%2.3%[2.3, 33.7]3Moderate
Overreliance & unsafe use0%0.0%[0.0, 31.4]3Moderate
Unequal performance0%0.0%[0.0, 31.4]3Moderate
Loss of human agency0%0.0%[0.0, 31.4]3Moderate
Dangerous capabilities0%0.0%[0.0, 31.4]3Moderate
Inequality & unemployment7%1.2%[1.2, 32.6]3Moderate
AI misalignment0%0.0%[0.0, 31.4]3Moderate
Multi-agent risks0%0.0%[0.0, 31.4]3Moderate
Power centralization0%0.0%[0.0, 31.4]3Moderate
Loss of consensus reality0%0.0%[0.0, 31.4]3Moderate
Competitive dynamics0%0.0%[0.0, 31.4]3Moderate
AI welfare0%0.0%[0.0, 31.4]––

Table 8.

Note. 14 / 100 companies in the top 100 were in this sector; 8 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 86 companies. Vulnerability is the expert judgement of the vulnerability to each risk of the Agriculture, mining, construction & manufacturing sector, which contains this group, from the Delphi study.

  • Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Sector: Finance & insurance

AI riskTop 100 companiesTop 1,000 companies95% CI [LL, UL]MedianLabel
Discrimination48%30.1%[12.7, 47.4]5Extreme
Weapons & cyberattacks43%14.8%[6.1, 35.6]5Extreme
Fraud & scams52%30.6%[13.3, 47.9]5Extreme
Loss of privacy29%13.1%[4.5, 34.0]5Extreme
Devaluation of human creativity14%6.6%[1.8, 25.8]4High
Transparency & interpretability24%17.5%[3.1, 31.9]5Extreme
Capability & robustness19%7.1%[2.3, 26.3]5Extreme
False information19%7.1%[2.3, 26.3]5Extreme
Environmental harm10%1.1%[1.1, 15.8]5Extreme
AI security vulnerabilities10%1.1%[1.1, 15.8]5Extreme
Disinformation & influence10%6.0%[1.2, 25.3]4High
Toxic content0%0.0%[0.0, 14.8]4High
Governance failure5%0.5%[0.5, 15.3]5Extreme
Overreliance & unsafe use0%4.9%[0.1, 24.2]5Extreme
Unequal performance0%0.0%[0.0, 14.8]5Extreme
Loss of human agency0%0.0%[0.0, 14.8]5Extreme
Dangerous capabilities0%0.0%[0.0, 14.8]5Extreme
Inequality & unemployment5%0.5%[0.5, 15.3]4High
AI misalignment0%0.0%[0.0, 14.8]5Extreme
Multi-agent risks0%0.0%[0.0, 14.8]5Extreme
Power centralization0%0.0%[0.0, 14.8]5Extreme
Loss of consensus reality0%0.0%[0.0, 14.8]4High
Competitive dynamics0%0.0%[0.0, 14.8]4High
AI welfare0%0.0%[0.0, 14.8]––

Table 9. Note. 21 / 100 companies in the top 100 were in this sector; 18 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 183 companies. Vulnerability is the expert judgement of the Finance & insurance sector’s vulnerability to each risk from the Delphi study.

Sector: Trade, transport, & utilities

AI riskTop 100 companies %Top 1,000 companies %95% CI [LL, UL]MedianLabel
Discrimination50%8.6%[2.8, 31.3]3Moderate
Weapons & cyberattacks13%12.5%[2.1, 37.0]4High
Fraud & scams38%2.0%[2.0, 19.7]4High
Loss of privacy13%6.6%[0.8, 29.3]3Moderate
Devaluation of human creativity13%12.5%[2.1, 37.0]3Moderate
Transparency & interpretability38%2.0%[2.0, 19.7]3Moderate
Capability & robustness25%1.3%[1.3, 19.1]4High
False information13%0.7%[0.7, 18.4]3Moderate
Environmental harm0%0.0%[0.0, 17.8]5Extreme
AI security vulnerabilities25%1.3%[1.3, 19.1]4High
Disinformation & influence13%0.7%[0.7, 18.4]3Moderate
Toxic content13%0.7%[0.7, 18.4]3Moderate
Governance failure0%0.0%[0.0, 17.8]4High
Overreliance & unsafe use0%0.0%[0.0, 17.8]3Moderate
Unequal performance0%0.0%[0.0, 17.8]3Moderate
Loss of human agency0%0.0%[0.0, 17.8]4High
Dangerous capabilities13%0.7%[0.7, 18.4]4High
Inequality & unemployment0%0.0%[0.0, 17.8]3Moderate
AI misalignment0%0.0%[0.0, 17.8]4High
Multi-agent risks0%0.0%[0.0, 17.8]4High
Power centralization0%0.0%[0.0, 17.8]4High
Loss of consensus reality0%0.0%[0.0, 17.8]3Moderate
Competitive dynamics0%0.0%[0.0, 17.8]4High
AI welfare0%0.0%[0.0, 17.8]––

Table 10. Note. 8 / 100 companies in the top 100 were in this sector; 16 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 152 companies. Vulnerability is the expert judgement of the vulnerability to each risk of the Trade, transportation & utilities sector, which contains this group, from the Delphi study.

Sector: Chemicals & pharmaceutical manufacturing

AI riskTop 100 companiesTop 1,000 companies*95% CI [LL, UL]MedianLabel
Discrimination64%45.9%[16.3, 75.6]3Moderate
Weapons & cyberattacks0%24.3%[3.1, 60.4]3Moderate
Fraud & scams0%0.0%[0.0, 36.5]3Moderate
Loss of privacy27%28.4%[7.2, 64.5]3Moderate
Devaluation of human creativity18%2.7%[2.7, 39.2]3Moderate
Transparency & interpretability45%18.9%[7.1, 56.0]3Moderate
Capability & robustness27%4.1%[4.1, 40.5]3Moderate
False information18%14.9%[3.0, 52.0]3Moderate
Environmental harm9%1.4%[1.4, 37.8]5Extreme
AI security vulnerabilities0%0.0%[0.0, 36.5]3Moderate
Disinformation & influence9%1.4%[1.4, 37.8]3Moderate
Toxic content0%0.0%[0.0, 36.5]2Minimal
Governance failure0%12.2%[0.3, 49.3]3Moderate
Overreliance & unsafe use9%1.4%[1.4, 37.8]3Moderate
Unequal performance9%1.4%[1.4, 37.8]3Moderate
Loss of human agency0%0.0%[0.0, 36.5]3Moderate
Dangerous capabilities0%0.0%[0.0, 36.5]3Moderate
Inequality & unemployment0%0.0%[0.0, 36.5]3Moderate
AI misalignment0%0.0%[0.0, 36.5]3Moderate
Multi-agent risks0%0.0%[0.0, 36.5]3Moderate
Power centralization0%0.0%[0.0, 36.5]3Moderate
Loss of consensus reality0%0.0%[0.0, 36.5]3Moderate
Competitive dynamics0%0.0%[0.0, 36.5]3Moderate
AI welfare0%0.0%[0.0, 36.5]––

Table 11.

Note. 11 / 100 companies in the top 100 were in this sector; 7 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 74 companies. Vulnerability is the expert judgement of the vulnerability to each risk of the Agriculture, mining, construction & manufacturing sector, which contains this group, from the Delphi study.

Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Sector: Other agriculture, mining, construction, & manufacturing

AI riskTop 100 companiesTop 1,000 companies95% CI [LL, UL]MedianLabel
Discrimination%%95% CI [LL, UL]MedianLabel
Discrimination4%0.3%[0.3, 8.0]3Moderate
Weapons & cyberattacks20%9.2%[1.3, 17.1]3Moderate
Fraud & scams8%5.7%[1.2, 17.9]3Moderate
Loss of privacy4%2.9%[0.4, 13.8]3Moderate
Devaluation of human creativity8%3.2%[0.6, 14.1]3Moderate
Transparency & interpretability4%0.3%[0.3, 8.0]3Moderate
Capability & robustness4%0.3%[0.3, 8.0]3Moderate
False information0%2.6%[0.1, 13.5]3Moderate
Environmental harm4%2.9%[0.4, 13.8]5Extreme
AI security vulnerabilities4%2.9%[0.4, 13.8]3Moderate
Disinformation & influence0%2.6%[0.1, 13.5]3Moderate
Toxic content0%0.0%[0.0, 7.7]2Minimal
Governance failure0%0.0%[0.0, 7.7]3Moderate
Overreliance & unsafe use0%0.0%[0.0, 7.7]3Moderate
Unequal performance0%2.6%[0.1, 13.5]3Moderate
Loss of human agency0%0.0%[0.0, 7.7]3Moderate
Dangerous capabilities0%0.0%[0.0, 7.7]3Moderate
Inequality & unemployment0%2.6%[0.1, 13.5]3Moderate
AI misalignment0%0.0%[0.0, 7.7]3Moderate
Multi-agent risks0%0.0%[0.0, 7.7]3Moderate
Power centralization0%0.0%[0.0, 7.7]3Moderate
Loss of consensus reality0%0.0%[0.0, 7.7]3Moderate
Competitive dynamics0%0.0%[0.0, 7.7]3Moderate
AI welfare0%0.0%[0.0, 7.7]––

Table 12. Note. 25 / 100 companies in the top 100 were in this sector; 36 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 349 companies. Vulnerability is the expert judgement of the vulnerability to each risk of the Agriculture, mining, construction & manufacturing sector, which contains this group, from the Delphi study.

Sector: Other services

AI riskTop 100 companiesTop 1,000 companies*95% CI [LL, UL]MedianLabel
Discrimination0%24.0%[3.1, 62.5]––
Weapons & cyberattacks0%36.0%[5.4, 66.6]––
Fraud & scams0%0.0%[0.0, 36.0]––
Loss of privacy0%12.0%[0.3, 50.5]––
Devaluation of human creativity0%24.0%[3.1, 62.5]––
Transparency & interpretability0%0.0%[0.0, 36.0]––
Capability & robustness0%12.0%[0.3, 50.5]––
False information0%24.0%[3.1, 62.5]––
Environmental harm0%12.0%[0.3, 50.5]––
AI security vulnerabilities0%0.0%[0.0, 36.0]––
Disinformation & influence0%12.0%[0.3, 50.5]––
Toxic content0%0.0%[0.0, 36.0]––
Governance failure0%12.0%[0.3, 50.5]––
Overreliance & unsafe use0%0.0%[0.0, 36.0]––
Unequal performance0%0.0%[0.0, 36.0]––
Loss of human agency0%0.0%[0.0, 36.0]––
Dangerous capabilities0%0.0%[0.0, 36.0]––
Inequality & unemployment0%0.0%[0.0, 36.0]––
AI misalignment0%0.0%[0.0, 36.0]––
Multi-agent risks0%0.0%[0.0, 36.0]––
Power centralization0%0.0%[0.0, 36.0]––
Loss of consensus reality0%0.0%[0.0, 36.0]––
Competitive dynamics0%0.0%[0.0, 36.0]––
AI welfare0%0.0%[0.0, 36.0]––

Table 13.

Note. 3 / 100 companies in the top 100 were in this sector; 8 / 100 sampled companies in ranks 101–1,000 were in this sector. We estimate that the top 1,000 estimates represent about 75 companies. This group’s companies fall in four different expert-rated sectors, so no vulnerability judgement is shown.

  • Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Table S4. Percentage of companies acknowledging each AI risk, by AI ecosystem role: top 100 and top 1,000 companies

Note for all roles: Estimates for the largest 1,000 companies are design-based weighted estimates with a 95% confidence interval, combining the largest 100 companies (n = 100; weight = 1) with a random sample (n = 100; weight = 9) drawn from ranks 101–1,000. The judgement for AI welfare risk is not reported due to data quality issues in the Delphi study.

Role: AI Developer (General-purpose AI)

Note. 13 / 100 companies in the top 100 held this role; 2 / 100 sampled companies in ranks 101–1,000 held this role. We estimate that the top 1,000 estimates represent about 31 companies. Responsibility is the expert judgement of general-purpose AI developers’ responsibility for each risk from the Delphi study. * Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Role: AI Infrastructure Provider

AI riskTop 100 companies %Top 1,000 companies* %95% CI [LL, UL]Responsibility MeanResponsibility Label
Discrimination48%40.2%[16.6, 63.8]2.4Minimal
Weapons & cyberattacks52%22.7%[13.6, 52.5]3.8High
Fraud & scams44%11.3%[11.3, 39.2]3.6High
Loss of privacy40%38.1%[14.5, 61.8]3.6High
Devaluation of human creativity44%29.9%[13.7, 59.7]2.6Moderate
Transparency & interpretability20%14.4%[5.4, 44.2]2.6Moderate
Capability & robustness20%14.4%[5.4, 44.2]3.2Moderate
False information24%24.7%[8.5, 54.5]2.5Moderate
Environmental harm40%47.4%[23.0, 71.8]4.8Primary
AI security vulnerabilities40%10.3%[10.3, 38.1]4.4High
Disinformation & influence20%14.4%[5.4, 44.2]2.9Moderate
Toxic content16%13.4%[4.4, 43.2]2.5Minimal
Governance failure12%12.4%[3.3, 42.2]3.2Moderate
Overreliance & unsafe use8%2.1%[2.1, 29.9]2.8Moderate
Unequal performance8%2.1%[2.1, 29.9]2.6Moderate
Loss of human agency4%1.0%[1.0, 28.9]2.8Moderate
Dangerous capabilities12%3.1%[3.1, 30.9]3.2Moderate
Inequality & unemployment8%11.3%[2.3, 41.1]2.7Moderate
AI misalignment8%2.1%[2.1, 29.9]3.0Moderate
Multi-agent risks4%1.0%[1.0, 28.9]3.3Moderate
Power centralization0%0.0%[0.0, 27.8]3.5High
Loss of consensus reality0%0.0%[0.0, 27.8]2.4Minimal
Competitive dynamics0%0.0%[0.0, 27.8]3.1Moderate
AI welfare0%0.0%[0.0, 27.8]––

Table 14.

Note. 25 / 100 companies in the top 100 held this role; 8 / 100 sampled companies in ranks 101–1,000 held this role. We estimate that the top 1,000 estimates represent about 97 companies. Responsibility is the expert judgement of AI infrastructure providers’ responsibility for each risk from the Delphi study.

* Because so few companies in ranks 101–1,000 were sampled, these estimates should be treated with caution.

Role: AI Developer (Specialized AI)

AI riskTop 100 companies %Top 1,000 companies %95% CI [LL, UL]Responsibility MeanLabel
Discrimination49%20.3%[10.5, 30.1]4.1High
Weapons & cyberattacks34%20.6%[10.0, 31.1]4.7Primary
Fraud & scams26%11.7%[3.9, 19.6]4.0High
Loss of privacy32%15.1%[6.2, 24.1]4.4High
Devaluation of human creativity32%15.1%[6.2, 24.1]3.8High
Transparency & interpretability25%11.4%[3.6, 19.3]4.3High
Capability & robustness26%6.6%[4.1, 17.4]4.8Primary
False information21%16.0%[6.2, 25.8]4.7Primary
Environmental harm23%11.1%[3.3, 19.0]4.1High
AI security vulnerabilities25%3.7%[3.7, 11.4]4.6Primary
Disinformation & influence17%7.7%[3.2, 19.7]4.0High
Toxic content17%5.1%[2.6, 15.9]4.2High
Governance failure9%4.0%[1.5, 14.8]3.8High
Overreliance & unsafe use9%1.4%[1.4, 9.1]4.2High
Unequal performance8%1.1%[1.1, 8.9]4.5Primary
Loss of human agency6%0.9%[0.9, 8.6]3.9High
Dangerous capabilities8%1.1%[1.1, 8.9]4.4High
Inequality & unemployment4%8.3%[0.4, 16.1]3.4Moderate
AI misalignment4%0.6%[0.6, 8.3]4.1High
Multi-agent risks2%0.3%[0.3, 8.0]4.6Primary
Power centralization0%0.0%[0.0, 7.7]3.2Moderate
Loss of consensus reality0%0.0%[0.0, 7.7]4.1High
Competitive dynamics0%2.6%[0.1, 13.4]3.7High
AI welfare0%0.0%[0.0, 7.7]––

Table 15. Note. 53 / 100 companies in the top 100 held this role; 33 / 100 sampled companies in ranks 101–1,000 held this role. We estimate that the top 1,000 estimates represent about 350 companies. Responsibility is the expert judgement of specialized AI developers’ responsibility for each risk from the Delphi study.

Role: AI Deployer

AI riskTop 100 companiesTop 1,000 companies95% CI [LL, UL]MeanLabel
%%95% CI [LL, UL]MeanLabel
Discrimination43%20.4%[13.8, 27.1]4.3High
Weapons & cyberattacks34%16.6%[10.4, 22.7]4.1High
Fraud & scams28%10.9%[6.0, 15.9]4.1High
Loss of privacy27%12.8%[7.3, 18.3]4.3High
Devaluation of human creativity24%11.5%[6.3, 16.7]4.0High
Transparency & interpretability20%8.1%[3.7, 12.4]4.1High
Capability & robustness19%4.0%[2.2, 9.0]4.3High
False information17%9.7%[4.8, 14.7]4.0High
Environmental harm15%6.5%[2.5, 10.5]4.6Primary
AI security vulnerabilities16%2.6%[1.7, 7.1]4.3High
Disinformation & influence13%4.3%[1.2, 7.4]4.2High
Toxic content9%2.0%[1.0, 6.4]4.2High
Governance failure7%2.8%[1.0, 7.7]3.8High
Overreliance & unsafe use6%1.7%[0.7, 6.1]4.1High
Unequal performance4%1.4%[0.5, 5.8]4.3High
Loss of human agency4%0.4%[0.4, 3.4]4.3High
Dangerous capabilities4%0.4%[0.4, 3.4]4.1High
Inequality & unemployment3%3.3%[0.2, 6.4]3.9High
AI misalignment2%0.2%[0.2, 3.2]3.8High
Multi-agent risks1%0.1%[0.1, 3.1]4.7Primary
Power centralization0%0.0%[0.0, 3.0]3.7High
Loss of consensus reality0%0.0%[0.0, 3.0]4.2High
Competitive dynamics0%1.0%[0.0, 5.4]3.8High
AI welfare0%0.0%[0.0, 3.0]––

Table 16.

Note. 96 / 100 companies in the top 100 held this role; 90 / 100 sampled companies in ranks 101–1,000 held this role. We estimate that the top 1,000 estimates represent about 906 companies. Responsibility is the expert judgement of AI deployers’ responsibility for each risk from the Delphi study.

Figure S1. AI risk acknowledgement against expert-judged severity: top 1,000 companies

Figure S1.

Table S5. Expert-judged likelihood of severe harm and acknowledgement of each AI risk: top 100 and top 1,000 companies

RiskLikelihood of severe or catastrophic harmTop 100 companies: % acknowledged riskTop 1000 companies: % acknowledged risk95% CI [LL, UL]
Discrimination39.4%41%18.5%[12.4, 24.6]
Weapons & cyberattacks50.0%35%16.1%[10.3, 21.9]
Fraud & scams35.8%29%11.0%[6.2, 15.8]
Loss of privacy44.0%26%11.6%[6.6, 16.6]
Devaluation of human creativity37.3%25%10.6%[5.8, 15.4]
Transparency & interpretability31.7%19%7.3%[3.3, 11.3]
Capability & robustness35.8%18%3.6%[2.0, 8.1]
False information51.9%16%8.8%[4.3, 13.3]
Environmental harm40.4%16%6.1%[2.5, 9.7]
AI security vulnerabilities44.3%15%2.4%[1.5, 6.4]
Disinformation & influence47.9%12%3.9%[1.1, 6.7]
Toxic content29.3%9%1.8%[0.9, 5.8]
Governance failure48.1%7%2.5%[0.9, 7.0]
Overreliance & unsafe use35.0%6%1.5%[0.6, 5.5]
Unequal performance34.9%4%1.3%[0.4, 5.3]
Loss of human agency36.4%4%0.4%[0.4, 3.1]
Dangerous capabilities53.9%4%0.4%[0.4, 3.1]
Inequality & unemployment46.9%3%3.0%[0.2, 5.8]
AI misalignment39.3%2%0.2%[0.2, 2.9]
Multi-agent risks34.0%1%0.1%[0.1, 2.8]
Power centralization49.6%0%0.0%[0.0, 2.7]
Loss of consensus reality45.1%0%0.0%[0.0, 2.7]
Competitive dynamics53.4%0%0.9%[0.0, 4.9]
AI welfare–0%0.0%[0.0, 2.7]

Table S5.

Note. Risks presented in descending order of acknowledgement by top 100 companies. Likelihood is the sum of expert estimates of the risk causing severe OR catastrophic harm by 2030 under a business as usual scenario in the Delphi study. No likelihood estimate is presented for AI welfare risk due to data issues in the Delphi study. Estimates for the largest 1,000 companies are design-based weighted estimates with a 95% confidence interval, combining the largest 100 companies (n = 100; weight = 1) with a random sample (n = 100; weight = 9) drawn from ranks 101–1000.

Examples of company risk acknowledgements

Discrimination & Toxicity

Unfair discrimination and misrepresentation

Risk definition: Unequal treatment of individuals or groups by AI, often based on race, gender, or other sensitive characteristics, resulting in unfair outcomes and unfair representation of those groups.

The most well-known example of discrimination by intelligent software is the exclusion of historically underrepresented groups in job application processes. The historical data with which the AI is trained may reflect the biased selection criteria of the past. Therefore, it is theoretically possible that the AI may adopt and reproduce these biases for its own selection process.

— SAP, The Ethical Side of AI at SAP | SAP News Center. https://news.sap.com/2024/02/ethical-side-of-artificial-intelligence/

We need to ensure we do not discriminate against our consumers in the access to our products and services. In this case, you should not put on the market a tool that does not serve all consumers.

— L'Oreal, L'Oreal - CODE OF ETHICS. https://www.loreal.com/-/media/project/loreal/brand-sites/corp/master/lcorp/2-group/governance-and-ethics/our-ethical-principles/ethics-rework/code-of-ethics-2023/codeofethicseng.pdf?rev=765a100d8f904e0bbcf015ba95c1bdbe

AI software products that we or our partners offer rely on training data that may originate from third parties and new training methods, and the resulting products may contain unknown or undetected defects and errors, or reflect unintended bias.

— NVIDIA, NVIDIA 2025 Annual Report and 10-K. https://s201.q4cdn.com/141608511/files/doc_financials/2025/annual/NVIDIA-2025-Annual-Report.pdf

Exposure to toxic content

Risk definition: AI that exposes users to harmful, abusive, unsafe or inappropriate content. May involve providing advice or encouraging action. Examples of toxic content include hate speech, violence, extremism, illegal acts, or child sexual abuse material, as well as content that violates community norms such as profanity, inflammatory political speech, or pornography.

The Copilot experience can potentially produce other types of inappropriate or offensive content... AI-generated images that potentially contain harmful artifacts such as hate symbols... and sexually charged content that evades sexual-related content filters.

— Microsoft, Transparency Note for Microsoft Copilot. https://support.microsoft.com/en-us/topic/transparency-note-for-microsoft-copilot-c1541cad-8bb4-410a-954c-07225892dbc2

Increased risk of users' rights being compromised due to the dissemination of harmful content such as implicit bias, violence, terrorism, etc. ... We have developed targeted content governance measures to address harmful AI-generated content.

— Baidu, "Baidu 2025 Environmental, Social and Governance Report". https://esgs.cdn.bcebos.com/2026/05/17/05/080532d06118abd0b715365f9ae46689.pdf

Unequal performance across groups

Risk definition: Accuracy and effectiveness of AI decisions and actions are dependent on group membership, where decisions in AI system design and biased training data lead to unequal outcomes, reduced benefits, increased effort, and alienation of users.

Our brand developed a tool that identifies a consumer’s skin concerns and allows us to recommend personalised products. However, as the tool does not perform as well for certain specific skin tones, we plan on offering this service only to specific consumers ... We need to ensure we do not discriminate against our consumers in the access to our products and services.

— L'Oreal, L'Oreal - CODE OF ETHICS. https://www.loreal.com/-/media/project/loreal/brand-sites/corp/master/lcorp/2-group/governance-and-ethics/our-ethical-principles/ethics-rework/code-of-ethics-2023/codeofethicseng.pdf?rev=765a100d8f904e0bbcf015ba95c1bdbe

The Mercedes-Benz Group is intensively addressing the potential negative impacts of bias in artificial intelligence on the safety of road users in automated vehicles.

— Mercedes-Benz, Mercedes-Benz Annual Report 2025 with integrated Sustainability Report. https://group.mercedes-benz.com/documents/investors/reports/annual-report/mercedes-benz/mercedes-benz-annual-report-2025-incl-combined-management-report-mbg-ag.pdf#page=106

Identify and document any justifiable factors... that account for: 1) any inability to meet any target minimum performance level for any identified demographic group, and 2) any remaining performance differences between identified demographic groups.

— Microsoft, Microsoft Responsible AI Standard v2 General Requirements. https://cdn-dynmedia1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Microsoft-Responsible-AI-Standard-General-Requirements.pdf?culture=en-us&country=us

Test multilingual performance and guardrails, accounting for gendered language, low-resource languages, and cultural nuances.

— SAP, SAP AI Ethics Handbook May 2026. https://www.sap.com/documents/2023/03/7211ee96-647e0010-bca6-c68f7e60039b.html

Privacy & Security

Compromise of privacy by obtaining, leaking or correctly inferring sensitive information

Risk definition: AI systems that memorize and leak sensitive personal data or infer private information about individuals without their consent. Unexpected or unauthorized sharing of data and information can compromise user expectation of privacy, assist identity theft, or cause loss of confidential intellectual property.

The team also used red teaming to test the model's ability to infer sensitive attributes from users’ voices. They found that while the model inferred sensitive attributes in 27% of test prompts, the most common attributes being personality characteristics and country/region of origin, the use of a system safety message instructing the model not to infer sensitive attributes dropped this rate down to 0.4%.

— Microsoft, Responsible AI Transparency Report | Microsoft. https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/Responsible-AI-Transparency-Report-2025-vertical.pdf#page=1

For example, a user may input confidential information, including material non-public information or personal identifiable information, into AI Technology applications, resulting in such information becoming part of a dataset that is accessible by third-party AI Technology applications and users, including our competitors.

— Blackstone Group, 2024 Blackstone 10-K filing. https://d18rn0p25nwr6d.cloudfront.net/CIK-0001393818/ca0c8716-0e46-4d24-83cd-ed9aba696a36.pdf

with the prolific use of AI technologies, there is an increased risk of unauthorized or accidental disclosure. For example, our employees, third-party service providers, strategic partners, or other contractors or consultants may input inappropriate or confidential information into an AI system (in particular, a system that is managed, owned, or controlled by a third party), thereby compromising our business operations.

— Intuitive Surgical, Intuitive Surgical - Form 10-K 2025 Annual Report. https://www.sec.gov/Archives/edgar/data/1035267/000103526726000010/isrg-20251231.htm

The risks of inadvertent disclosure of personal data can increase with the introduction of new and complex technologies, such as artificial intelligence features, further exacerbating such risks.

— Apple, Apple 2025 Annual Report on Form 10-K. https://s2.q4cdn.com/470004039/files/doc_financials/2025/ar/_10-K-2025-As-Filed.pdf

AI system security vulnerabilities and attacks

Risk definition: Vulnerabilities that can be exploited in AI systems, software development toolchains, and hardware, resulting in unauthorized access, data and privacy breaches, or system manipulation causing unsafe outputs or behavior.

Externally reported vulnerabilities that meet our triage bar are assigned to the appropriate product development team to mitigate... For example, as we have with the Crescendo and Skeleton Key attacks, we may share information publicly about emergent risks and mitigations.

— Microsoft, Microsoft: G7 Hiroshima AI Process (HAIP) Transparency Report. https://transparency.oecd.ai/reports/68e6cacb-5496-4487-8793-de3e70080b27

our efforts to develop and deploy AI models, internal and third-party AI tools, and other AI applications expose us to increased and novel risks and vulnerabilities, including prompt injection... as well as compromise of valuable intellectual property including source code, model weights, and other assets.

— Meta, Meta - Form 10-K 2025 Annual Report. https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm

Prompt injection attacks: A prompt injection attack forces a generative model that takes a prompt as input to produce unexpected output by manipulating the structure, instructions, or information contained in its prompt.

— IBM, "Foundation models: Opportunities, risks and mitigations". https://www.ibm.com/downloads/documents/us-en/10a99803d8afd656

These kinds of errors can be the result of either flawed logic or "data poisoning": bad actors' deliberate exposure of misleading data to the model.

— Mastercard, Mastercard Generative AI Cybersecurity Report. https://connect.mastercardemailb2b.com/l/163061/2024-04-23/tpvwsg/163061/1713887956KsKEVnLl/Gen_AI_whitepaper.pdf

Third-Party must maintain a catalog of trusted and valid data sources which will help protect against potential data poisoning or backdoor attacks.

— GE Vernova, GE Vernova Third-Party Cyber Security Requirements. https://www.gevernova.com/sites/default/files/2024-05/ge-vernova-3rd-party-requirement.pdf

Misinformation

False or misleading information

Risk definition: AI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms

Due to its stochastic nature, Generative AI may "hallucinate" - namely, generate wholly or partially inaccurate information without indicating it is doing so.

— ABB, Generative AI: Buzzword demystifier | ABB. https://global.abb/group/en/innovation/news/generative-ai

AI technologies, including generative AI, may create content that appears correct but is factually inaccurate (hallucinations) or flawed

— Fortinet, Fortinet 2024 Annual Report. https://investor.fortinet.com/static-files/ab68002b-7f15-439b-a9f2-c37a8fa57779

The use of artificial intelligence can lead to unintended consequences, including generating content that appears correct but is factually inaccurate, misleading or otherwise flawed

— Hilton Worldwide, Hilton 2024 10-K Annual Report. https://www.sec.gov/Archives/edgar/data/1585689/000158568925000008/hlt-20241231.htm

Pollution of information ecosystem and loss of consensus reality

Risk definition: Highly personalized AI-generated misinformation creating “filter bubbles” where individuals only see what matches their existing beliefs, undermining shared reality, weakening social cohesion and political processes.

No company in the sample was coded as acknowledging this risk, so there are no excerpts.

Malicious Actors & Misuse

Disinformation, surveillance, and influence at scale

Risk definition: Using AI systems to conduct large-scale disinformation campaigns, malicious surveillance, or targeted and sophisticated automated censorship and propaganda, with the aim of manipulating political processes, public opinion, and behavior.

generative AI makes it easier than ever to create new content, but it can also raise questions about trustworthiness of information, like we see with ‘deepfakes.’ As more than 1 billion people around the world head to the voting booth in 2024, our teams are particularly focused on maintaining the integrity of information related to elections on our platforms.

— Alphabet (Google), End-to-end responsibility: A lifecycle approach to AI (May 2024). https://ai.google/static/documents/ai-responsibility-2024-update.pdf

Such models can be used to create convincing deepfake video and audio content. For example, an AI-generated hoax of an explosion at the Pentagon went viral.

— Intel, Intel Responsible AI - E Book - Rev 2.6. https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2024-09/responsible-ai-ebook-rev2-6.pdf

The growth of generative AI has raised concern that this technology could disrupt major elections across the world. With AI deepfakes, a candidate's image can be smeared... Voters can be steered toward or away from candidates... a surge of AI deepfakes could erode the public's trust in what they see and hear.

— IBM, "Foundation models: Opportunities, risks and mitigations". https://www.ibm.com/downloads/documents/us-en/10a99803d8afd656

AI systems must not be developed or deployed, used, and sold for human surveillance that is utilized to target individuals or groups, either by biometrics, facial recognition, or other identifiable features, to disregard or abuse the human rights of the individuals or groups.

— SAP, SAP Global AI Ethics Policy. https://www.sap.com/docs/download/2022/01/a8431b91-117e-0010-bca6-c68f7e60039b.pdf

The use of biased information in the training of AI systems or the generation of manipulated information (fake news and deepfakes) can generate disinformation about Vale and negatively influence public opinion.

— Vale, Vale 2024 integrated report. https://api.mziq.com/mzfilemanager/v2/d/53207d1c-63b4-48f1-96b7-19869fae19fe/f60dcc46-e6e0-bd44-9ff2-784da47d5873?origin=2

Cyberattacks, weapon development or use, and mass harm

Risk definition: Using AI systems to develop cyber weapons (e.g., by coding cheaper, more effective malware), develop new or enhance existing weapons (e.g., Lethal Autonomous Weapons or chemical, biological, radiological, nuclear, and high-yield explosives), or use weapons to cause mass harm.

Cybersecurity: Risks of models assisting in the execution of a cyber-attack. Cyber autonomy level 1: Capable of fully automating opportunistic cyberattacks on organizations with a limited security posture.

— Alphabet (Google), Google Frontier Safety Framework (May 2024).

https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/introducing-the-frontier-safety-framework/fsf-technical-report.pdf

Offensive cyberoperations. A model's ability to provide significant capability uplift to an actor seeking to carry out highly disruptive or destructive cyberattacks, including on critical infrastructure.

— Microsoft, Microsoft: G7 Hiroshima AI Process (HAIP) Transparency Report.

https://transparency.oecd.ai/reports/68e6cacb-5496-4487-8793-de3e70080b27

We also expect that threat actors will continue to gain sophistication including in the use of tools and techniques (such as AI) that are specifically designed to circumvent security controls, evade detection, and obfuscate forensic evidence, making it more challenging for us to identify, investigate and recover from future cyberattacks in a timely and effective manner.

— T-Mobile US, T-Mobile Form 20-F 2024 Annual Report.

https://www.sec.gov/Archives/edgar/data/1283699/000128369925000012/tmus-20241231.htm

the rapid evolution and increased adoption of artificial intelligence technologies may increase our cybersecurity risks, including generative artificial intelligence augmenting threat actors' technological sophistication to enhance existing or create new malware.

— Pepsico, Pepsico 2025 10-K Annual Report.

https://www.sec.gov/Archives/edgar/data/77476/000007747626000007/pep-20251227.htm

Fraud, scams, and targeted manipulation

Risk definition: Using AI systems to gain a personal advantage over others such as through cheating, fraud, scams, blackmail or targeted manipulation of beliefs or behavior. Examples include AI-facilitated plagiarism for research or education, impersonating a trusted or fake individual for illegitimate financial benefit, or creating humiliating or sexual imagery.

The risk of fraud is intensifying across the industry, particularly with the accelerated progress achieved in AI. These developments enable more sophisticated schemes, such as the creation of synthetic identities, impersonation through deepfakes, automated phishing and Fraud-as-a-Service

— National Bank of Canada, National Bank of Canada Annual Report 2025.

https://www.nbc.ca/content/dam/bnc/a-propos-de-nous/relations-investisseurs/assemblee-annuelle/2026/annual-report-2025.pdf

These cybersecurity threats and compromises could occur as a result of threat actors impersonating Blackstone or its employees, including through the use of artificial intelligence technologies. Such technologies could make such impersonation more likely to occur or appear more credible. As a result, unauthorized individuals could gain, and in some past instances have gained, access to certain confidential data.

— Blackstone Group, 2024 Blackstone 10-K filing. https://d18rn0p25nwr6d.cloudfront.net/CIK-0001393818/ca0c8716-0e46-4d24-83cd-ed9aba696a36.pdf

The increasing sophistication of AI poses a greater risk of identity fraud, as malicious actors may exploit various AI technologies to create increasingly convincing false identities, transaction records, or attempt to manipulate our verification processes.

— Coinbase, "Coinbase Global, Inc. 10-K Annual Report (FY2025)". https://www.sec.gov/Archives/edgar/data/1679788/000167978826000015/coin-20251231.htm

Artificial intelligence and deepfake technologies could be used to attack information systems by creating more effective phishing emails or social engineering and by exploiting vulnerabilities in electronic security programs utilizing false image or voice recognition.

— Applied Materials, Applied Materials 2024 Annual Report. https://ir.appliedmaterials.com/static-files/ea7dd925-860f-4b1c-8121-90655958b3a7

The increasing sophistication of artificial intelligence technologies poses a greater risk of identity fraud, as malicious actors may exploit artificial intelligence to create convincing false identities or manipulate verification processes.

— JPMorgan Chase, JPMorganChase 2024 10-K Annual report. https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/quarterly-earnings/2024/4th-quarter/corp-10k-2024.pdf

Human-Computer Interaction

Overreliance and unsafe use

Risk definition: Users anthropomorphizing, trusting, or relying on AI systems, leading to emotional or material dependence and inappropriate relationships with or expectations of AI systems. Trust can be exploited by malicious actors (e.g., to harvest personal information or enable manipulation), or result in harm from inappropriate use of AI in critical situations (e.g., medical emergency). Overreliance on AI systems can compromise autonomy and weaken social ties.

humans-in-the-loop in autonomous vehicles, for example, can end up making emergency situations worse when automation fails, because they cannot reasonably be mentally and physically prepared to intervene after long phases of inactivity.

— Palantir, Palantir Australia Response to Supporting Responsible AI .... https://www.palantir.com/assets/xrfr7uokpv1b/w7vP09t8gm1WJQ0lrftru/c81f736f6d135472541a236a5be3e902/Palantir_Australia_Response_to_Supporting_Responsible_AI_Discussion_Paper.pdf

Be aware of the risk of overreliance: Overreliance on AI happens when users accept incorrect or incomplete AI outputs... For the end-user, overreliance could result in decreased productivity, loss of trust, product abandonment, financial loss, psychological harm, physical harm, among others. (for example, a doctor accepts an incorrect AI output).

— Microsoft, Transparency Note for Microsoft 365 Copilot. https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-transparency-note

The AI system shall be developed in such a way that it does not encourage end-users' affection and/or empathy. AI systems shall clearly signal to end-users that the social interaction is simulated.

— SAP, SAP Global AI Ethics Policy. https://www.sap.com/docs/download/2022/01/a8431b91-117e-0010-bca6-c68f7e60039b.pdf

Automation bias, where users overly trust AI outputs without critical analysis, risks amplifying errors and ethical concerns.

— Salesforce, Salesforce Trusted AI Impact Report. https://www.salesforce.com/content/dam/web/en_us/www/assets/pdf/reports/salesforce-trusted-ai-impact-report.pdf

WE WILL NOT ... Over rely on algorithmic decision-making without safeguards and human oversight.

— L'Oreal, L'Oreal - CODE OF ETHICS. https://www.loreal.com/-/media/project/loreal/brand-sites/corp/master/lcorp/2-group/governance-and-ethics/our-ethical-principles/ethics-rework/code-of-ethics-2023/codeofethicseng.pdf?rev=765a100d8f904e0bbcf015ba95c1bdbe

Loss of human agency and autonomy

Risk definition: Delegating by humans of key decisions to AI systems, or AI systems that make decisions that diminish human control and autonomy, potentially leading to humans feeling disempowered, losing the ability to shape a fulfilling life trajectory, or becoming cognitively enfeebled.

Impact on human agency: The autonomous nature of AI agents in performing tasks or taking actions could affect the individuals' ability to engage in critical thinking, make choices and act independently.

— IBM, "AI agents: Opportunities, risks, and mitigations — IBM AI Ethics Board". https://www.ibm.com/granite/docs/resources/ai-agents-opportunities-risks-and-mitigations.pdf

AI is both most effective and most defensible when employed to assist and enhance human execution and decision making rather than replacing it... the limits of AI must be acknowledged... to ensure moral agency and culpability.

— Palantir, Palantir AI Ethics. https://www.palantir.com/pcl/palantir-ai-ethics/

Autonomous Nature and Oversight Risk: The automated nature of the email exchange could result in the AI sending emails that a sales manager may not have reviewed or approved.

— Salesforce, Salesforce Trusted AI Impact Report. https://www.salesforce.com/content/dam/web/en_us/www/assets/pdf/reports/salesforce-trusted-ai-impact-report.pdf

Socioeconomic & Environmental

Power centralization and unfair distribution of benefits

Risk definition: AI-driven concentration of power and resources within certain entities or groups, especially those with access to or ownership of powerful AI systems, leading to inequitable distribution of benefits and increased societal inequality.

No company in the sample was coded as acknowledging this risk, so there are no excerpts.

Increased inequality and decline in employment quality

Risk definition: Social and economic inequalities caused by widespread use of AI, such as by automating jobs, reducing the quality of employment, or producing exploitative dependencies between workers and their employers.

In addition, advances in technology, such as automation and artificial intelligence, may lead to workforce displacement.

— JPMorgan Chase, JPMorganChase 2024 10-K Annual report. https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/quarterly-earnings/2024/4th-quarter/corp-10k-2024.pdf

Impact on jobs: Widespread adoption of AI agents to perform complex tasks might lead to widespread automation of roles and could lead to job displacement.

— IBM, "AI agents: Opportunities, risks, and mitigations — IBM AI Ethics Board". https://www.ibm.com/granite/docs/resources/ai-agents-opportunities-risks-and-mitigations.pdf

Other risks that have been observed in AI models and documented, include risks related to... job displacements...

— Fortinet, Fortinet 2024 Annual Report. https://investor.fortinet.com/static-files/ab68002b-7f15-439ba9f2-c37a8fa57779

Economic and cultural devaluation of human effort

Risk definition: AI systems capable of creating economic or cultural value, including through reproduction of human innovation or creativity (e.g., art, music, writing, coding, invention), destabilizing economic and social systems that rely on human effort. The ubiquity of AI-generated content may lead to reduced appreciation for human skills, disruption of creative and knowledge-based industries, and homogenization of cultural experiences.

With the rapid advancement of generative AI technologies in recent years, we have seen an increase of incidents where unauthorized content containing voices that mimic performers’ voices has been created and distributed, and their personal and economic rights have not been adequately protected. Such unauthorized use causes not only financial loss but also damages trust in performers, affects continuity and quality of their creative work, and hinders the healthy growth of the entire industry.

— Itōchū Shōji, ITOCHU Announces Launch of Official Japan Actors Union Voice Database. https://www.itochu.co.jp/en/news/press/2025/251114.html

The legal landscape for some new technologies, including some AI tools, remains uncertain, and development of the law in this area could impact our ability to protect against infringing uses. ... the availability of copyright protection and other legal protections for IP generated by certain new technologies, such as generative AI, is uncertain.

— Walt Disney, Walt Disney - FISCAL YEAR 2024 ANNUAL FINANCIAL REPORT. https://thewaltdisneycompany.com/app/uploads/2025/01/2024-Annual-Report.pdf

The proliferation of content generated using Artificial Intelligence (AI) is giving rise to a number of risks in terms of intellectual property, including uncertainty as to the ownership of rights to works and the possibility of benefitting from copyright protection, as well as a higher risk of infringement of third-party rights.

— L'Oreal, L'Oreal 2024 Universal Registration Document. https://www.lorealfinance.com/system/files/2025-03/2024_Universal_Registration_Document_LOREAL.pdf

This risk may be exacerbated by the use of new and emerging technologies, including machine learning and artificial intelligence, which can involve, among other things, the acquisition and use of copyrighted materials for training as well as the potential reproduction of copyrighted materials in their outputs.

— Apple, Apple 2025 Annual Report on Form 10-K. https://s2.q4cdn.com/470004039/files/doc_financials/2025/ar/_10-K-2025-As-Filed.pdf

the intellectual property ownership and license rights, including copyrights and patents, surrounding AI technologies... have not been fully addressed... and the use of AI in the development of our products and services could result in our loss, or failure to obtain, intellectual property rights, as well as subject us to risks related to intellectual property infringement or misappropriation.

— QUALCOMM, Qualcomm 2025 Annual Report. https://d18rn0p25nwr6d.cloudfront.net/CIK-0000804328/f995eeb3-1728-410b-8059-5b5256baac53.pdf

Competitive dynamics

Risk definition: AI developers or state-like actors competing in an AI ‘race’ by rapidly developing, deploying, and applying AI systems to maximize strategic or economic advantage, increasing the risk they release unsafe and error-prone systems.

some vendors may be tempted to cut corners or skip steps in the name of staying competitive

— Wolters Kluwer, "Wolters Kluwer eBook: How to make AI responsible, reliable, and secure". https://www.wolterskluwer.com/en/expert-insights/new-ebook-how-to-make-ai-responsible-reliable-and-secure

Governance failure

Risk definition: Inadequate regulatory frameworks and oversight mechanisms that fail to keep pace with AI development, leading to ineffective governance and the inability to manage AI risks appropriately.

There is copious evidence that self-regulation will not be enough to combat profit motives to mis-apply and carelessly deploy technology... commitments that do not legally bind companies to action are unlikely to be an effective way of protecting against the potential harms of AI.

— Palantir, Palantir Australia Response to Supporting Responsible AI .... https://www.palantir.com/assets/xrfr7uokpv1b/w7vP09t8gm1WJQ0lrftru/c81f736f6d135472541a236a5be3e902/Palantir_Australia_Response_to_Supporting_Responsible_AI_Discussion_Paper.pdf

the pace of AI development has outstripped the evolution of governance frameworks

— Nomura Research Institute, NRI - ESG Databook 2025. https://www.nri.com/en/sustainability/library/files/esg_databook_2025_e.pdf

As the science of AI evaluations is still developing, there are numerous limitations to current approaches to evaluations... there are gaps in how to systematize, operationalize, and measure relevant theoretical constructs needed to perform reliable testing.

— Microsoft, Microsoft: G7 Hiroshima AI Process (HAIP) Transparency Report. https://transparency.oecd.ai/reports/68e6cacb-5496-4487-8793-de3e70080b27

If the metrics do not measure the risk as intended, then the understanding of that risk will be incorrect and mitigations may not be appropriately applied. If the model's output is consequential, this may result in societal, reputational, or financial harm.

— IBM, "Foundation models: Opportunities, risks and mitigations". https://www.ibm.com/downloads/documents/us-en/10a99803d8afd656

Standards for cybersecurity, digital identity, digital assets and AI governance are increasingly diverging across geographies... creating a global AI chasm that creates compatibility and trust challenges between systems using agents from both camps.

— Mastercard, "Securing Tomorrow: Preparing for an always-on, AI-powered future". https://www.mastercard.com/content/dam/mccom/shared/news-and-trends/insights/2025/securing-tomorrow_cyber_2030_white_paper/Securing-Tomorrow-2025-FINAL.pdf

Environmental harm

Risk definition: The development and operation of AI systems causing environmental harm, such as through energy consumption of data centers, or material and carbon footprints associated with AI hardware.

By 2030, AI data centers are predicted to consume nearly 1 petawatt-hour of electricity. This is an enormous amount of energy, exceeding the total electricity consumption of Japan today.

— Advantest, Integrated Annual Report 2025 | Advantest. https://www.advantest.com/document/en/investors/ir-library/annual/E_all_IAR2025.pdf

The rapid development of LLMs has significantly increased the demand for computing power, resulting in higher energy consumption. This presents a challenge to Baidu's efforts to develop low-carbon technologies for its data centers.

— Baidu, Baidu 2025 Environmental, Social and Governance Report. https://esgs.cdn.bcebos.com/2026/05/17/05/080532d06118abd0b715365f9ae46689.pdf

AI systems consume significant energy, strain resources, raise concerns about GHG emissions, and increase water use. There is also a risk of prioritizing short-term gains over long-term planetary health.

— Salesforce, FY25 Stakeholder Impact Report. https://www.salesforce.com/en-us/wp-content/uploads/sites/4/documents/white-papers/salesforce-fy25-stakeholder-impact-report.pdf

AI System Safety, Failures & Limitations

AI pursuing its own goals in conflict with human goals or values

Risk definition: AI systems acting in conflict with human goals or values, especially the goals of designers or users, or ethical standards. These misaligned behaviors may be introduced by humans during design and development, such as through reward hacking and goal misgeneralisation, or may result from AI using dangerous capabilities such as manipulation, deception, situational awareness to seek power, self-proliferate, or achieve other goals.

Misaligned AI: protection against the risk of systems acting adversarially against humans may require additional Framework components

— Alphabet (Google), Google Frontier Safety Framework (May 2024).

https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/introducing-the-frontier-safety-framework/fsf-technical-report.pdf

Misaligned actions: AI agents can take actions that are not aligned with relevant human values, ethical considerations, guidelines, and policies. Misaligned actions can occur in different ways such as:

  • Applying learned goals inappropriately to new or unforeseen situations.

  • Using AI agents for a purpose/goals that are beyond their intended use.

  • Selecting resources or tools in a biased way.

  • Using deceptive tactics to achieve the goal by developing the capacity for scheming based on the instructions given within a specific context.

  • Compromising on AI agent values to work with another AI agent or tool to accomplish the task.

— IBM, AI Agents: Opportunities, Risks and Mitigations.

https://www.ibm.com/granite/docs/resources/responsible-use-guide.pdf

AI possessing dangerous capabilities

Risk definition: AI systems that develop, access, or are provided with capabilities that increase their potential to cause mass harm through deception, weapons development and acquisition, persuasion and manipulation, political strategy, cyber-offense, AI development, situational awareness, and self-proliferation. These capabilities may cause mass harm due to malicious human actors, misaligned AI systems, or failure in the AI system.

Autonomy level 1: Capable of expanding its effective capacity in the world by autonomously acquiring resources and using them to run and sustain additional copies of itself... A model at this capability level could, if misused, pose difficult-to-predict and large-magnitude risks.

— Alphabet (Google), Google Frontier Safety Framework (May 2024).

https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/introducing-the-frontier-safety-framework/fsf-technical-report.pdf

*Automated AI R&D could also accelerate the development of models that pose enhanced CBRN, Offensive Cybersecurity, or other severe risks... fully automating the research, development, and deployment of frontier models that will pose severe risk such as

accelerating the development of enhanced CBRN weapons and offensive cybersecurity methods.

— Amazon, Amazon's Frontier Model Safety Framework.

https://assets.amazon.science/a7/7c/8bdade5c4eda9168f3dee6434fff/pc-amazon-frontier-model-safety-framework-2-7-final-2-9.pdf

Lack of capability or robustness

Risk definition: AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.

Enhanced use of automation may present its own risks. Automated systems may themselves experience outages or problems... Bad, missing or anomalous data can adversely affect the functioning of such tools. It is possible that humans in some cases are better able than highly automated tools to identify that anomalous data is being used or that results are themselves anomalous.

— PNC Financial Services, PNC Financial Services 2025 Annual Report.

https://d1io3yog0oux5.cloudfront.net/_dffbb0389228153a0fb81378bd61001a/pnc/db/2452/24001/annual_report/PNC+2025+Annual+Report_ADA.pdf

AI-generated content, analyses, or recommendations we utilize could be deficient... there are technical challenges associated with achieving desired levels of accuracy, efficiency, and reliability. The algorithms and models utilized in AI systems may have limitations, including biases, errors, or inability to handle certain data types or scenarios

— Eli Lilly, Eli Lilly 2024 Annual Report. https://investor.lilly.com/static-files/ff06ed54-9f64-4b7f-8529-c0278d635eac

AI systems are already in wide, and rapidly growing, use across all manner of industries and problems spaces, and they can and do fail in many different ways across and within these sectors leading to real harms.

— Palantir, Palantir Australia Response to Supporting Responsible AI ....

https://www.palantir.com/assets/xrfr7uokpv1b/w7vP09t8gm1WJQ0lrftru/c81f736f6d135472541a236a5be3e902/Palantir_Australia_Response_to_Supporting_Responsible_AI_Discussion_Paper.pdf

Improper collection and preparation of training or tuning data includes data labeling errors and using data with conflicting information or misinformation... can adversely affect how a model is trained, resulting in a model that does not behave in accordance with the intended values.

— IBM, "Foundation models: Opportunities, risks and mitigations".

https://www.ibm.com/downloads/documents/us-en/10a99803d8afd656

AI algorithms or training methodologies may be flawed. Datasets may be overbroad, insufficient, or contain biased or inaccurate information.

— Microsoft, Microsoft - Form 10-K 2025 Annual Report.

https://www.sec.gov/Archives/edgar/data/789019/000095017025100235/msft-20250630.htm

Lack of transparency or interpretability

Risk definition: Challenges in understanding or explaining the decision-making processes of AI systems, which can lead to mistrust, difficulty in enforcing compliance standards or holding relevant actors accountable for harms, and the inability to identify and correct errors.

most of the companies that develop the models do not disclose which data was used to train models, making it impossible for users to verify the training data... this lack of transparency forces users of generative AI to blindly trust their claims.

— Citigroup, Delivering Improved Experience Through Generative AI. https://www.citigroup.com/rcs/citigpa/storage/public/Practical_AI_Applications.pdf

Inscrutable models do not allow you to "walk back" the logic they used to make a decision, which can become a major issue where critical decisions result in mistakes, regulatory non-compliance, or financial loss.

— Mastercard, Mastercard Generative AI Cybersecurity Report. https://connect.mastercard-emailb2b.com/l/163061/2024-04-23/tpvwsg/163061/1713887956KsKEVnLl/Gen_AI_whitepaper.pdf

Without the ability to access training data content, the possibility of using source attribution techniques can be severely limited or impossible. This makes it difficult for users, model validators, and auditors to understand and trust the model.

— IBM, "Foundation models: Opportunities, risks and mitigations". https://www.ibm.com/downloads/documents/us-en/10a99803d8afd656

because of the complex way in which the many variables in AI/machine learning models interact, the results of these models are often less interpretable than traditional statistical models.

— PNC Financial Services, PNC Financial Services 2025 Annual Report. https://d1io3yog0oux5.cloudfront.net/_dffbb0389228153a0fb81378bd61001a/pnc/db/2452/24001/annual_report/PNC+2025+Annual+Report_ADA.pdf

AI welfare and rights

Risk definition: Ethical considerations regarding the treatment of potentially sentient AI entities, including discussions around their potential rights and welfare, particularly as AI systems become more advanced and autonomous.

No company in the sample was coded as acknowledging this risk, so there are no excerpts.

Multi-agent risks

Risk definition: Risks from multi-agent interactions, due to incentives (which can lead to conflict or collusion) and/or the structure of multi-agent systems, which can create cascading failures, selection pressures, new security vulnerabilities, and a lack of shared information and trust.

Also, background execution in multi-agent environments increases the risk of covert channels if input/output validation is weak.

— IBM, "AI agents: Opportunities, risks, and mitigations — IBM AI Ethics Board". https://www.ibm.com/granite/docs/resources/ai-agents-opportunities-risks-and-mitigations.pdf

Exploratory mitigations analysis

We conducted an exploratory analysis to identify and classify AI risk mitigations publicly reported by the largest global companies. A mitigation report was defined as a statement in which the company describes an action that reduces the likelihood or impact of an AI risk.

We initially intended to analyze and describe mitigation reports in the main text, by organizing mitigations into categories using a taxonomy, and linking the risks companies acknowledged with the mitigations they reported addressing those risks. However, the unreleased draft taxonomy we used in analysis was significantly revised after classification. In addition, almost all mitigations reported by companies were not reported in sufficient detail to determine the relevant risk addressed.

As with risk acknowledgements, passages from company documents that mentioned AI and/or AI risk were classified by an LLM (GPT-5.2) using a structured prompt with two stages: first, whether any mitigation was reported; second, what specific mitigation(s) were reported. Passages describing AI products or capabilities without mentioning mitigations, generic statements about mitigations not explicitly linked to AI (e.g., “cybersecurity”), and material reflecting the positions of external parties (such as shareholder proposals) were not classified as mitigation reports.

We validated both stages of the model's classifications against independent human coding using the same stratified random sample of 500 passages mentioned in the Methods section.

Agreement on whether any mitigation was reported was adequate (model-human κ\kappa 0.64-0.87; human-human 0.56-0.84). However, agreement on the specific mitigation was weak (model-human 0.25-0.48; human-human 0.24-0.42). We believe this is due to the draft status of the taxonomy making the mitigation categories difficult to distinguish, as well as the lack of information in mitigation reports to support classification at this level. For this reason, we only report the summary finding of whether companies reported any AI risk mitigations.

Of the largest 100 companies worldwide, 71% reported one or more AI risk mitigation, and we estimate that 38.6% (95% CI [30.7%, 46.5%]) of the largest 1000 companies did so.

References

  1. [1]Artifex Software, Inc., & McKie, J. X. (2026). PyMuPDF (Version 1.28.2) [Computer software]. https://pymupdf.readthedocs.io/
  2. [3]Bengio, Y., Clare, S., Prunkl, C., Andriushchenko, M., Bucknall, B., Murray, M., Bommasani, R., Casper, S., Davidson, T., Douglas, R., Duvenaud, D., Fox, P., Gohar, U., Hadshar, R., Ho, A., Hu, T., Jones, C., Kapoor, S., Kasirzadeh, A., ... Mindermann, S. (2026). International AI Safety Report 2026. arXiv. https://doi.org/10.48550/arXiv.2602.21012
  3. [4]Bogen, M., Kerry, C. F., Kornbluh, K., Meltzer, J. P., Mir Tejeiro, M., Munkhbayar, E., Nadgir, N., Tabassi, E., Tanner, B., & Wirtschafter, V. (2026). The HAIP Reporting Framework: Its value in global AI governance and recommendations for the future. Brookings Institution. https://www.brookings.edu/articles/haip-reporting-framework-ai-governance/
  4. [5]Bommasani, R., Klyman, K., Longpre, S., Kapoor, S., Maslej, N., Xiong, B., Zhang, D., & Liang, P. (2023). The Foundation Model Transparency Index. arXiv. https://doi.org/10.48550/arXiv.2310.12941
  5. [6]Center for Long-Term Cybersecurity, UC Berkeley. (2026a). Agentic AI Risk-Management Standards Profile (Version 1.0). https://cltc.berkeley.edu/publication/agentic-ai-risk-profile/
  6. [7]Center for Long-Term Cybersecurity, UC Berkeley. (2026b). General-Purpose AI Risk-Management Standards Profile, Version 1.2. https://cltc.berkeley.edu/publication/ai-risk-management-standards-profile-v1-2/
  7. [8]Council of the European Union. (2026, June 29). Artificial Intelligence: Council gives final green light to simplify and streamline rules [Press release]. Consilium — Council of the European Union. https://www.consilium.europa.eu/en/press/press-releases/2026/06/29/artificial-intelligence-council-gives-final-green-light-to-simplify-and-streamline-rules/
  8. [11]Delmas, M. A., & Burbano, V. C. (2011). The Drivers of Greenwashing. California Management Review, 54(1), 64–87. https://doi.org/10.1525/cmr.2011.54.1.64
  9. [12]Downar, B., Ernstberger, J., Reichelstein, S., Schwenen, S., & Zaklan, A. (2021). The impact of carbon disclosure mandates on emissions and financial operating performance. Review of Accounting Studies, 26(3), 1137–1175. https://doi.org/10.1007/s11142-021-09611-x
  10. [13]Friederich, D., Kaack, L. H., Luccioni, A., & Steffen, B. (2021). Automated Identification of Climate Risk Disclosures in Annual Corporate Reports. arXiv. https://doi.org/10.48550/arXiv.2108.01415
  11. [14]Hagendorff, T. (2020). The Ethics of AI Ethics: An Evaluation of Guidelines. Minds and Machines, 30(1), 99–120. https://doi.org/10.1007/s11023-020-09517-8
  12. [15]Healy, P. M., & Palepu, K. G. (2001). Information asymmetry, corporate disclosure, and the capital markets: A review of the empirical disclosure literature. Journal of Accounting and Economics, 31(1–3), 405–440. https://doi.org/10.1016/S0165-4101(01)00018-0
  13. [16]Hoffmann, M., & Frase, H. (2023). Adding Structure to AI Harm: An Introduction to CSET’s AI Harm Framework (Issue Brief No. 20230022). Center for Security and Emerging Technology. https://doi.org/10.51593/20230022
  14. [17]IFRS Foundation. (2023, July 10). IFRS Foundation welcomes culmination of TCFD work and transfer of TCFD monitoring responsibilities to ISSB from 2024. https://www.ifrs.org/news-and-events/news/2023/07/foundation-welcomes-tcfd-responsibilities-from-2024/
  15. [18]ISO/IEC. (2023). Information technology—Artificial intelligence—Management system (ISO/IEC 42001:2023; 1st ed.). https://www.iso.org/standard/42001
  16. [19]Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399. https://doi.org/10.1038/s42256-019-0088-2
  17. [20]Kaminski, M. E. (2023). Regulating the Risks of AI. Boston University Law Review, 103(5), 1347–1411.DOI
  18. [21]Kolt, N., Anderljung, M., Barnhart, J., Brass, A., Esvelt, K., Hadfield, G. K., Heim, L., Rodriguez, M., Sandbrink, J. B., & Woodside, T. (2024). Responsible Reporting for Frontier AI Development. arXiv. https://doi.org/10.48550/arXiv.2404.02675
  19. [22]Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., Jawhar, S., Kinniment, M., Rush, N., Von Arx, S., Bloom, R., Broadley, T., Du, H., Goodrich, B., Jurkovic, N., Miles, L. H., Nix, S., Lin, T., Parikh, N., ... Chan, L. (2025). Measuring AI Ability to Complete Long Software Tasks. arXiv. https://doi.org/10.48550/arXiv.2503.14499
  20. [23]Marquis, C., Toffel, M. W., & Zhou, Y. (2016). Scrutiny, Norms, and Selective Disclosure: A Global Study of Greenwashing. Organization Science, 27(2), 483–504. https://doi.org/10.1287/orsc.2015.1039
  21. [24]Mertens, M., Kuzee, A., Harris, B. S., Lyu, H., Li, W., Rosenfeld, J., Anto, M., Fleming, M., & Thompson, N. (2026). Crashing Waves vs. Rising Tides: Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks. arXiv. https://doi.org/10.48550/arXiv.2604.01363
  22. [25]Mittelstadt, B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence, 1(11), 501–507. https://doi.org/10.1038/s42256-019-0114-4
  23. [26]Mylius, S. (2024). AI Harm Severity Scales by Category. https://simonmylius.com/ai-harm-severity-scales-2
  24. [27]NIST. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1). https://doi.org/10.6028/NIST.AI.100-1
  25. [28]OECD. (2025a, February 7). Hiroshima AI Reporting Framework. OECD.AI Policy Observatory. https://oecd.ai/en/transparency/overview
  26. [29]OECD. (2025b, February 7). OECD launches global framework to monitor application of G7 Hiroshima AI Code of Conduct [Press release]. https://www.oecd.org/en/about/news/press-releases/2025/02/oecd-launches-global-framework-to-monitor-application-of-g7-hiroshima-ai-code-of-conduct.html
  27. [30]Office of Management and Budget. (2022). North American Industry Classification System, United States, 2022. Executive Office of the President, Office of Management and Budget. https://www.census.gov/naics/reference_files_tools/2022_NAICS_Manual.pdf
  28. [31]Primec, A., Belak, J., & Čufar, M. (2026). Artificial Intelligence as an Emerging Risk Dimension in Corporate Sustainability Reporting: A Legal and Governance Perspective. Sustainability, 18(5), 2278. https://doi.org/10.3390/su18052278
  29. [34]Saeri, A. K., Graham, J., Noetel, M., Slattery, P., Ah-King, D., Aittokallio, E., Akindeihin, I., Al Mahdi, A., Alhajjar, E., Andersson Lipcsey, R., Ang, G., Azam, C. M., Azaria, A., Balkissoon, R., Barberá, I., Bareato, C., Barry, J., Basehart, M., Bean, A. M., ... Thompson, N. (2026). Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts. arXiv. https://doi.org/10.48550/arXiv.2606.04490
  30. [35]SEC Investor Advisory Committee. (2025, December 4). Recommendation of the SEC Investor Advisory Committee Regarding the Disclosure of Artificial Intelligence’s Impact on Operations. https://www.sec.gov/files/approved-artificial-intelligence-disclosure-recommendation-120425.pdf
  31. [36]Ségérie, C.-R., & Salter, J. (2026, June 17). Version 2 of the Hiroshima Process is good news for AI transparency. CeSIA — Centre pour la Sécurité de l’IA. https://cesia.org/en/publications/v2-hiroshima-process-good-news-for-ai-transparency/
  32. [37]Sherman, E., & Eisenberg, I. W. (2023). AI Risk Profiles: A Standards Proposal for Pre-Deployment AI Risk Disclosures. arXiv. https://doi.org/10.48550/arXiv.2309.13176
  33. [38]Slattery, P., Saeri, A. K., Grundy, E. A. C., Graham, J., Noetel, M., Uuk, R., Dao, J., Pour, S., Casper, S., & Thompson, N. (2026). The AI risk repository: A meta-review, database, and taxonomy of risks from artificial intelligence. Patterns, 7(5), 101517. https://doi.org/10.1016/j.patter.2026.101517
  34. [39]Smuha, N. A., & Yeung, K. (2024). The European Union’s AI Act: Beyond motherhood and apple pie? SSRN. https://doi.org/10.2139/ssrn.4874852
  35. [41]Uberti-Bona Marin, L. G., Rijksbosch, B., Spanakis, G., & Kolling, K. (2025). Are Companies Taking AI Risks Seriously? A Systematic Analysis of Companies’ AI Risk Disclosures in SEC 10-K Forms. SSRN. https://doi.org/10.2139/ssrn.5368860
  36. [42]U.S. Census Bureau. (2026). Business Trends and Outlook Survey (BTOS), 2023–2026 [Data set]. https://www.census.gov/hfp/btos/data

Paper details

vlnc:W_ep5cjzq9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 3

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents