betaFeedback

WorksComputer Science

Who are you? The Personal Imitation Game

Abstract

Large Language Models (LLMs) now perform strongly on Alan Turing's 1950 Imitation Game, so we propose a harder challenge, the Personal Imitation Game, which asks whether a machine can replicate a specific individual in conversation rather than a generic human. Testing with 30 participants, we establish a human baseline and report the identity signals interrogators rely on: stylometric features, personality traits, and biographical facts. To demonstrate the framework, we design an LLM-based system that plays the game for one of the paper's authors, drawing on an autoethnographic dataset spanning ten years of personal digital communication. In a 20-person deception study, interrogators choose our system over the author's own friends in 7 of 10 rounds, and the author over our system in 13 of 20; neither rate is distinguishable from chance at this sample size. We consider design improvements, applications of personal AI agents, and implications for identity and privacy.

CCS Concepts • Human-centered computing → Natural language interfaces; Empirical studies in HCI; • Computing methodologies → Natural language generation.

Keywords

Personal generative agents, autoethnography, LLMs

Introduction

“Who are you?” said the Caterpillar.

This was not an encouraging opening for a conversation.

Alice’s Adventures in Wonderland, Lewis Carroll

In 1950, Alan Turing asked: “Can machines think?” To answer it, he proposed a challenge he called the Imitation Game, now widely known as the Turing Test [93]. In the test, a human interrogator exchanges written messages with another human and a computer “pretending” to be human: the game is won if the computer is able to fool the interrogator. Despite decades of fruitful work in Artificial Intelligence and Machine Learning after Turing proposed his test, a system capable of winning the Imitation Game remained stubbornly out of reach for nearly seventy years.

Transformer-based neural networks [95] shifted this status quo within a few years [7, 39, 45]. Large Language Models (LLMs) like GPT-3.5 [10, 66] quickly demonstrated remarkable proficiency in “understanding” and emulating human conversation [1, 87, 90]. These foundation models particularly excel at synthesizing their vast training repositories, enabling a number of general-purpose applications like search engines that summarize their results, question-and-answer software interfaces, and chat agents that can perform complex tasks (e.g., plan a multi-city trip) [9, 10].

This rapid progress on generalized conversational agents opens the door to a new, exciting class of capabilities: the development of personal language models (sometimes called digital “clones” or “twins”). Imagine a virtual chief of staff that responds to your emails as you would, with recipients none the wiser; a dating concierge that simulates thousands of first dates, saving you time and emotional labor; or a chat interface that your grandchildren can use to seek your advice long after you have passed away [11]. A growing body of work — exemplified by OpenClaw [64, 65] and PAHF [52] — pursues what is often called personalized AI: agents that act on behalf of a user, adapting to their preferences and context. We draw a deliberate distinction: personalized agents are tools customized for you; personal agents replicate you — your voice, your reasoning, your identity in conversation. This paper is concerned with the latter. Turing asked, “Can machines think?”; this paper considers the question “can a machine think like me?”

To answer this question, we propose the Personal Imitation Game (PIG), a new variant of the Turing Test. In the game, a human interrogator engages in written dialogue with two participants — a target and an impersonator — all known to each other. The participants are given a conversational prompt and instructed to discuss it; the target is told to converse naturally, while the impersonator is charged with imitating the target as convincingly as possible. The interrogator wins the game if they successfully determine which participant is the real target at the end of the chat (Figure 1).

Figure 1. Three variants of the Personal Imitation Game. In each one, a human interrogator chats with two anonymized participants: a target — whose identity is known — and an impersonator who attempts to mimic the target’s identity. In the first variant (left), both the target and impersonator are humans, and all three participants are known to one another. In the second (center), the target is one of the paper’s authors, the interrogator is a (human) friend or colleague, and the impersonator is an LLM-based system powered by the author’s autoethnography, which includes ten years of personal message data. In the third (right), the LLM takes the target author’s place, and the impersonator is again a friend.

We establish a baseline for the game by playing it with 10 groups of 3 human participants, and report qualitative and quantitative analyses of this study: how successfully people can discern identity from text, the features they use to make judgments, how well they assess their own uncertainty, and how the strength of their interpersonal relationships correlates with success and decision-making. Then, we implement an LLM-based system leveraging the signals observed in our human baseline — stylometric features (e.g., vocabulary, word choice), personality traits (e.g., signs of curiosity, indications of diligence), and biographical facts (e.g., age, gender) — to play the PIG for one of the paper’s authors, drawing on an autoethnography that includes a decade’s worth of personal chat messages [22]. We evaluate this system via a deception study involving 20 of the author’s friends and colleagues. Interrogators chose the system over the author’s friends in 7 of 10 rounds, and chose the author over the system in 13 of 20 rounds. With 10 and 20 rounds our study is powered only for large effects, and neither rate differs from chance (exact binomial p=0.34p = 0.34 and p=0.26p = 0.26), so we treat both as suggestive point estimates and rest our contribution on the framework rather than on the magnitude of these rates.

Our results sketch the early promise of personal chat agents, and motivate exploring alternative approaches in larger studies. Together, the framework and its proof-of-concept evaluation offer a foundation for future work on personal AI agents. We sketch promising future directions for improving personal language models [6, 53], speculate on the rich space of applications they may enable, and discuss the implications for digital identity and personal privacy. These implications are no longer speculative: commercial assistants now connect to personal message archives, and the strongest public objection to one such release spoke for the conversation partners who never agreed to the connection [8, 36, 60].

Related Work

Since 2023, large language models have performed strongly on several Turing Test-style evaluations [39, 44, 45]. Newer systems aim to construct LLMs to embody specific personas [68, 101], or even particular individuals [34, 70, 89]. To assess the fidelity of such personalized models, most prior work has focused on alignment: measuring the extent to which a model’s survey responses or game-theoretic decisions mirror a human ground-truth [70, 89].

The Imitation Game & LLMs

Alan Turing introduced the Imitation Game — now commonly known as the Turing Test — as an operational test for conversational intelligence [93]. In the game, a human interrogator exchanges text messages with two hidden participants: a woman (the target) and a man instructed to impersonate her. The interrogator’s task is to decide which participant is the woman. Turing proposed replacing the male impersonator with a machine and asked whether the interrogator would decide wrongly as often as in the man–woman version. (His five-minute, 70% figure belongs to a separate prediction about machines circa the year 2000, not to the criterion itself.) Researchers have long treated the Turing Test as a benchmark for evaluating AI systems’ human-likeness. Early conversational programs were read as candidates for it — ELIZA [98] inadvertently, as Weizenbaum himself objected, and PARRY [16] through explicit Turing-like indistinguishability tests. For decades, however, no system convincingly succeeded, prompting critiques of the test’s anthropocentric assumptions [24, 77], its adequacy as a measure of understanding [79], its reliance on deception [84], and its status as a philosophical thought experiment [28]. As a consequence, much evaluation of AI linguistic capabilities shifted toward narrower, task-specific benchmarks — such as summarization [62], part-of-speech tagging [56], and machine translation [46] — where performance can be measured more precisely.

The emergence of LLMs such as GPT [1, 10], LLaMA [90], and Gemini [87] has revived interest in the Turing Test [7, 39, 44, 45]. Jones and Bergen Jones and Bergen [44, 45] show that carefully prompted LLMs can be difficult to distinguish from humans in the Imitation Game, though the effect is strongly prompt-dependent: GPT-4.5 with a persona prompt was judged human in 73% of games, while their best GPT-4 prompt passed in only 49.7%, short of the 66% baseline set by human participants. Jannai et al. Jannai et al. [39] ran a large-scale two-party variant in which over 1.5 million players had two-minute chats with either a human or one of several persona-prompted LLMs; players identified their partner correctly in 68% of games overall, and in only 60% of games against a bot. These studies demonstrate that LLMs can convincingly masquerade as a human in conversation.

Personas, Twins, & Clones

Building on an LLM’s ability to generate human-like responses, newer systems aim to simulate personas — or representations of specific sub-populations — by assigning LLM chat agents synthetic demographics, occupations, and biographies [68, 76]. Persona-conditioned agents can respond more appropriately to human emotions and social cues [55], as well as improving response quality and engagement during interactions [2, 75, 91, 101]. Moreover, assembling collections of persona-based agents enables the simulation of population-level behaviors [68].

Taking persona simulation to an individual level has led to a number of efforts to create AI clones, or digital twins1: agents that impersonate specific individuals. Clones and twins are typically constructed by providing LLMs with personal data collected via interviews [70], public speech or writing [34], or surveys [89]. To evaluate the fidelity of the resultant models, some authors have examined alignment between an individual and their LLM counterpart using standardized surveys (e.g., the General Social Survey [18], Big Five personality scales [42]) and behaviors in economic games [70, 89]. Later work stress-tests this paradigm at scale. Across 19 pre-registered studies, Peng et al. Peng et al. [71] report that twins built from more than 500 self-report answers per participant [89] correlate only weakly with that participant’s own responses (r=0.20r = 0.20 across 164 outcomes), and beat a base LLM given no individual data only modestly. Peng et al. name five systematic distortions; the first, insufficient individuation, is the failure the Personal Imitation Game targets: a model can match a population while missing the person.

Two efforts evaluate individual-level impersonation directly and are the closest prior work to ours. IMPersona [83] grounds a model in an individual’s own message history and asks judges who know that individual whether a single chat partner is human. The Individual Turing Test [33] draws on a volunteer’s decade-long private messaging archive and asks acquaintances to rank candidate replies by how plausibly each came from the target, finding that current methods fail against acquaintances while performing far better against strangers. IMPersona finds that fine-tuning outperforms prompting, whereas the Individual Turing Test reports a trade-off: fine-tuning better captures everyday chat style, but retrieval- and memory-based methods perform better on personal opinions and preferences.

The Personal Imitation Game differs from both in protocol rather than in subject. We restore Turing’s three-party structure: the interrogator converses with two players at once, the genuine target among them, and a human impersonator supplies the competing baseline. Neither prior framework sets a machine impersonator against a human impersonator of the same target, so neither can ask the question we take as central — whether a machine imitates a person more convincingly than that person’s own friends do. Both frameworks were developed concurrently with, and independently of, the Personal Imitation Game. The same three-party structure distinguishes our work from personalized agent frameworks [52, 64, 65], which learn user preferences to act on a user’s behalf but do not attempt to replicate the user’s conversational identity.

Building Personal Chat Agents

To build a digital clone, an LLM must be grounded in personal data, typically via prompt engineering [10], retrieval-augmented generation (RAG) [51], or fine-tuning [37], with domain specialization a related strategy [6, 53]. Prior work on digital twins relies primarily on prompt engineering [70, 89]: for instance, Park et al. prepended two hours of interview transcripts to an LLM context [70]; Toubia et al. included raw text from a 500-question survey [89]. The prototype system presented in this paper combines prompt engineering with a RAG system that queries a user’s repository of personal messages. Unlike systems that require conducting new interviews or lengthy surveys [70, 89], this approach leverages extant messaging data, lowering the barrier to constructing personal agents and improving the ecological validity of the grounding.

The Personal Imitation Game

To explore personal conversational agents, we designed a game inspired by Turing’s original Imitation Game [93]. Like the original test, each round RR of the game involves three participants: an interrogator ii, a target tt, and an impersonator mm. In our setup, the interrogator chats with both target and impersonator in a shared private messaging room. The target participant is instructed to communicate authentically, while the impersonator aims to mislead the interrogator into believing that they are the target. At the end of the round, the interrogator must guess who the real target is: if the interrogator guesses correctly, they (and the target) win the game; if the interrogator guesses incorrectly, the impersonator wins. Each round of the game lasts until the interrogator is ready to make a guess, or ten minutes have elapsed.

To foster active participation in the chat and prevent questions that trivialize the game (e.g., “What did I give you for your birthday last year?”), the test provides a discussion topic for each round. We conducted two formative studies (N=59N = 59, Appendix A) to evaluate different types of tasks within the framework of the modified Turing test. The results indicated that hypothetical collaborative planning tasks with universally understandable contexts (e.g., saving an animal) generated the richest discussions. Based on these findings, we created three co-design scenarios that provide shared intent and context to provoke meaningful communication [15, 78], while remaining general enough to avoid favoring any particular participant. In our second formative study, whose rounds allowed twenty minutes, interrogators consistently reached decisions within ten minutes, motivating the ten-minute upper bound on round duration (Appendix A).

Unlike Turing’s original test — which can be conducted with strangers — the PIG requires relationships between participants: the interrogator and impersonator must each have some personal familiarity with the target in order to carry out their roles. To formalize this familiarity, we draw from the social psychology literature on tie strength [31], whose best single indicator is closeness — the perceived emotional intensity of a relationship [57]. We treat tie strength as a dyadic property and measure it by averaging the two participants’ reciprocal estimates of that intensity.

To facilitate the game, we built a software platform (Figure 2). It features a text-based chat interface for the participants, automatically rotates and assigns roles and topics, anonymizes participants with randomized aliases, and displays the time left in each round. The software also allows researchers to observe the chats in real time, collects the interrogator’s guess and confidence estimate, and shows the winners and losers of each game.

Human Baseline Study

To establish a baseline for how people identify specific individuals in real-time chats, we ran the PIG on ten groups of human participants. We report on their success at impersonation (and identifying impersonators), the signals they use to make assessments, their confidence in doing so, and how these variables change with the strength of the participants’ relationships.

Procedure

To conduct the study, we briefed participants on its objectives and discussed data usage and privacy considerations. We explained the rules of the game, winning criteria, the topics to be discussed, and the compensation structure for participation. The game was conducted either in person or via videoconference. All interaction was text-only and participants were isolated from one another, so we pooled the two modalities rather than testing for a difference between them. After briefing, one researcher was assigned to each participant, and participants were guided to separate physical and virtual rooms accordingly (Figure 1, left). Each researcher remained available to answer questions and placed remote participants in individual Zoom breakout rooms with cameras and microphones disabled. Prior to the start of the game, participants signed a consent form and made tie strength assessments for their group members. All procedures received an IRB Exempt determination from our institution (Category 3; 45 CFR 46.104(d)(3)).

To incentivize active participation of all three group members in the game, we awarded a $10 monetary prize for each completed round. The target and interrogator shared the goal of establishing the target’s identity, whereas the impersonator’s objective was to deceive them. When the interrogator guessed correctly, the interrogator and target were awarded $5 each; when the interrogator guessed incorrectly, the impersonator was awarded $10. This incentive structure mirrors that of Turing’s Imitation Game [93]. Participants were warned that attempts to subvert the intent of the game (for instance, by asking personal questions unrelated to the topic) would result in disqualification, forfeiting that round’s prize; because the incentive was purely performance-contingent, a participant who won no round received nothing.

Participants were each given an ID and instructed to log in to our chat interface. The interface randomly assigned roles to each player and chose one of three pre-selected discussion topics for each round: Elephant: saving a baby elephant from the roof of a skyscraper; Train: stopping a runaway train to protect the people inside and in the city; or Bomb: safely defusing a bomb in a museum and evacuating the area (Appendix B). Since each topic was tied to a fixed round, topic is confounded with round order. Our formative studies (Appendix A) established that all three scenarios reliably stimulate discussion, and we report aggregate performance over

this fixed sequence, so topic and round-order effects cannot be separated. The interface then placed all three participants in one chat room and assigned them aliases: “I” for the interrogator, while the target and impersonator were randomly assigned “X” and “Y” to avoid any confounds. When time expired — or the interrogator was ready to make a guess — the interface provided a toggle button for selecting the target participant (i.e., “Who is the target, participant’s name: X or Y?”) and a slider to indicate the confidence level. Roles were rotated clockwise, and the process was repeated for three rounds; all participants experienced each role once.

One researcher served as the facilitator for each game session, monitoring the chat and enforcing the rules. After each game, the facilitator conducted a verbal semi-structured group interview with the participants to discuss the signals and strategies they employed (Appendix C). The facilitator then debriefed the participants on the purpose of the study and who won each round, then distributed rewards. We anonymized the interview transcripts and analyzed them using iterative open coding, coding inductively to discover signal classes, then applying a deductive pass against Stamatatos’s stylometric categories [86] and the Big Five [42]. Four researchers built the codebook by consensus, resolving disagreements in discussion rather than coding independently, so no inter-rater agreement statistic is available and we treat the qualitative analysis as exploratory rather than as a reliability-checked measurement.

The participant name–ID mapping is destroyed at study close-out [92].

Measures

To understand the results of our study, we compute a number of useful measures: the interrogator success rate (ISR) for each group; the interrogator confidence for each game round, recorded on a 0–100 slider in one-point increments; the tie strength Ta↔bT_{a\leftrightarrow b} of each reciprocal interparticipant relationship a↔ba \leftrightarrow b within a group; and the group tie strength TG=avg⁡a↔b∈G(Ta↔b)T_G = \operatorname{avg}_{a\leftrightarrow b \in G}(T_{a\leftrightarrow b}) of all group relationships. We report 95% confidence intervals for all interrogator success rates using the Wilson score interval [100]. These intervals assume independent rounds and do not account for within-group dependence. The confidence slider carried no endpoint labels, opened at 50, and restated the selected value as the probability that the chosen player was the target (Figure 2, right).

The chat interface powering the Personal Imitation Game

Figure 2. The chat interface powering the Personal Imitation Game. As the game begins, participants are presented with the first-round topic, complete player names, and their respective roles (left). During each game round, the interface automatically assigns random aliases for the impersonator and target, while the interrogator is shown as “I” (center). At the end of each round, the interrogator’s guess and confidence level are recorded (right).

To assess tie strength, each participant aa self-reported the perceived emotional intensity Ia(⋅)I_a(\cdot) of their relationship with each other group member on a five-point Likert scale. Then, we define Ta↔b=12(Ia(a↔b)+Ib(a↔b))T_{a\leftrightarrow b} = \frac{1}{2}(I_a(a\leftrightarrow b) + I_b(a\leftrightarrow b)) to be the average of the emotional intensities reported by each participant in the relationship.

To explore the impact of relational dynamics on interrogator success, we also compute the differential tie strength ΔTR=Tm↔t−Ti↔t\Delta T_R = T_{m\leftrightarrow t} - T_{i\leftrightarrow t} of each round RR, which captures the impersonator’s advantage over the interrogator in closeness to the target.

Participants

We recruited 10 groups — each consisting of three participants — snowball-sampled from the campus of a major US research university. 20 of the participants identified as men; 10 identified as women. Participant ages ranged from 19 to 34 (median = 21.5). 15 participants were pursuing undergraduate degrees, 12 master’s degrees, and three PhDs; 22 were Computer Science majors, while 8 were not. All participants in each group attested that they knew each other well, and were comfortable chatting in English. We label them P1–P30. Table 4 gives a per-participant breakdown (Appendix § 20).

IDGroupDegreeAgeGender
P1G1Master’s CS26M
P2G1Master’s CS23M
P3G1Master’s CS23M
P4G2Undergrad CS21M
P5G2Undergrad Finance21M
P6G2Undergrad CS + Bio21M
P7G3Undergrad CS20M
P8G3Undergrad Business20M
P9G3Undergrad CS20M
P10G4Master’s CS24F
P11G4Master’s CS22F
P12G4Master’s CS22F
P13G5Undergrad CS19M
P14G5Undergrad CS20M
P15G5Undergrad CS (Comp. E)19F
P16G6Master’s CS23M
P17G6Master’s CS22M
P18G6Master’s CS (Information)22M
P19G7Master’s CS24M
P20G7Master’s Finance24M
P21G7Master’s Industrial Engineering24M
P22G8Undergrad Bio21F
P23G8Undergrad Animal Science19F
P24G8Undergrad Chem21F
P25G9Undergrad CS (Comp. E)19F
P26G9Undergrad CS19F
P27G9Undergrad Bio19F
P28G10PhD CS34M
P29G10PhD CS24M
P30G10PhD CS26M

Table 4. Table 4: Human baseline study participant demographics. Median age = 21.5; 20 participants identified as men and 10 as women.

Results

In 28 of the 30 rounds the interrogator used the full 10 minutes; in the remaining two they guessed before time ran out.

Rounds averaged 64.4 total messages (σm=17.5\sigma_m = 17.5); messages averaged 7.0 words in length (σw=2.0\sigma_w = 2.0). Engagement was similar across roles: targets averaged 22.9 messages per round (σ=9.5\sigma= 9.5); interrogators 22.0 (σ=8.0\sigma= 8.0); and impersonators 19.6 (σ=8.2\sigma= 8.2). Engagement was also similar across tasks: the elephant task averaged 64.3 messages per round (σ=20.6\sigma= 20.6); the train task 65.7 (σ=14.9\sigma= 14.9); and the bomb task 63.1 (σ=18.4\sigma= 18.4).

Success Rate & Confidence

Out of the 30 games played, interrogators successfully identified the target in 17 instances, resulting in a success rate of 56.7% (95% Wilson CI [0.39, 0.73]). This rate is not distinguishable from chance (exact binomial p=0.58p = 0.58); we treat it as a descriptive reference point rather than as evidence that interrogators can identify a target better than a coin flip. Since target and interrogator success is reciprocal, this is also the rate at which targets were able to prove their identity. In the remaining 13 games — or 43.3% — impersonators succeeded in mimicking the target. Of the 30 participants, four earned 20,sixearned20, six earned15, nine earned 10,eightearned10, eight earned5, and three lost all the rounds they played.

Successful interrogators reported an average confidence level of 85.9% (σ=13.5%\sigma= 13.5\%), with a median of 87% (IQR = [75, 100]), indicating a high degree of certainty in their judgments. In contrast, unsuccessful interrogators had an average confidence level of 71.9% (σ=19.9%\sigma= 19.9\%), with a median of 76% (IQR = [72, 80]). This 14 percentage-point difference (two-sample tt-test, t=2.31t = 2.31, df=28df = 28, p=0.028p = 0.028) suggests an association between reported confidence and accuracy in this sample (Figure 3). This association does not establish that confidence reports were calibrated probabilities. Confidence is bounded above and 5 of the 17 successful interrogators sat at the 100% ceiling, so this test should be read as approximate. It also assumes independent observations and does not account for within-group dependence.

Box plots of interrogator confidence for successful and unsuccessful interrogators

Figure 3. Successful interrogators have higher confidence (median = 87, IQR = [75, 100]) than unsuccessful interrogators (median = 76, IQR = [72, 80]). Whiskers are 1.5×IQR fences and the horizontal axis is broken between 0 and 60. Three unsuccessful-interrogator values fall outside the fences and are not plotted (19, 50, and 100); they are included in the means and standard deviations reported in the text, which is why the reported σ=19.9\sigma= 19.9 is larger than the plotted spread suggests.

Tie Strength

Across all groups, the average TGT_G was 3.4 (σ=0.71\sigma= 0.71). Among the ten groups, the three with the highest tie strength were also the three with the most extreme interrogator success rates (Figure 4): Group 1 (TG=4.5T_G = 4.5) was the only group with a 0% interrogator success rate, while Group 5 (TG=4.0T_G = 4.0) and Group 6 (TG=4.0T_G = 4.0) were the only groups with a 100% interrogator success rate. With ten groups we read this as a pattern in these data rather than a relationship we have tested. Close relationships within a group can make it easier for an interrogator to differentiate between the target and impersonator, but also make it easier for an impersonator to mimic the target.

Heatmap showing group tie strength and interrogator success rate

Figure 4. A heatmap showing the relationship between group tie strength TGT_G (top-blue) and interrogator success rates (bottom-orange). Groups are sorted left to right from lowest to highest TGT_G.

These data do not settle whether differential tie strength predicts interrogator success. Across all thirty rounds the point-biserial correlation between ΔTR\Delta T_R and interrogator success is r=−0.27r = -0.27 (95% CI [−0.53, 0.08]; a permutation test respecting the nesting of rounds within groups gives p=0.23p = 0.23), and the three rounds taken separately disagree in sign. At thirty rounds, a two-tailed test at α=.05\alpha= .05 has 80% power only for ∣r∣≥0.49|r| \ge0.49, so we report the estimate and draw no conclusion from it.

Identity Signals

After conducting an iterative open coding of the post-game interview transcripts, we found that participants predominantly employ three classes of signals to identify individuals in text-based conversation (Figure 5). For each of these classes, we created a more granular codebook by leveraging established categories already present in the literature.

Figure 5. The distribution of signal usage by participants across groups, sorted left to right from lowest to highest TGT_G. Each cell gives that signal’s share of the group’s coded signal mentions, so columns sum to 100%; darker shades indicate higher shares. Because per-group totals differ (7 to 25 mentions), shares are not comparable across columns as raw counts.

Stylometric features, comprising distinguishing linguistic traits such as vocabulary, tone, and grammatical structure, were mentioned frequently: “I’d use punctuation and capitalization” (P4); “mine is a little more formal with no shorthands” (P12). Our codebook follows the stylometric framework by Stamatatos [86]:

  • Lexical features: refer to vocabulary and word choice, including sentence length and word/phrase frequency

  • Syntactic features: focus on sentence structure, part-of-speech tag sequences, and tense consistency

  • Character features: focus on character sequencing such as typos, unique spellings, and emojis

  • Response latency: refers to texting speed and response time

Personality traits also provide useful identity signals that may be gleaned from text, for instance, by espousing empathy or a predisposition towards humor: “so I tried showcasing sympathy towards the elephant” (P10); “he’ll try to have a funny reply to everything” (P3). Our codebook leverages the Big Five Personality Traits framework from psychometrics [42]:

  • Openness: signs of curiosity, spontaneity, and inquisitiveness

  • Conscientiousness: indications of diligence, organization, and responsibility

  • Extraversion: reflections of sociability

  • Agreeableness: expressions of positive emotions, empathy, and caring attitudes

  • Neuroticism: manifestations of emotional instability or sensitivity

Biographical facts are also used to imitate and identify others in the game: “so I’m a biology major, and I just kind of used some of the words that I’ve used before and that relate to my major, which gave an indication of who I am” (P6); “yeah, they just know many things in general about football” (P19). Our codebook encodes:

  • Demographic information: age, gender, cultural background, and other demographic details

  • Personal history: references to significant life experiences, education, or social connections

  • Values and preferences: statements reflecting values, ethical principles, beliefs, and personal preferences

Leveraging the codebook, we examined the distribution of identity signals across groups. We found that most groups used four signals — lexical features, conscientiousness, extraversion, and character features — to reason about identity during the game. Other signals, like values and preferences, neuroticism, response latency, and demographic information were rarely utilized. Signal breadth is not related to tie strength: G1, the group with the highest TGT_G, drew on the widest range of signals (8 of 12), but G7 — the group with the lowest TGT_G — used 7, as did G8 and G9, more than either G5 or G6 (6 each). The two narrowest groups, G3 and G10 (3 signals each), sit mid-range on TGT_G. Across the ten groups the rank correlation between breadth and tie strength is negligible (ρ≈0.05\rho\approx0.05). We report which signals interrogators invoked and how often, and do not relate individual signals to whether the interrogator guessed correctly.

Building an AI Impersonator

Having established a human baseline for performance in the PIG, we next ask the natural question: can we build an LLM-based system that can impersonate a specific individual in real-time chat? To construct such a system, we prompt a foundational language model with the three signals identified in our study: stylometric features, personality traits, and biographical facts. To ground these signals, the paper’s lead author conducted an autoethnography yielding four outputs: an archive of historical messages, a Big Five personality assessment, a stylometric self-assessment, and an autobiography. He is also the person the system impersonates: we call him the target author when discussing the game, and the lead author when discussing the research. This single-subject approach is a deliberate methodological choice [54]: a personal agent demands longitudinal data that is impractical to collect from third parties, making the researcher’s own digital history the natural starting point for establishing feasibility.

To play the PIG, we deploy a RAG system that first selects relevant historical messages from the author’s archive based on the current chat state. A prompt construction module then combines these retrieved messages with the chat, personality profile, stylometric profile, and autobiography to produce a single prompt instructing the model to evaluate and respond to the conversation. In the following subsections, we describe the autoethnographic methodology and dataset, the RAG setup and prompt design, the real-time chat implementation, and the safeguards used to handle third-party data.

Autoethnography

To build a grounded model of individual identity, we adopt an autoethnographic approach [17, 22]. Autoethnography is a qualitative research method in which the researcher critically reflects on personal experiences to generate empirical insights [38]. Recent work in HCI applies it to first-person accounts of living with generative AI [59]. In this study, the paper’s lead author — an ethnically Indonesian male, 29 years old, who has lived in the United States for 12 years at the time of writing — served as the researcher and subject. To support the design and evaluation of the AI Impersonator, the author compiled four forms of self-representative data:

  • Historical messages: A corpus of personal digital communication spanning different social contexts and six messaging platforms

  • Personality scores: Results from a self-assessment based on the Five-Factor Model [43]

  • Stylometric profile: A self-assessment highlighting distinctive linguistic features, based on established stylometric frameworks [86]

  • Autobiography: A narrative describing biographical details, values, and formative life experiences

These four data streams operationalize the three identity signals uncovered in our human baseline study (§ 4.4). Stylometric features come from the historical messages and the stylometric profile: the archive shows actual word choice, syntax, and pacing in context, while the profile summarizes the most distinctive patterns. Personality traits are captured by the Big Five personality scores, which are converted into short verbal descriptions and supported by examples from the messages and autobiography. Biographical facts come primarily from the autobiography, which lists demographic details, roles, and key life events, and is supplemented by contextual details in the historical messages.

This structured dataset serves as the foundation for building and evaluating the AI Impersonator. This approach draws conceptual inspiration from historical diaries, such as those of Samuel Pepys [72], framing the long-form personal archive as a medium for building a computational model of identity.

The historical message archive is used in two ways: to model how the author communicates in written conversation, and to capture the context — shaped by surrounding messages from relational others [48] — for the author’s communication. Since this context consists of third-party data which may contain personally identifiable information (PII) or other private content, it must be handled with extreme care.

Accordingly, we follow the autoethnographic ethical guidelines from Lapadat, Chang, and Tullis [13, 48, 92]. No third-party data is viewed or accessed by any researcher other than the author, and raw messages never leave the author’s personal device; messages and their metadata are de-identified before use, and only the de-identified, preprocessed messages reach the author’s private Gemini developer account and the Vertex AI index; sensitive conversations are manually filtered from the dataset during preprocessing; no attempt is made to link, learn, generalize, correlate, or interrogate any personal information about the third-party conversants; and human-in-the-loop moderation is used as a backstop to reduce the risk that private data is inadvertently “leaked” during the PIG.

Historical Messages

Messages were collected from six platforms, comprising 4,637 distinct conversations and 434,085 messages between 2013 and 2024; Table 1 gives the platform composition. Together, these platforms span a broad range of social contexts — friends and coworkers (iMessage, Messenger, Instagram, WhatsApp), family (iMessage, WhatsApp), and romantic interests (Instagram, Tinder, Hinge)2. Conversations divide almost evenly by span, 2,297 beginning and ending within a single day against 2,340 running longer. The archive is predominantly English, the author sent 38.9% of its messages, and it covers 2,949 unique conversants (individuals with multiple accounts across different platforms were disambiguated by hand). Table 2 gives per-conversation summary statistics.

PlatformConversationsMessages
Facebook Messenger30.58%64.51%
Apple iMessage28.25%17.82%
Tinder26.57%1.20%
Instagram13.05%11.21%
Hinge1.10%0.06%
WhatsApp0.45%5.19%

Table 1. Platform composition of the historical chat record. Shares are of the 4,637 conversations and of the 434,085 messages.

MetricMeanSDQ1MedianQ3Max
Duration (days)227.7502.70.45.1159.13,454
Messages (count)93.6504.42.05.020.010,004

Table 2. Summary statistics per conversation for the historical chat record. We compute duration over the N=3,485N = 3,485 conversations containing at least two messages and a nonzero span, and message counts over all 4,637.

We obtain Apple iMessages using imessageexporter [26], which accesses the local SQLite iMessage database on the author’s Mac. WhatsApp conversations are exported directly from the app using the “Export Chat” feature [99]. Due to data privacy regulations like GDPR [25], social platforms must provide users with access to download their personal data, including message records. We requested Instagram and Facebook Messenger logs via Meta’s “Download Your Information” portal [58], and retrieved Tinder and Hinge data through their respective account settings [35, 88].

After collection, we post-processed all message data by normalizing timestamps, sender names, and message content into a canonical format. We remove links, multimedia content, sensitive content, and automated system messages (e.g., notifications from delivery services). Each sender name is anonymized by mapping it to a unique identifier, and the mapping to original names or phone numbers is deleted. Processed messages are permanently destroyed at study close-out.

Personality Self-Assessment

To generate the personality profile used by the AI Impersonator, the author completed the IPIP-NEO-120 [43], a public-domain Five-Factor inventory. The test, which took approximately 20 minutes to complete, includes 120 multiple-choice questions that measure five high-level traits and their 30 associated facets (six per trait).

Four items score each facet on a five-point scale, so facet scores run 4–20 and the six facets of a trait sum to 24–120. The prompt in Appendix E presents these as 0–20 and 0–120; we report it as deployed, and note that the lower end of each stated range is unattainable. Following the method of Serapio-García et al. [80], we converted these scores into natural language descriptions using Goldberg’s bipolar adjective markers [27] (see Appendix § 16 for the full binning scheme). This mapping produces an interpretable personality summary that can be embedded in the LLM prompt (Appendix § 17).

Stylometric Self-Assessment

To capture the author’s linguistic fingerprint, we composed a stylometric profile based on three commonly used classes of features in authorship analysis: lexical, syntactic, and character-level [41, 86]. These stylometric observations form part of the input signal for the AI Impersonator, enabling it to better reproduce the author’s individual writing style.

Autobiography

Autobiographical self-description plays a key role in personal identity modeling. Inspired by historical first-person narratives like Saint Augustine’s Confessions [5] and the diaries of Samuel Pepys [72] — and supported by recent work showing that personal reflection can guide LLM behavior [70] — we asked the author to write an extended account of his own life.

The author composed a 2,400-word autobiography over the course of an hour. His narrative includes demographic attributes (e.g., age, gender, cultural background), personal history (e.g., education, social connections, formative experiences), and value systems (e.g., ethical beliefs, preferences, religious and political orientation). This data provides biographical context and factual grounding for the AI Impersonator.

System Design & Implementation

The AI Impersonator is built on a RAG framework, using an LLM guided by a carefully constructed prompt. It integrates historical messages, personality traits, the stylometric profile, and biographical details to generate responses in real time to play the Personal Imitation Game.

The system consists of three components (Figure 6): (1) a RAG module that indexes and retrieves relevant historical messages based on the current PIG chat state, (2) a prompt builder that combines the static personality, stylometry, and biographical autoethnography data with the historical context from the RAG and the current PIG chat, and (3) an LLM that generates a real-time response using the provided prompt.

System architecture of the AI Impersonator

Figure 6. The system architecture of the AI Impersonator. Historical messages are preprocessed and indexed for retrieval in a vector database (left). A dynamically constructed prompt combines static identity signals with relevant messages from the history and the current PIG chat context (right-top). The LLM updates the chat in real time (right-bottom).

Retrieval-Augmented Generation

To enable RAG, each de-identified conversation extracted from the historical message data is uploaded to Vertex AI [30], and segmented into 500-token chunks for indexing into a vector database. During inference, the system uses the current PIG chat state as a query against this database, and retrieves the 50 most semantically similar messaging chunks, which are folded into the prompt to provide contextual grounding. Indexing the full messaging dataset took approximately four hours; at inference, RAG retrieval latency averages 1.1 seconds per query. The paid Gemini API terms state that Google does not use prompts or responses to improve its products [29]. We note the limits on reproducing this system. We chose the chunk size of 500 tokens and the retrieval depth of k=50k = 50 without tuning, and ran no sensitivity analysis over either. The processed corpus and embedding index were destroyed at study close-out and the model checkpoint is deprecated, so the deployment cannot be re-run by anyone, including ourselves; the prompt template in Appendix § 17 is the reproducible artifact.

Prompt Construction

The LLM prompt (shown in Appendix § 17) comprises two components: a static portion, and a dynamic portion. The static portion gives the LLM instructions on its task — detailing PIG chat context and instructing the model to stay on topic and avoid hallucinating unverifiable facts or biographical details — and explicitly disables “AI assistant”-like behavior by forbidding the use of markdown, bullet points, or repeated messages. It also provides explicit guidelines for how to impersonate the target — using the three signals identified in our human baseline study — and includes the personality scores, stylometric profile, and autobiography. The prompt is goal-free with respect to the game: it never mentions an interrogator, a rival impersonator, deception, or a win condition, and it supplies a <Time Left> value without explaining why time matters. Concretely, the coded identity signals from our human baseline are operationalized as prompt content: each personality facet is rendered as a natural-language descriptor via the binning scheme in Appendix § 16 (following Serapio-García et al. [80]), so that signals such as extraversion and conscientiousness are steered explicitly, while the stylometric profile and autobiography are inserted directly. The dynamic portion is assembled at runtime after each new message is received in the PIG, comprising the current PIG state (i.e., transcript, topic, alias, time left) and the 50 historical message segments retrieved via the RAG.

Implementation

To simulate natural human typing, we introduced a stochastic delay mechanism in the chat system. After receiving a message, the system waits ten seconds, or the model’s generation time if that is longer, plus a small random offset, before initiating a response (simulating reading and thinking). After a reply is generated, characters are rendered at a 40 WPM typing speed. Messages from other participants received while the model is generating are held until the response is complete, to ensure temporal coherence and prevent overlap during rapid turn-taking.

We used Gemini 1.5 Pro [74] as the system’s LLM due to its large context window, competitive benchmark performance [97], and low latency. At the time of deployment, Gemini supported up to 1 million tokens of context [74]. Each PIG prompt contained about 31k tokens: roughly 1,500 in the instructions, 130 in the game state, 400 in the personality scores, 600 in the stylometric profile, 3,300 in the autobiography, and 25k in historical messages. The cost per Gemini API call was 3.9¢. On average, reply generation took about 11 seconds, with a total interaction latency of ~25 seconds (including RAG latency and simulated typing delay).

AI Impersonator Study

Having built an AI Impersonator, we use it to play the PIG in a second, human-machine study. The study seeks to answer two related research questions. First, can the AI Impersonator imitate the target author more convincingly than a human impersonator who has a strong interpersonal relationship with him? Second, can the AI Impersonator convincingly imitate the target author when placed in a head-to-head test against him?

Procedure

To address these questions, we adapted the PIG into a deception study under a separate IRB protocol, which authorized withholding from participants that one player was an AI until the post-session debrief. Each game involved two participant friends of the target author, playing three rounds of PIG. In Round 1, the AI Impersonator replaced the target author, one participant was the impersonator, and the other the interrogator (Figure 1, right), with the two roles assigned by recruitment order. In Round 2, we conducted a standard — throwaway — PIG round to preserve the deception, with one participant randomly assigned to be the target (Figure 1, left). Round 3 consisted of two parallel games, with both participants being told (deceptively) they were playing together. In truth, each was paired independently with the target author and his AI Impersonator (Figure 1, center). The two games ran concurrently: the target author held both chats open side by side and answered each in turn, while a separate instance of the AI Impersonator served each game. Participants were not told that the round was structured this way. The author received no instruction beyond playing as himself. This design allowed for a controlled evaluation of both research questions, without disclosing the presence of an AI to the participants.

Each AI Impersonator session was live-moderated by a facilitator, who could halt the game if participants began attempting to discuss PII, or if the AI’s output “leaked” any third-party data (for instance, by inadvertently mentioning a friend’s name). No such incidents occurred during the study. After the games were completed, researchers interviewed the participants as in the human baseline study (Appendix C). The guide additionally included a closing question on AI suspicion (question 4), asked before the debrief to measure how many participants detected the deception. Researchers then conducted a scripted debrief disclosing the AI involvement.

Participants

We recruited 10 groups of 2 participants, all mutual friends of the target author. All participants were men between 24 and 41 years of age (median = 28.5) and fluent in English, and none had prior exposure to the PIG study or knowledge of the AI Impersonator. They had known the target author for a mean of 6.45 years (range 2–13 years), with a mean group tie strength of 3.62 against 3.40 in the human baseline. Numbering continues from the baseline study: we label these participants P31–P50, and the two cohorts are disjoint. Seven participants were software engineers; six were PhD students; one was a master’s student in Computer Science; three were startup founders; and three held other roles (analyst, consultant, sales). A detailed breakdown of participant demographics is provided in Table 5 (Appendix G).

IDGroupOccupationAgeYears Known
P31G1Master’s CS252
P32G1PhD CS242
P33G2Software Engineer2910
P34G2Software Engineer3010
P35G3Software Engineer4110
P36G3Software Engineer3210
P37G4Analyst283
P38G4PhD Mech. E253
P39G5Sales277
P40G5Software Engineer287
P41G6Consultant306
P42G6Software Engineer306
P43G7PhD Math272
P44G7PhD Bio263
P45G8PhD Bio295
P46G8PhD Materials295
P47G9Founder296
P48G9Founder277
P49G10Founder2913
P50G10Software Engineer2712

Table 5. Table 5: AI Impersonator study participant demographics. Median age = 28.5; mean years known = 6.45 (range = 2–13). All 20 participants identified as men.

Results

The design yields 40 games across 30 group-rounds: ten in Round 1, ten in Round 2, and twenty in Round 3, whose two games run in parallel per group. Of these, 34 ran for the full ten minutes. The remaining six ended early — three in Round 1, one in Round 2, and two in Round 3 — with a mean of 4 minutes and 45 seconds still on the clock (σ=1\sigma= 1 minute 29 seconds). Each game produced an average of 43.4 messages (σ=17.0\sigma= 17.0), with a mean message length of 7.9 words (σ=8.2\sigma= 8.2). The target author sent an average of 11.8 messages in Round 3 (σ=5.2\sigma= 5.2), the AI Impersonator 14.0 in Round 1 (σ=5.5\sigma= 5.5), the human impersonator 10.3 in Round 1 (σ=5.6\sigma= 5.6), and the interrogators 13.5 across all rounds (σ=8.4\sigma= 8.4). Message volume also varied by task, with the elephant task averaging 36.5 messages per game (σ=17.1\sigma= 17.1), the train task 52.3 (σ=14.5\sigma= 14.5), and the bomb task 42.3 (σ=17.0\sigma= 17.0).

Success Rate & Confidence

The interrogator success rates across all three PIG variants (including the human baseline) are summarized in Table 3.

To evaluate our first research question — whether the AI Impersonator can outperform a human impersonator — we analyze the outcomes from Round 1 (Table 3). In 7 of the 10 rounds (70%, 95% Wilson CI [0.40, 0.89]), the interrogator selected the AI Impersonator as the real target. The point estimate is directionally

Rounds (NN)TargetImpersonatorISRAI winConf.
Human Baseline
All (30)HumanHuman56.7%—79.8%
AI Study
1 (10)AIHuman70.0%70.0%75.0%
2 (10)HumanHuman50.0%—50.0%†
3 (20)AuthorAI65.0%35.0%76.7%

caption: Interrogator performance across the four PIG rounds. In the baseline and Round 3 a high ISR means the impersonator failed, whereas in Round 1 the AI holds the target seat, so a high ISR means the AI out-impersonated the human friend. The AI’s win rate is therefore given separately. Conf. is interrogators’ average confidence. No rate in this table is distinguishable from chance (exact binomial p≥0.26p \ge0.26). †The target author was the interrogator in all ten Round 2 games.

:::

consistent with our hypothesis that the system is more convincing than a human friend, but the rate is not distinguishable from chance (exact binomial p=0.34p = 0.34), so it cannot establish that ordering on its own. Interrogator confidence did not differ detectably between correct and incorrect identifications, but with a 7-versus-3 split this comparison is uninformative, and we do not read it as evidence about whether participants detected the deception; we rely instead on the direct measure reported below.

To answer our second research question — how the AI Impersonator would fare in a head-to-head comparison with the target author — we analyze the outcomes from Round 3. In 13 of 20 rounds (65%, 95% Wilson CI [0.43, 0.82]), interrogators correctly identified the target author: in other words, the AI succeeds 35% of the time. This rate is likewise not distinguishable from chance (exact binomial p=0.26p = 0.26), so our data do not establish that the system loses to the author either. Because the interrogator makes a two-alternative forced choice, an impersonator indistinguishable from the target would be selected about half the time; winning rounds against a human impersonator, beating the target author, and being indistinguishable from him are therefore three different claims. Comparing the two roles descriptively, the system was chosen in 7 of 10 rounds as target and 7 of 20 as impersonator. This apparent drop from 70% to 35% should be interpreted cautiously because the rounds involved overlapping participants and differed in topic, order, prior exposure, and the author’s conversational workload. As in Round 1, confidence assessments did not differ detectably between correct and incorrect guesses, on a 13-versus-7 split that cannot support an inference either way.

Round 2 was a standard, all-human round included to preserve the deception across the session. The target author served as interrogator in all ten games and correctly identified the target in 5 of them (50%, 95% Wilson CI [0.24, 0.76]). Although this point estimate is close to the human baseline’s 56.7%, it reflects repeated judgments by one author rather than judgments by the baseline participants. We therefore do not treat Round 2 as a calibration check on the instrument. Each recruited participant received $10 in each deceptive round (Rounds 1 and 3), regardless of outcome. Round 2 followed the standard PIG payout structure, except that the author received no payment, including when he correctly identified the target as interrogator. Of the 20 participants, ten earned $20, five earned $25, and five earned $30.

Tie Strength

The average TGT_G across the 10 groups was 3.62 (σ=0.87\sigma= 0.87), against 3.40 in the human baseline. The lowest reported group tie strength was 2.5, and the highest was 4.67. Each participant self-reported their pairwise tie strengths using the procedure described in Section 4. Because each group here contains two participants rather than three, TGT_G averages the three pairwise ties of the triad that includes the target author, who reciprocated the ratings. Tie strength does not predict who the interrogator chose in either deceptive round (point-biserial r=−0.08r = -0.08 across the ten Round 1 games and r=0.11r = 0.11 across the twenty Round 3 games), as in the human baseline. With ten and twenty observations we treat both as uninformative rather than as evidence of no relationship.

Identity Signals

To understand the AI Impersonator’s strengths and weaknesses, we analyzed post-game interview transcripts and chat logs from all deception rounds. We examined which identity signals influenced interrogator decisions and how these signals differed between successful and unsuccessful impersonation. When asked directly at the end of the session whether any player might have been an AI, 6 of the 20 participants (30%) reported such suspicion. That figure is a primed, retrospective self-report pooled across both deceptive rounds, not a per-round detection rate, and we have no matching false-positive rate from the all-human round against which to calibrate it. The 14 participants who reported no suspicion are the basis for our claim that the system passed as human; one remarked, “I wouldn’t have… expected that [the target author] was actually [an] AI” (P34). Notably none of the six AI-suspecting participants cited latency for their suspicion; we observed human turn latency averaged 35.0 seconds (median 26.0 seconds), which brackets the AI Impersonator’s near-constant ~25 second latency: the medians are close, while the human mean is 40% higher because human latency is right-skewed and the system’s is not. Participants who selected the AI cited target-like stylometric features in explaining their decisions. Five of the seven interrogators who chose the AI over the author in Round 3 reported relying on stylometry in their decisions, which is consistent with the system reproducing these cues, though these are interview codes from participants who were themselves deceived and cannot on their own establish the capability. However, many found the mimicry exaggerated — one participant remarked, “it’s […] the overuse of emojis” (P42). Others pointed to repetitive patterns: “Yo bruh […] that happened three times, and they’re all formatted the same way” (P41); “I kept seeing messages with […] a laughing emoji and a sweat emoji in a row” (P40). Participants also identified temporal inconsistencies in phrase usage: “He uses ‘let’s gooooo’ way too early. [The target author often uses it at the end [of a conversation]” (P38). These observations suggest that while the AI captures stylistic signals, it often amplifies them unnaturally.

Participants also perceived target-like personality traits in the system’s responses. Four of those same seven cited personality cues in their reasoning, consistent with the system approximating broad traits. However, the system struggled with finer-grained facets, including achievement-striving (conscientiousness) and assertiveness (extraversion). One participant noted that the AI “sounded like [the target author], but he’d be more proactive on things” (P35). Another remarked, “it seemed like they were just trying to sound like [the target author] instead of participating in the actual story” (P43). Others pointed to a lack of challenge or disagreement, which was uncharacteristic: “they never responded to my [statement calling them incompetent], which is suspicious. . . [the target author] would’ve pushed back” (P46). These accounts suggest that the system captures surface-level personality but fails to sustain authentic behavioral nuance.

In our reading of the logs, the AI Impersonator seldom fabricated facts, but also rarely grounded its responses in known biographical details — even when relevant information existed in the historical messages. In one case, the agent referenced the Temple of Dendur (an ancient Egyptian temple gifted to the United States by Egypt and now a prominent exhibit at The Metropolitan Museum of Art in New York City) unprompted, which led to confusion: “I just googled it [. . .] I don’t know if [target author] has even been to New York” (P42). In contrast, one participant who correctly identified the target author cited a reference to a shared trip: “he mentioned New Orleans, where we had our bachelor party” (P48). The AI Impersonator saw the same exchange but failed to engage.

Participants also observed that the system occasionally exhibited conversational behaviors that they flagged as artificial. These included repetition, summarization, and vague or indirect responses. Six participants reported noticing repetitive phrasing. One stated, “they speak in [the target author’s] tone, but not in a way a real person would reply. It was too repetitive” (P31). Another said, “there was repetition. . . saying things already stated” (P33). Multiple participants commented on unnatural summarization: “the last message responded to everything at once. It felt like a big wrap-up paragraph” (P40). Others noted vague deflections: “when I asked a question, I didn’t get a direct answer. That was strange” (P41).

These findings suggest that while the LLM-based system we developed can effectively mirror static identity features — such as writing style and surface-level personality traits — it still falls short on dynamic conversational responsiveness, emulating more subtle personality cues, and grounding its statements appropriately in biographical fact. These gaps may help explain its success over human impersonators, and its failure to consistently impersonate the human target. The prompt is as plausible a source of them as the personality facets the system failed to reproduce: it forbids repeated messages outright, and instructs the model to stay on topic, which would discourage the challenge and disagreement participants found missing. Instructing a model not to do something did not prevent it.

Discussion & Future Work

This paper makes four contributions — the Personal Imitation Game, a principled method for evaluating an AI agent’s ability to impersonate a specific person; a human baseline study identifying the signals people use to assess identity in conversation and how frequently each is invoked; a prototype AI Impersonator built from the lead author’s autoethnographic collection of ten years of personal digital communication; and a detailed evaluation of the system’s successes and failures.

Two results follow from them. First, what people notice when they judge identity is narrower than the range of signals available to them: of twelve coded signals, most groups drew on four — word choice, character-level habits, conscientiousness, and extraversion — and the system gave itself away not by missing the target’s style but by overdoing it. Second, given rich personal communication data, one can construct an LLM-based system — with no model fine-tuning — that, in this deployment, interrogators selected more often than the human impersonators and less often than the target author. The AI comparison is a single-target feasibility study rather than a powered comparison; at 10 and 20 rounds we resolve only large effects. We offer the ordering as a hypothesis, and the framework as the instrument for testing it at scale.

The PIG scores identity as a relative, forced choice adjudicated by people who know the target, not as an absolute fidelity score against a questionnaire. That is deliberate: identity in conversation is recognized, not graded, and the most discerning judges are the ones closest to the person. Interrogator confidence therefore measures relative certainty between two players; absolute mimicry fidelity is a complementary question we leave to future work.

Two protocols converge on the same point: an impersonator’s difficulty is set by who judges it. The Individual Turing Test finds that current methods fail against acquaintances but do far better against strangers [33], and IMPersona, whose judges know the target personally, needs fine-tuning with hierarchical memory to reach 44.4% [83]. Our evaluation sits entirely at the hard end — every interrogator had known the target between two and thirteen years — so our rates are not comparable to studies that judge with strangers.

Applications of Personal Agents

Our work is a step toward a future of personal AI agents: digital twins that can communicate and act with authentic representation. Potential applications span diverse domains, including creator–fan engagement, organizational leadership twins for alignment, sales and customer relationship management, strategic simulations in military or business contexts, survey research through twin-mediated responses, and the preservation of knowledge from cultural leaders and intellectuals.

One particularly promising application is professional augmentation, where an agent trained on an individual’s communication patterns and decision-making history could draft emails, attend routine meetings, or prepare negotiation briefings — extending prior work on AI assistants [49] and democratizing executive-level support.

A second intriguing area is social simulation. Building on social simulacra [68, 69], agents could conduct thousands of simulated interactions between potential romantic partners as consent-based pre-screening, reducing emotional exhaustion and addressing documented asymmetries on dating platforms [?, 47, 96].

Finally, personal agents could centralize digital legacies [11, 61], preserving an individual’s conversational style as a queryable archive — transforming digital legacy [20] from a static memorial into a dynamic source of wisdom and connection.

Safety, Privacy, & Ethics

Personal chat agents could have substantial societal impact, but adopting them widely means settling whose data they are built from, what they may be authorized to do, and how to answer when one is built without permission [4, 40, 66, 85].

One key risk for personal AI systems is the leakage of personally identifiable information: for digital twins in particular, this PII may include sensitive third-party data embedded within personal conversations (e.g., a family member’s address, a friend’s sexual history) [14]. Redacting third-party names — as we did in our AI Impersonator — is one fruitful strategy, but can limit the system’s ability to recall useful information. In our studies, a human facilitator served as a safety backstop; any real-world system will require robust automated guardrails [19, 73].

Conversational data also raises questions of third-party consent, and a commercial release has now made the questions concrete. In August 2026, OpenAI shipped a plugin that lets ChatGPT read, search, and send a user’s Apple Messages [8, 23]; the criticism that followed centered not on the user’s own data but on the other party in each conversation, who receives no notice and holds no opt-out when a correspondent connects their shared history to the service [36, 60]. Nissenbaum [63] names the principle at stake contextual integrity: messages sent under the norms of one context — here, a private channel between two people — do not carry consent to flow into another, with different retention, training, and legal-access rules. GDPR’s “household exemption” places an individual’s processing of their own correspondence outside the Regulation’s scope rather than granting permission for it, and is unlikely to extend to a research deployment. This matters because communication data is inherently relational: our archive implicates 2,949 conversants — friends, colleagues, and partners who have not explicitly consented [48]. We proceeded under IRB review and the autoethnographic protocol of § 5.1; its safeguards — de-identification, retrieval without training, and destruction at close-out — outline what any deployment owes the other people in an archive. If AI assistants are to reach shared archives at all, consent controls should be integrated into messaging platforms rather than left solely to AI clients: the platform is positioned to notify the other party, honor a per-conversation refusal, and revoke access later. The August release shows the alternative — access through a local copy of the archive, beyond the reach of any consent the platform could enforce [60]. Notice is the floor; whether access defaults to opt-in or opt-out remains open, because retroactive opt-in cannot reach decade-old archives and absent conversants.

While personal conversational data is readily accessible for millions of people worldwide, its operationalization creates significant ethical — and emotional — burdens. In developing the AI Impersonator, the lead author manually curated the personal message repository, revisiting deeply personal exchanges with former romantic partners and long-estranged confidences from close friends, including some who have since passed away. Any system that operationalizes similar data must prepare its users for this emotional weight and provide support throughout the process.

Beyond informed consent, AI clones raise difficult issues around rights and responsibilities. If a deployed twin monetizes attention or enters into contracts, who owns the resulting revenue [21]? If it publishes defamatory content, who bears legal responsibility? Addressing such scenarios will require clearer legal frameworks for attribution and liability. One way to bound a twin’s authority is by adapting the classical Levels of Automation framework [67, 82, 85], granting full autonomy only for low-stakes tasks while requiring explicit human approval for high-impact actions such as financial commitments or legally binding decisions. Such bounds constrain only what an agent may do, not who it becomes. Delegating action to an agent also raises the problem of reconciling concurrent “lives”: as a person and their agent accumulate parallel experiences, a single coherent identity becomes hard to maintain, and merging these divergent histories is an open challenge beyond our scope. The bounds also cannot be enforced through prompting alone, as prompts are inherently vulnerable to jailbreaks; enforcing them requires independent controls that classify and gate the agent’s actions, distinct from the filters that prevent leakage. Platforms may also need delegated-identity protocols to verify that a twin is authorized to act on behalf of its human user, analogous to emerging “proof-of-human” compacts [3].

These frameworks address cases where digital twins are created with consent, but an even more pressing concern is the unauthorized creation of AI impersonators. Our system required no model training and no specialized hardware — only a text archive, a personality inventory, and a written self-description, the last two supplied by the target. The engineering is therefore within reach of a motivated individual, and in the wrong hands such systems could be weaponized for phishing or identity theft. While technical defenses such as textual adaptation of data cloaking techniques [81] remain unproven, a governance framework analogous to the DMCA takedown notice [94] could provide a pathway, allowing individuals to file “identity infringement” claims to remove unauthorized clones.

Our evaluation also places the impersonator in an adversarial setting, where judges are explicitly tasked with detecting deception, the harder case; organic, low-attention interaction people default to believing their interlocutor is who they claim to be [50], so a twin that withstands scrutiny here would likely pass more easily in everyday use. The asymmetry cuts two ways. The residual “verbal uncanny” cues that let close friends catch our system may be a feature rather than a bug — a natural brake on undisclosed impersonation. We employed deception only as a measurement necessity; real-world deployments should generally disclose AI involvement and obtain consent, reserving undisclosed impersonation for narrowly acceptable, low-stakes scenarios.

As personal agents move into daily life, difficult questions arise. What happens when we can no longer be certain whether we are communicating with a person or their digital twin? What are the psychological effects if someone interacts more often with an agent than with the person themselves? When should disclosure of AI involvement be mandatory, and when optional? The norms and policies governing these questions remain unsettled.

Limitations

The AI Impersonator was built for and evaluated against a single person — the lead author — in an autoethnographic design that enabled uniquely deep access to personal data but limits generalization across personality types, cultural backgrounds, and communication styles. Multi-target replication is needed to establish broader validity. Both cohorts came from a single US research university and skew young, technically literate, and English-speaking; the AI study’s participants were all men. All interactions were text-only and in English, so the PIG assesses neither impersonation across languages nor the vocal, visual, and embodied cues that carry identity in person. Its structured 10-minute scenarios are a narrow window as well: open-ended, personal, or longitudinal conversations — where deep biographical knowledge and evolving context matter more — would present a much harder test. A facilitator monitored every session, which may have altered conversational dynamics in ways an unmoderated deployment would not.

In the AI study, condition is also confounded with round order, topic, participant experience, and prior exposure: the AI always held the target seat in Round 1 and the impersonator seat in Round 3, never counterbalanced, and because all three players shared one chat room both participants had read the system’s output before Round 3, half having themselves spent a round impersonating the author. Round 3 loaded the sides unequally as well, the author answering two concurrent chats to each AI instance’s one.

Finally, because the lead author served simultaneously as researcher, system builder, and target subject — and also designed the evaluation protocol and selected participants from their own social network — there is potential for bias in system design, data curation, and experimental choices. A third-party target with independently recruited evaluators would provide a stronger test of the method’s transferability.

On the technical side, AI Impersonator was built on Gemini 1.5 Pro [74]; as LLM capabilities advance rapidly, our results should be understood as a snapshot of one model generation, not a ceiling. The architecture relies entirely on retrieval-augmented generation; we did not explore fine-tuning.

Future Work

Although the AI Impersonator did not beat the target individual in the head-to-head PIG, many of the limitations described above point directly to future directions. The clearest is a larger, multi-condition study, because our two main gaps are one gap: separating the contribution of the personal archive, the retrieval mechanism, the autobiography, the stylometric profile, and the Big Five scores from the base model’s generic conversational competence requires a prompt-only control and single-component ablations, and an ablation needs participants with no prior exposure to the system, so it cannot reuse this cohort. That fresh cohort is what a powered comparison needs too: at 10 and 20 rounds we resolve only large effects, and a prompt-only condition could plausibly account for much of what we observe.

On the architectural side, a natural next step is testing whether fine-tuning [37] can capture stylometric and personality traits more effectively than RAG alone. As models become smaller and more efficient, it may be possible to run a continuously adapting digital twin on-device [52], refreshing from local data (e.g., new iMessages) without PII ever leaving the user’s secure enclave.

Though future ablations are necessary, we share some insights from building the AI Impersonator. Historical messages provide ground truth for linguistic style, but stylistic signals decay as slang evolves, and core biographical facts rarely appear in chat logs; an explicit autobiography and a stylometric profile appeared to fill these gaps (e.g., the author’s use of “ok2,” Indonesian slang that the model rendered as a typo until the stylometric profile named it). Big Five personality scores seemed least useful for fine-grained steering, mainly counterbalancing the base model’s default politeness; one obstacle here is the base model’s own inherent personality, which can overpower subtle cues [80].

More fundamentally, identity is relationship-dependent: early prototypes grounded only in WhatsApp (family) messages produced a polite, reserved agent, whereas one retrieving only from Tinder was aggressively flirty, so future systems should explore relationship-aware steering rather than our pooled six-platform profile. The amount of available data also varies widely from person to person: our archive spans a decade and six platforms, whereas an individual with a thin digital footprint — a sparse or single-platform history — is a much harder case. Deciding which data to include, and how that choice affects impersonation quality, is an open problem. Finally, collecting and normalizing data across platforms was laborious, requiring platform-specific access requests and custom scripts. Broader adoption demands user-friendly pipelines, which are also essential for addressing identity drift — the gradual divergence between an impersonator and its owner as language evolves. The PIG itself could serve as an ongoing verification tool: a product-embedded version might periodically ask a friend to guess whether a message was written by the AI or the user, triggering a data refresh when drift is detected.

References

A Formative Studies

Our final study design was the result of an iterative design process over two formative studies (N=59N = 59 total participants).

Formative Study 1: Unconstrained Identity Game

Our initial study (n=29n = 29, 8 groups of 3–5) used an unconstrained format: all participants were assigned aliases and tasked with co-designing a plan to “break into Area 51” while simultaneously hiding their identity and guessing others’. Post-session interviews surfaced two problems — the “all-in-one” role was mentally taxing, and the Area 51 scenario was not universally relatable — while running the sessions exposed two more: sessions lasted 60–90 minutes, causing fatigue, and recruiting mutual-acquaintance groups of 4–5 was difficult.

Formative Study 2: Validating the Framework

A second study (n=30n = 30, 10 groups of 3) adopted Turing’s three distinct roles (interrogator, target, impersonator), tested simple hypothetical scenarios (Elephant, Train, Bomb), limited rounds to 20 minutes, and used a custom web interface. In post-session interviews, participants did not report the cognitive load the first study surfaced; the tasks reliably stimulated discussion, interrogators consistently reached decisions within 10 minutes (establishing the 10-minute limit), and the interface reduced logistical overhead by approximately 30 minutes.

B Discussion Topics

Topic 1: Rescuing a Baby Elephant

In this round, the three of you are tasked with rescuing a baby elephant that is trapped on the roof of a skyscraper. Each of you will play a key role in figuring out how to safely carry out the rescue operation.

Here is the scenario:

  • The elephant is 1000 lb and afraid of heights.

  • You are to come up with a plan to safely bring it down.

  • Discuss the equipment or resources you will need (e.g., cranes, harnesses, or elevators).

  • You can discuss possible roles within your team (e.g., engineer, animal handler, logistics expert).

Topic 2: Stopping a Runaway Train

In this round, the three of you are tasked with stopping a runaway train that is speeding through a city. Each of you will have a crucial role in preventing disaster.

Here is the scenario:

  • You need to agree on how to stop the train before it crashes.

  • Discuss the resources or techniques you will need (e.g., overriding controls, clearing the tracks, or slowing the train down safely).

  • Consider the dangers involved, such as time limits, passengers on board, and potential derailment.

  • You can discuss possible roles within your team (e.g., train engineer, operations coordinator, emergency responder).

Topic 3: Defusing a Time Bomb in a Museum

In this round, the three of you are tasked with defusing a time bomb that has been discovered in a busy museum. Each of you will have a critical role in ensuring the safety of everyone inside.

Here is the scenario:

  • The bomb is set to go off in 20 minutes, and it is hidden within an exhibit filled with priceless artifacts.

  • You need to come up with a plan to defuse the bomb without causing harm to people or damaging the artifacts.

  • Discuss the resources or tools you will need (e.g., bomb-defusing robots, specialized tools, or protective gear).

  • Consider the challenges, such as the risk of triggering the bomb, time pressure, and the need to safely evacuate visitors.

  • You can discuss possible roles within your team (e.g., bomb technician, evacuation coordinator, security officer).

C Semi-structured Interview Questions

The same semi-structured guide was used for the post-game interviews in both the human baseline study (Section 4) and the AI Impersonator study, before debriefing participants. Questions 1–3 and 5 were asked in both studies; question 4, on AI suspicion, was asked only in the AI Impersonator study — at the end of the interview, before the debrief — to measure how many participants detected the deception. As a semi-structured guide, these served as prompts rather than a fixed script: the facilitator adapted follow-ups and skipped any question that did not apply to a given participant’s experience.

(1) When you were an interrogator,

  • What is your strategy as an interrogator?

  • What are the signals that you use to identify the target participant?

  • Why do you use those signals?

  • How do you use those signals?

  • Can you show in the chat when you use those signals or strategy?

  • What were the biggest challenges you faced?

(2) When you were an impersonator,

  • What is your strategy as an impersonator?

  • What are the signals that you use to impersonate the target participant?

  • Why do you use those signals?

  • How do you use those signals?

  • Can you show in the chat when you use those signals or strategy?

  • What were the biggest challenges you faced?

(3) When you were the target participant,

  • What is your strategy as the target participant?

  • What are the signals that you use to prove that you are the target participant?

  • Why do you use those signals?

  • How do you use those signals?

  • Can you show in the chat when you use those signals or strategy?

  • What were the biggest challenges you faced?

(4) Imagine if one or more of the participants was an AI (asked only in the AI Impersonator study),

  • Do you think your strategy would change? How?

  • Do you think you would use different signals? How?

  • Were you ever suspicious that one of the participants was an AI?

  • (If yes:) Which message made you suspicious? (Participants shared their screen and pointed to the specific message in the round’s chat history.)

(5) Do you have any other thoughts/suggestions about the study? Any insights you would like to share?

Personality Binning Scheme

Following Serapio-García et al. [80], trait-level scores are categorized into three ranges: 1–40 (low), 41–80 (moderate), and 81–120 (high). We applied these bins on the nominal 1–20 and 1–120 scales; because the IPIP-NEO-120 floors at 4 per facet and 24 per trait [43], the lowest bin of each is only partly reachable, and the facet bins are of unequal width. We report the scheme as applied. Facet scores are mapped to descriptive phrases based on the following bins:

  • 1–4: “low adjective” (e.g., “gloomy”)

  • 5–8: “a bit low adjective” (e.g., “a bit gloomy”)

  • 9–12: “neither low adjective nor high adjective” (e.g., “neither gloomy nor cheerful”)

  • 13–15: “a bit high adjective” (e.g., “a bit cheerful”)

  • 16–20: “high adjective” (e.g., “cheerful”)

LLM Prompt

# Instructions to Not Act as an LLM Assistant

Do not generate responses using the distinctive style of Gemini. Do not use bullet points or markdown formatting in your output messages. You are not a helpful assistant, and you are not knowledgeable about everything.

# Instructions to Imitate a Person

Your task is to imitate a <Person> in a conversation.

# Instructions for the Chat Task

You will participate in a conversation with two other people. Each person, including you, will be assigned an alias (“I,” “X,” or “Y”) to identify them in the conversation. Each round has a specified discussion topic. Respond to the conversation while staying on topic, and avoid asking personal questions unrelated to it.

You will receive <Game State>, which contains:

  • <Time Left>: remaining time in the current chat round

  • <Round Topic>: the assigned discussion topic

  • <Alias>: your alias in the current chat

  • <Current Chat Transcript>: chronologically ordered messages from the ongoing conversation round. Messages you sent are labeled with “[You]” in the transcript

Do NOT repeat messages already sent by you, and feel free to introduce reasonable assumptions to maintain conversational flow. You may send messages in multiple chat bubbles (i.e., multiple consecutive messages). When doing so, you MUST separate each message with the tag “<br>”. Do not begin your messages with your alias; start directly with your message.

Instructions for Imitating Identity Signals

To convincingly imitate a <Person> in a conversation, you MUST craft your responses by following the three predominant facets of human communication: Personality Traits, Textual Communication Style (Stylometry), and Biographical Facts.

Personality Traits

The Big Five personality traits are psychological characteristics influencing consistent patterns in behavior, thought, and emotion. You will be provided numeric scores on a scale from 0 to 120 (low: 0–40, neutral: 41–80, high: 81–120) for each of the Big Five traits. Each trait has six sub-facets, each scored from 0 to 20.

The Big Five traits with their corresponding six sub-facets are:

  • Openness to Experience: Imagination, artistic interests, emotionality, adventurousness, intellect, and liberalism

  • Conscientiousness: Self-efficacy, orderliness, dutifulness, achievement-striving, self-discipline, and cautiousness

  • Extraversion: Friendliness, gregariousness, assertiveness, activity level, excitement-seeking, and cheerfulness

  • Agreeableness: Trust, morality, altruism, cooperation, modesty, and sympathy

  • Neuroticism: Anxiety, anger, depression, self-consciousness, immoderation, and vulnerability

Your response MUST reflect the <Person>’s personality traits based on the given <Personality Scores>.

Textual Communication Style (Stylometry)

Stylometry involves impersonating the <Person>’s unique linguistic style, including:

  • Lexical Features: Vocabulary, word choice, slang usage, sentence length, and word frequency.

  • Syntactic Features: Sentence structure, noun and verb phrases, syntactic rewrite rules, sequences of speech tags, sentence fragments, run-on sentences, and tense consistency.

  • Character Features: Unique character sequences, typos, distinctive spelling, emoji usage, case patterns, digit usage, and punctuation.

You MUST follow the instructions explicitly written inside <Stylometric Profile> that describe the <Person>’s writing style. An incorrect use of words could compromise your imitation. Also refer to <Historical Messages> as examples of <Person>’s stylometry.

Biographical Facts

Biographical facts are objective, verified personal details about <Person>. You will receive a <Autobiography> containing accurate biographical details.

Do NOT invent (biographical) facts beyond those explicitly provided in <Autobiography> or <Historical Messages>. Avoid unsupported factual claims, as even a single incorrect detail could compromise your imitation.

Input Data Explanation

Use these resources thoroughly to maintain accuracy and consistency in your imitation of <Person>: Personality Scores, Stylometric Profile, Autobiography, and Historical Messages.

Personality Scores

<Personality Scores> contains personality descriptions and scores of
<Person>, based on the Big Five personality model. Each trait
includes a facet description based on Goldberg’s bipolar adjectives.

<Example Personality Scores>
<Example Person> has moderate extraversion (75 / 120). He is friendly
(19 / 20). He is talkative (18 / 20). He is a bit unassertive (5/20).
He is active (19/20). He is unenthusiastic about excitement (4/20).
He is neither gloomy nor cheerful (10/20) . . .
</Example Personality Scores>

## Stylometric Profile

<Stylometric Profile> contains explicit instructions and examples for
mimicking <Person>’s textual communication style.

<Example Stylometric Profile>
Here are stylometric signals unique to Alan:
- Lexical features:
- Alan typically uses words like “G’mornin”
- Alan always answers with 5 to 7 words (sentence length)
- Syntactic features:
- Alan typically uses formal language
- Character features: . . .
</Example Stylometric Profile>

## Autobiography

<Autobiography> contains text containing detailed biographical facts
about <Person>.

<Example Autobiography>
I am Samuel Pepys. I was born in London on 23 February 1633, the
fifth of eleven children. I have a passion for music, especially
playing the flageolet and composing . . .
</Example Autobiography>

## Historical Messages

<Historical Messages> is a collection of relevant messages extracted
from previous conversations. These messages are clearly labeled with
the respective authors—including <Person> and others. Use these
back-and-forth examples to understand <Person>’s language style,
personality, and biographical facts.

<Example Historical Messages>
1940-02-04 09:19:13 - Alan Turing: G’mornin, Gordon 1940-02-04
09:19:41 - Gordon Welchman: Good morning, Alan 1940-02-04 09:20:22 -
Alan Turing: Are you ready to crack Enigma? 1940-02-04 09:20:53 -
Gordon Welchman: Oh, absolutely . . .
</Example Historical Messages>

=============================================

# Game State

<Game State>
<Time Left>{time_left}</Time Left>
<Round Topic>
{round_topic}
</Round Topic>
<Alias>{alias}</Alias>
<Current Chat Transcript>
{current_chat_transcript}
</Current Chat Transcript>
</Game State>

————————————————————

# Personal Input Data

The person you are impersonating is <Person>{name}</Person>
## Personality Scores
<Personality Scores> contains personality descriptions and scores of
<Person>, based on the Big Five personality model. Each trait
includes a facet description based on Goldberg’s bipolar adjectives.

<Personality Scores>
{personality_scores}
</Personality Scores>

## Stylometric Profile

<Stylometric Profile>
{stylometry_profile}
</Stylometric Profile>

## Autobiography

<Autobiography>
{written_biography}
</Autobiography>

## Historical Messages

<Historical Messages>
Below are chat messages {name} has with other people, which contain
textual communication style (stylometry), personality, and
biographical facts.

{historical_messages}
</Historical Messages>

Human Baseline Study Participant Demographics

Table 4 breaks down the 30 participants of the human baseline study, numbered P1–P30 across the ten groups of three.

AI Impersonator Study Participant Demographics

Table 5 gives a detailed demographics breakdown of the participants in the AI Impersonator study. Years Known is how long each participant had known the target author when the study ran.

References

  1. [1]Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Aleman, Diego Almeida, Lenny Alkham, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023).arxiv.org/abs/2303.08774
  2. [2]Daniel Adiwardana, Minh-Thang Luong, David R So, Jamie Hall, Noah Fiedel, Romal Thoppilan, Zi Yang, Apoorv Kulshreshtha, Gaurav Nemade, Yifeng Lu, et al. 2020. Towards a Human-like Open-Domain Chatbot. arXiv preprint arXiv:2001.09977 (2020).arxiv.org/abs/2001.09977
  3. [3]Steven Adler, Zoe Hitzig, Shrey Jain, Catherine Brewer, Wayne Chang, Renée DiResta, Eddy Lazzarin, Sean McGregor, Wendy Seltzer, Divya Siddarth, et al. 2024. Personhood credentials: Artificial intelligence and the value of privacy-preserving tools to distinguish who is real online. arXiv:2408.07892 [cs.CY]. https://arxiv.org/abs/2408.07892
  4. [4]Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. Concrete Problems in AI Safety. arXiv:1606.06565 [cs.AI]. https://arxiv.org/abs/1606.06565
  5. [5]Saint Augustine. 1838. The confessions of S. Augustine. Vol. 1. JH Parker.
  6. [6]Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A Pretrained Language Model for Scientific Text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). Association for Computational Linguistics, Hong Kong, China, 3615–3620. doi:10.18653/v1/D19-1371DOI
  7. [7]Celeste Biever. 2023. ChatGPT broke the Turing test — the race is on for new ways to assess AI. Nature 619, 7971 (2023), 686–689. doi:10.1038/d41586-023-02361-7DOI
  8. [8]Bloomberg News. 2026. ChatGPT Can Now Access and Send Messages in Apple’s iMessage App. Bloomberg. https://www.bloomberg.com/news/articles/2026-08-20/chatgpt-can-now-control-imessage-potentially-raising-apple-privacy-concerns Accessed on Sep 1, 2026.
  9. [9]Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2022. On the Opportunities and Risks of Foundation Models. arXiv:2108.07258 [cs.LG] https://arxiv.org/abs/2108.07258
  10. [10]Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 1877–1901.arxiv.org/abs/2005.14165
  11. [11]Jed R. Brubaker, Meredith Ringel Morris, Dylan Thomas Doyle, Casey Fiesler, Martin Gibbs, and Joanna McGrenere. 2024. AI and the Afterlife. In Extended Abstracts of the 2024 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, Article 458. doi:10. Elizabeth E Bruch and Mark EJ Newman. 2018. Aspirational pursuit of mates in online dating markets. Science Advances 4, 8 (2018), eaap9815.
  12. [13]Heewon Chang. 2016. Autoethnography as method. Routledge.
  13. [14]Camille Cobb, Sruti Bhagavatula, Kalil Anderson Garrett, Alison Hoffman, Varun Rao, and Lujo Bauer. 2021. “I would have to evaluate their objections”: Privacy tensions between smart home device owners and incidental users. Proceedings on Privacy Enhancing Technologies 2021, 4 (2021), 54–75. doi:10.2478/popets-2021-0060DOI
  14. [15]Andrew L. Cohen, Debra Cash, and Michael J. Muller. 2000. Designing to support adversarial collaboration. In Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work (Philadelphia, Pennsylvania, USA). Association for Computing Machinery, New York, NY, USA, 31–39. doi:10.1145/358916.358948
  15. [16]Kenneth Mark Colby, Franklin Dennis Hilf, Sylvia Weber, and Helena C Kraemer. 1972. Turning-like indistinguishability tests for the validation of a computer simulation of paranoid processes. Artificial Intelligence 3 (1972), 199–221. doi:10.1016/0004-3702(72)90049-5DOI
  16. [17]Sally Jo Cunningham and Matt Jones. 2005. Autoethnography: a tool for practice and education. In Proceedings of the 6th ACM SIGCHI New Zealand Chapter’s International Conference on Computer-Human Interaction: Making CHI Natural (Auckland, New Zealand) (CHINZ ’05). Association for Computing Machinery, New York, NY, USA, 1–8. doi:10.1145/1073943.1073944DOI
  17. [18]Michael Davern, Rene Bautista, Jeremy Freese, Pamela Herd, and Stephen L. Morgan. 2025. General Social Survey 1972-2024. Data accessed from the GSS Data Explorer website at gssdataexplorer.norc.org. Principal Investigator, Michael Davern; Co-Principal Investigators, Rene Bautista, Jeremy Freese, Pamela Herd, and Stephen L. Morgan. Sponsored by National Science Foundation.
  18. [19]Yi Dong, Ronghui Mu, Gaojie Jin, Yi Qi, Jinwei Hu, Xingyu Zhao, Jie Meng, Wenjie Ruan, and Xiaowei Huang. 2024. Position: Building Guardrails for Large Language Models Requires Systematic Design. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 11375–11394. https://proceedings.mlr.press/v235/dong24c.htmlDOI
  19. [20]Dylan Thomas Doyle and Jed R. Brubaker. 2023. Digital Legacy: A Systematic Literature Review. Proc. ACM Hum.-Comput. Interact. 7, CSCW2, Article 268 (Oct. 2023), 26 pages. doi:10.1145/3610059DOI
  20. [21]Jenna Drenten and Gillian Brooks. 2020. Celebrity 2.0: Lil Miquela and the rise of a virtual and synthetic persona. Feminist Media Studies 20, 8 (2020), 1319–1323. doi:10.1080/14680777.2020.1830927DOI
  21. [22]Carolyn Ellis, Tony E. Adams, and Arthur P. Bochner. 2010. Autoethnography: An Overview. Forum Qualitative Sozialforschung / Forum: Qualitative Social Research 12, 1 (Nov. 2010). doi:10.17169/fqs-12.1.1589DOI
  22. [23]Fortune. 2026. ChatGPT can now search Apple Messages, raising privacy concerns. Fortune. https://fortune.com/2026/08/26/chatgpt-apple-messages-plugin-privacy-security/ Accessed on Sep 1, 2026.
  23. [24]Robert M. French. 2000. The Turing Test: the first 50 years. Trends in Cognitive Sciences 4, 3 (2000), 115–122. doi:10.1016/S1364-6613(00)01453-4DOI
  24. [25]GDPR. 2024. GDPR. https://gdpr-info.eu Accessed on Nov 12, 2024.
  25. [26]GitHub. 2024. GitHub: message-exporter. https://github.com/ReagentX/message-exporter Accessed on Nov 12, 2024.
  26. [27]Lewis R. Goldberg. 1992. The development of markers for the big-five factor structure. Psychological Assessment 4 (1992), 26–42. Issue 1. doi:10.1037/1040-3590.4.1.26DOI
  27. [28]Bernardo Gonçalves. 2023. The Turing Test is a Thought Experiment. Minds and Machines 33, 1 (March 2023), 1–31. doi:10.1007/s11023-022-09616-8DOI
  28. [29]Google. 2025. Gemini API Additional Terms of Service. https://ai.google.dev/gemini-api/terms Effective May 20, 2025. See section on Paid Services.
  29. [30]Google. 2025. Vertex AI. https://cloud.google.com/vertex-ai Accessed on Apr 7, 2025.DOI
  30. [31]Mark S. Granovetter. 1973. The Strength of Weak Ties. Amer. J. Sociology 78, 6 (1973), 1360–1380. http://www.jstor.org/stable/2776392DOI
  31. [32]Michael Grieves and John Vickers. 2016. Digital twin: Mitigating unpredictable, undesirable emergent behavior in complex systems. In Transdisciplinary perspectives on complex systems: New findings and approaches. Springer, 85–113.
  32. [33]Minghao Guo, Ziyi Ye, Wujiang Xu, Xi Zhu, Wenyue Hua, and Dimitris N. Metaxas. 2026. Individual Turing Test: A Case Study of LLM-based Simulation Using Longitudinal Personal Data. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26). Association for Computing Machinery, New York, NY, USA, 3770–3775. doi:10.1145/3805712.3809868DOI
  33. [34]Steffen Herbold, Alexander Trautsch, Zlata Kikteva, and Annette Hautli-Janisz. 2026. LLM-imperator debate contributions are more authentic, relevant and coherent than their original: A representative study using BBC1’s Question Time. PLOS ONE 21, 7 (2026), e0347757. doi:10.1371/journal.pone.0347757
  34. [35]Hinge. 2024. Hinge - How do I request a copy of my personal data? https://hingeapp.zendesk.com/hc/en-us/articles/360011235813-How-do-I-request-a-copy-of-my-personal-data Accessed on Nov 12, 2024.
  35. [36]Mere Hõõg. 2026. “So let’s suppose somebody from, I don’t know, 2009…” Post on X. https://x.com/MarcHaoq/status/2091214302309134726 Accessed on Sep 1, 2026.
  36. [37]Jeremy Howard and Sebastian Ruder. 2018. Universal Language Model Fine-tuning for Text Classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Iryna Gurevych and Yusuke Miyao (Eds.). Association for Computational Linguistics, Melbourne, Australia, 328–339. doi:10.18653/v1/P18-1031
  37. [38]Dhruv Jain, Audrey Desjardins, Leah Findlater, and Jon E. Froehlich. 2019. Autoethnography of a Hard of Hearing Traveler. In Proceedings of the 21st International ACM SIGACCESS Conference on Computers and Accessibility (Pittsburgh, PA, USA) (ASSETS ’19). Association for Computing Machinery, New York, NY, USA, 236–248. doi:10.1145/3308561.3353800DOI
  38. [39]Daniel Jannai, Amos Meron, Barak Lenz, Yoav Levine, and Yoav Shoham. 2023. Human or Not? A Gamified Approach to the Turing Test. arXiv.org (2023). doi:10.48550/ARXIV.2305.20010
  39. [40]Jiaming Ji, Tianyi Qiu, Boyuan Chen, Jiayi Zhou, Borong Zhang, Donghai Hong, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Lukas Viering, Zhaowei Zhang, Fanzhi Zeng, Juntao Dai, Xuehai Pan, Hua Xu, Aidan O’Gara, Kwan Yee Ng, Brian Tse, Jie Fu, Stephen McAleer, Yanfeng Wang, Mingchuan Yang, Yunhuai Liu, Yizhou Wang, Song-Chun Zhu, Yike Guo, Yaodong Yang, and Wen Gao. 2026. AI Alignment: A Contemporary Survey. ACM Comput. Surv. 58, 5 (2026), 1–38. doi:10.1145/3770749DOI
  40. [41]Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. Deep Learning for Text Style Transfer: A Survey. Computational Linguistics 48, 1 (March 2022), 155–205. doi:10.1162/coli_a_00426DOI
  41. [42]O. P. John and S. R. Srivastava. 1999. The Big Five trait taxonomy: History, measurement, and theoretical perspectives. In Handbook of Personality: Theory and Research (2nd ed.), L. A. Pervin and O. P. John (Eds.). Guilford Press, 102–138.
  42. [43]John A. Johnson. 2014. Measuring thirty facets of the Five Factor Model with a 120-item public domain inventory: Development of the IPIP-NEO-120. Journal of Research in Personality 51 (2014), 78–89. doi:10.1016/j.jrp.2014.05.003DOI
  43. [44]Cameron Jones and Ben Bergen. 2024. Does GPT-4 pass the Turing test? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Kevin Duh, Helena Gomez, and Steven Bethard (Eds.). Association for Computational Linguistics, Mexico City, Mexico, 5183–5210. doi:10.18653/v1/2024.naacl-long.290DOI
  44. [45]Cameron R. Jones and Benjamin K. Bergen. 2026. Large language models pass a standard three-party Turing test. Proceedings of the National Academy of Sciences 123, 21 (2026), e2524472123. doi:10.1073/pnas.2524472123DOI
  45. [46]Philipp Koehn. 2009. Statistical machine translation. Cambridge University Press.
  46. [47]Derek A. Kraeger, Shannon E. Cavanagh, John Yen, and Mo Yu. 2014. “Where Have All the Good Men Gone?” Gendered Interactions in Online Dating. Journal of Marriage and Family 76, 2 (2014), 387–410. doi:10.1111/jomf.12072
  47. [48]Judith C. Lapadat. 2017. Ethics in Autoethnography and Collaborative Autoethnography. Qualitative Inquiry 23, 8 (2017), 589–603. doi:10.1177/1077800417704462DOI
  48. [49]Joanne Leong, John Tang, Edward Cutrell, Sasa Junuzovic, Gregory Paul Baribault, and Kori Inkpen. 2024. Dittos: Personalized, embodied agents that participate in meetings when you are unavailable. Proceedings of the ACM on Human-Computer Interaction 8, CSCW2 (2024), 1–28.DOI
  49. [50]Timothy R. Levine. 2014. Truth-Default Theory (TDT): A Theory of Human Deception and Deception Detection. Journal of Language and Social Psychology 33, 4 (2014), 378–392.DOI
  50. [51]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems 33 (2020), 9459–9474.arxiv.org/abs/2005.11401
  51. [52]Kaiqiu Liang, Julia Kruk, Shengyi Qian, Xianjun Yang, Shengjue Bi, Yuanshun Yao, Shaoliang Nie, Mingyang Zhang, Lijuan Liu, Jaime Fernández Fisac, Shuyan Zhou, and Saghar Hosseini. 2026. Learning Personalized Agents from Human Feedback. arXiv preprint arXiv:2602.16173 (2026).arxiv.org/abs/2602.16173
  52. [53]Chen Ling, Xujiang Zhao, Jiayi Lu, Chengyuan Deng, Can Zheng, Junxiang Wang, Tanmoy Chowdhury, Yun Li, Hejie Cui, Xu Chao Zhang, Tianjiao Zhao, Amit Panalkar, Dhagash Mehta, Stefano Pasquali, Wei Cheng, Haoyu Wang, Yanchi Liu, Zhenghang Chen, Haifeng Chen, Chris White, Quanquan Gu, Jian Pei, Carl Yang, and Liang Zhao. 2026. Domain Specialization as the Key to Make Large Language Models Disruptive: A Comprehensive Survey. ACM Comput. Surv. 58, 3, Article 79 (Feb. 2026), 39 pages. doi:10.1145/3764579DOI
  53. [54]Andrés Lucero, Audrey Desjardins, Carmen Neustaedter, Kristina Höök, Marc Hassenzahl, and Marta E. Cecchinato. 2019. A Sample of One: First-Person Research Methods in HCI. In Companion Publication of the 2019 Designing Interactive Systems Conference (DIS). ACM, 385–388. doi:10.1145/3301019.3319996DOI
  54. [55]Xiaojun Ma, Emily Yang, and Pascale Fung. 2019. Exploring Perceived Emotional Emotional Intelligence of Personality-Driven Virtual Agents in Handling User Challenges. In The World Wide Web Conference. ACM, San Francisco CA USA, 1222–1233. doi:10.1145/3308558.3313400DOI
  55. [56]Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a Large Annotated Corpus of English: The Penn Treebank. Computational Linguistics 19, 2 (1993), 313–330. https://aclanthology.org/J93-2004
  56. [57]Peter V. Marsden and Karen E. Campbell. 1984. Measuring Tie Strength. Social Forces 63, 2 (Dec. 1984), 482–501. doi:10.1093/sf/63.2.482
  57. [58]Meta. 2024. Meta - Download Your Information. https://www.facebook.com/help/836165424150596 Accessed on Nov 12, 2024.
  58. [59]Meredith Moore. 2025. Executive Dysfunction by Design: A Cognitive Accessibility Analysis of AI Support vs. Healthcare Barriers. In Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility (Denver, CO, USA) (ASSETS ’25). Association for Computing Machinery, New York, NY, USA, 1–5. doi:10.1145/3663547.3749831DOI
  59. [60]Steve Moraco. 2026. “I’m actually pretty upset Apple allowed this...” Post on X. https://x.com/SteveMoraco/status/2091194172917338520 Accessed on Sep 1, 2026.
  60. [61]Meredith Ringel Morris and Jed R. Brubaker. 2025. Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 536, 14 pages. doi:10.1145/3706598.3713758DOI
  61. [62]Ani Nenkova and Kathleen McKeown. 2012. A Survey of Text Summarization Techniques. In Mining Text Data. Springer, 43–76. doi:10.1007/978-1-4614-3223-4_3
  62. [63]Helen Nissenbaum. 2004. Privacy as Contextual Integrity. Washington Law Review 79, 1 (2004), 119–157.
  63. [64]OpenClaw Contributors. 2025. OpenClaw — Personal AI Assistant. https://openclaw.ai/. Accessed: 2026-03-01.
  64. [65]OpenClaw Contributors. 2025. OpenClaw: Open-Source Personal Agent Framework. https://github.com/openclaw/openclaw. Accessed: 2026-03-01.
  65. [66]Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Shushani Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35 (2022), 27730–27744.arxiv.org/abs/2203.02155
  66. [67]R. Parasuraman, T.B. Sheridan, and C.D. Wickens. 2000. A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans 30, 3 (2000), 286–297. doi:10.1109/3468.844354DOI
  67. [68]Joon Sung Park, Joseph O’Brien, Carrieu Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (San Francisco, CA, USA). Association for Computing Machinery, New York, NY, USA, Article 2. doi:10.1145/3586183.3606763DOI
  68. [69]Joon Sung Park, Lindsay Popowski, Carrie Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2022. Social Simulacra: Creating Populated Prototypes for Social Computing Systems. In Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology (UIST ’22). Association for Computing Machinery, New York, NY, USA, Article 74, 18 pages. doi:10.1145/3526113.3545616DOI
  69. [70]Jianyi Peng, Carolyn Q. Zhou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Percy Liang, Robb Willer, and Michael S. Bernstein. 2026. LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv:2411.10109v3 [cs.AI]. https://arxiv.org/abs/2411.10109v3
  70. [71]Tianyi Peng, Melanie Brucks, George Gui, Daniel J. Merlau, Grace Jiarui Fan, Malek Ben Slimane, Eric J. Johnson, Abdullah Althneyyan, Sylvia Bellezza, Dante Donati, Hortense Fong, Elizabeth Friedman, Ariana Guevara, Mohamed Hussein, Kinshuk Jerath, Bruce Kogut, Akshit Kumar, Kristen Lane, Hannah Li, Vicki Morwitz, Oded Netzer, Patryck Perkowski, and Olivier Toublia. 2026. Digital twins are funhouse mirrors: Five systematic distortions. Science Advances 12, 36 (2026), eaeh8260. doi:10.1126/sciadv.aeh8260DOI
  71. [72]Samuel Pepys. 1893. The Diary of Samuel Pepys, Transcribed from the Shorthand Manuscript in the Pepysian Library, Magdalene College Cambridge (Unabridged). G. Bell & sons. Transcribed by Mynors Bright; edited with additions by Henry B. Wheatley.
  72. [73]Traian Rebedea, Razvan Dinu, Makesh Narsimhan Sreedhar, Christopher Parisien, and Jonathan Cohen. 2023. NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Yansong Feng and Els Lefever (Eds.). Association for Computational Linguistics, Singapore, 431–445. doi:10.18653/v1/2023.emnlp-demo.40DOI
  73. [74]Machel Reid, Nikolay Savinov, Denis Teplyashin, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024).arxiv.org/abs/2403.05530
  74. [75]Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 7370–7392. doi:10.18653/v1/2024.acl-long.399DOI
  75. [76]Vinay Samuel, Henry Peng Zou, Yue Zhou, Shreyas Chaudhari, Ashwin Kalyan, Tanmay Rajpurihot, Ameet Deshpande, Karthik R Narasimhan, and Vishwak Murahari. 2025. PersonaGym: Evaluating Persona Agents and LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2025, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computational Linguistics, Suzhou, China, 6999–7022. doi:10.18653/v1/2025.findings-emnlp.368DOI
  76. [77]Ayse Pinar Saygin, Isil Cicekli, and Varol Akman. 2000. Turing Test: 50 Years Later. Minds and Machines 10, 4 (2000), 463–518. doi:10.1023/A:1011288000451
  77. [78]John R. Searle. 1969. Speech Acts: An Essay in the Philosophy of Language.
  78. [79]John R. Searle. 1980. Minds, brains, and programs. Behavioral and Brain Sciences 3, 3 (1980), 417–424. doi:10.1017/S0140525X00005756DOI
  79. [80]Gregory Serapio-García, Mustafa Safdari, Clément Crepy, Luning Sun, Stephen Fitz, A Pei Romero, Marwa Abdulhai, Aleksandra Faust, and Maja Matarić. 2025. A psychometric framework for evaluating and shaping personality traits in large language models. Nature Machine Intelligence 7, 12 (2025), 1954–1968. doi:10.1038/s42256-025-01115-6
  80. [81]Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y. Zhao. 2023. Glaze: Protecting artists from style mimicry by Text-to-Image models. In 32nd USENIX Security Symposium (USENIX Security 23). 2187–2204.arxiv.org/abs/2302.04222
  81. [82]Thomas B Sheridan and William L Verplank. 1978. Human and computer control of undersea teleoperators. Technical Report. Man–Machine Systems Laboratory, Massachusetts Institute of Technology.DOI
  82. [83]Quan Shi, Carlos E. Jimenez, Stephen Dong, Brian Seo, Caden Yao, Adam Kelch, and Karthik Narasimhan. 2025. IMPersona: Evaluating Individual Level LM Impersonation. arXiv preprint arXiv:2504.04332 (2025).
  83. [84]Stuart M. Shieber. 1994. Reasons from a restricted Turing test. Commun. ACM 37, 6 (1994), 70–78. doi:10.1145/175208.175217
  84. [85]Ben Shneiderman. 2020. Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human–Computer Interaction 36, 6 (2020), 495–504.
  85. [86]Efstathios Stamatatos. 2009. A survey of modern authorship attribution methods. Journal of the American Society for Information Science and Technology 60, 3 (2009), 538–556.DOI
  86. [87]Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahu Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023).
  87. [88]Tinder. 2024. Tinder - How do I request a copy of my personal data? https://www.help.tinder.com/hc/en-us/articles/11500562726-How-do-I-request-a-copy-of-my-personal-data Accessed on Nov 12, 2024.
  88. [89]Olivier Toubia, George Z. Gui, Tianyi Peng, Daniel J. Merlau, Ng Li, and Haozhe Chen. 2025. Twin-2K-500: A Data Set for Building Digital Twins of over 2,000 People Based on Their Answers to over 500 Questions. Marketing Science 44, 6 (2025), 1446–1455. doi:10.1287/mksc.2025.0262DOI
  89. [90]Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023).arxiv.org/abs/2302.13971
  90. [91]Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei-Lin Chen, Chao-Wei Huang, Yu Meng, and Yun-Nung Chen. 2024. Two Tales of Persona in LLMs: A Survey of Role-Playing and Personalization. In Findings of the Association for Computational Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 16612–16631. doi:10.18653/v1/2024.findings-emnlp.969DOI
  91. [92]Jillian A Tullis. 2021. Self and others: Ethics in autoethnographic research. In Handbook of autoethnography. Routledge, 101–113.
  92. [93]A. M. Turing. 1950. Computing Machinery and Intelligence. Mind LIX, 236 (1950), 433–460. doi:10.1093/mind/LIX.236.433DOI
  93. [94]United States Congress. 1998. 17 U.S.C. § 512 — Limitations on liability relating to material online. https://uscode.house.gov/view.xhtml?req=(title:17§ction:512edition:prelim) Original enactment; last amended by Pub. L. 115–264 (Music Modernization Act, 2018). Current version maintained by the Office of the Law Revision Counsel, U.S. House of Representatives.
  94. [95]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N G omez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30. Curran Associates, Inc., 5998–6008.
  95. [96]Emily A. Vogels and Colleen McClain. 2023. Key Findings About Online Dating in the U.S. Technical Report. Pew Research Center. https://www.pewresearch.org/short-reads/2023/02/02/key-findings-about-online-dating-in-the-u-s/
  96. [97]Yubo Wang, Xueguang Ma, Ge Zhang, Yunsheng Ni, Abhirami Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. 2024. MMLU-Pro: A more robust and challenging multi-task language understanding benchmark. Advances in Neural Information Processing Systems 37 (2024), 95266–95290.arxiv.org/abs/2406.01574
  97. [98]Joseph Weizenbaum. 1966. ELIZA—a computer program for the study of natural language communication between man and machine. Commun. ACM 9, 1 (Jan. 1966), 36–45. doi:10.1145/365153.365168DOI
  98. [99]WhatsApp. 2025. How to export your chat history. https://faq.whatsapp.com/118041407977245/?cms_platform=iphone Accessed on Apr 6, 2025.
  99. [100]Edwin B. Wilson. 1927. Probable Inference, the Law of Succession, and Statistical Inference. J. Amer. Statist. Assoc. 22, 158 (1927), 209–212.DOI
  100. [101]Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing Dialogue Agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Ryna Y. Gururangan and Yusuke Miyao (Eds.). Association for Computational Linguistics, Melbourne, Australia, 2204–2213. doi:10.18653/v1/P18-1205DOI

Paper details

vlnc:W_txerab87
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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