betaFeedback

WorksEconomics, Econometrics and Finance

Artificial Intelligence Clones

Abstract

Large language models, trained on personal data, are increasingly able to mimic individual personalities. These AI clones'' or AI agents'' have the potential to transform how people search for matches in contexts ranging from marriage to employment. This paper presents a theoretical framework to study the tradeoff between the substantially expanded search capacity of AI representations and their imperfect representation of humans. An individual's personality is modeled as a point in kk-dimensional Euclidean space, and an individual's AI representation is modeled as a noisy approximation of that personality. I compare two search regimes: Under in person search, each person randomly meets some number of individuals and matches to the most compatible among them; under AI-mediated search, individuals match to the person with the most compatible AI representation. I show that a finite number of in-person encounters yields a better expected match than search over infinite AI representations. Moreover, when personality is sufficiently high-dimensional, simply meeting two people in person is more effective than search on an AI platform, regardless of the size of its candidate pool.

Introduction

Recent advances in large language models have brought us closer to a world in which individuals are represented by “AI clones” or “AI delegates” trained on their personal data [36]. This technology has the potential to transform how we search and match over human candidates, particularly in settings where direct engagement is costly. Labor markets are already moving in this direction: Roughly 90% of employers already use an automated system to screen candidates [3], with four in ten employers reporting in a survey that they intend to use AI to “talk to” candidates [13]. These changes are facilitated by new AI platforms: Replicant enables job candidates to create digital representatives of themselves (see Figure 1), while Ribbon AI and Alex enable firms to automate their job recruiter [39, 44]. A similar shift is underway for dating markets. Bumble’s founder has described a future in which “your dating concierge could go and date for you with other dating concierges” [48], and platforms such as Volar Dating and Teaser AI have already adopted business models explicitly oriented around this approach [33, 29].12

Replicant AI training, clone creation, and smart matching

Figure 1. Replicant helps job candidates create AI clones to represent them in conversations with recruiters and potential employees. Source: Replicant website [38].

Across these domains, the appeal of AI-mediated matching is scale: AI representations enable search over far more candidates than a person could feasibly meet. But representations of complex human personalities based on limited training data are necessarily imperfect. Thus, there is a tradeoff between the platform’s improved scale of search and its reduced fidelity of representation.

This model is inspired by AI representations but applies broadly to search problems with two defining features. First, match quality depends on many attributes that are difficult to measure. Second, human evaluators have an intuitive and holistic understanding of match quality that does not require explicit assessment of those underlying attributes. For example, a recruiter can know that a candidate interviewed well without having precisely assessed the candidate on component dimensions. In such environments, automated search is constrained by imperfect representations of the candidates and the target, while human search is constrained by the cost of direct engagement. The central tradeoff is between the errors of automated search and the capacity constraints of human evaluation.

My results focus on a quantity I call the AI-equivalent sample size, which measures the value of the AI platform. The AI-equivalent sample size is the smallest mm such that if Subject is permitted mm draws in the in-person regime, then no search size advantage on the AI platform can compensate for the error introduced by the AI representations. That is, Subject would rather meet mm individuals in person than search over an infinite number of individuals on the AI platform.

§ 3 presents the main results. The first result, Proposition 1, says that the AI-equivalent sample size is always finite. In other words, for any level of AI approximation error, there exists a finite number of in-person encounters sufficient to outperform any search size advantage offered by the AI platform. To show this, I demonstrate an intuitive but novel monotone-likelihood ratio property: smaller Euclidean distances between AI representations are associated with smaller Euclidean distances between the actual points. Thus the best-case match on the AI platform is an individual whose AI representation perfectly matches the representation of the target. But because these representations imperfectly represent the underlying points, the expected actual distance between them is bounded away from zero. In contrast, under in-person search, the best of mm draws will eventually approach a perfect match, and thus sufficiently many draws must outperform the AI platform.

My second main result, Theorem 1, says that when the number of dimensions of personality is large, then the AI-equivalent sample size is simply two. That is, in-person search yields a higher expected payoff so long as Subject searches over at least two people.

This theorem is a consequence of counterintuitive properties of high dimensional geometry: As the number of dimensions grows large, random points in the unit ball isolate away from one another and amass near the boundary. In the context of my model, this means that each individual becomes increasingly unique in the space of personalities. While the scarcity of good matches raises the value of searching in a large pool, the best match identified on the AI platform is increasingly driven by noise in the representation rather than true compatibility. Theorem 1 shows that the breakdown of the platform’s selection mechanism outweighs the benefits of its expanded search.

The final part of the paper considers an extension in which the AI platform has access to different quantities of data for different individuals. Individuals are categorized into two groups, a “data-rich” and a “data-poor” group, which are differentiated by how well their AI representations mimic them. Despite the groups being otherwise identical in characteristics, I show that Subject is matched to a data-rich individual with probability exceeding one-half. Moreover, this probability converges to 1 as either the number of dimensions, or the disparity in estimation errors, grows large. These results suggests that use of AI representations may systematically advantage individuals that the AI platform represents more accurately.

The remainder of the paper is organized as follows. § 2 describes the model. § 3 presents the main results. § 4 presents the extension to heterogeneous individuals, and § 5 concludes.

Technological Context

This paper is motivated by recent advances in artificial intelligence that make it possible for algorithms to function as personalized representations of specific individuals. Two developments are central to this shift. First, early evidence suggests that large language models reproduce general patterns of human conversation, and can be further adapted to a particular individual’s style when fine-tuned on their personal data. Second, unlike earlier machine-learning systems that were trained for narrowly defined prediction tasks (such as predicting medical conditions from an image scan), large language models can engage in open-ended interaction both with humans and with one another. Together, these developments make large-scale automated search over human candidates technologically plausible. They are discussed in further detail below.

Digital Representations of Specific Individuals

Large language models are trained on large-scale corpora of text and other media drawn from the internet and related sources. While the first wave of such models adopted generic “helpful assistant” personas,4 researchers and firms are increasingly exploring the possibility that such models can be fine-tuned on personal data—such as written text, social media posts, and recorded speech and video—to produce systems that mimic stable features of a specific individual’s linguistic style, preferences, and judgment patterns.

Several firms already build and market such systems, often described as “AI clones” or “digital twins” of individuals. For example, Delphi AI enables public-facing figures to deploy AI agents that interact with audiences in their distinctive style; Replicant trains AI representatives for job candidates to participate in preliminary screening or interviews on their behalf; and Personal AI allows users to use their written communications to train a system that can, for example, write emails in their voice.

These generative agents are imperfect, but there are indications that they can already approximate some stable individual preferences and decision patterns with meaningful fidelity. For example, Park et al. [36] find that LLM-based agents trained on short personal interviews reproduce individuals’ General Social Survey survey responses and experimental choices with accuracy close to human test–retest reliability. As models improve and richer personal data become available, it is likely that such representations will continue to improve while never achieving perfect accuracy.

Open-Ended Interaction and Scalable Search

Large language models are capable of open-ended communication both with humans and with each other. For example, AI agents can interview one another, simulate collaboration, or engage in conversation on behalf of the individuals they represent. The economic significance of these developments is that interaction itself becomes scalable. Platforms such as Ribbon AI and related ‘talent twin” systems are using conversational agents to conduct preliminary job interviews, and dating services including Volar Dating and Teaser AI are experimenting with AI-mediated exchanges in which users’ agents converse to assess compatibility. Evaluation on these platforms is no longer constrained by human time or attention, allowing them to search across vastly larger pools of potential matches.

At the same time, digital representations are inevitably imperfect, and their fidelity may vary systematically across individuals as a function of data availability, online activity, and willingness to share personal information. § 3 establishes baseline limitations on AI-mediated search as a consequence of this imperfect representation. § 4 subsequently studies environments in which representation quality is heterogeneous and examines how such asymmetries shape outcomes.

Related Literature

This paper contributes to a growing literature on the social implications of artificial intelligence (AI), in particular to research comparing human and AI evaluation. AI predictions have been shown to outperform human experts across various prediction problems [27 Rajpurkar et al.,] 2017; Jung et al., 2017; Angelova et al., 2022; Agarwal et al., 2023). These papers all consider settings—such as medical diagnosis—where humans possess limited intuitive knowledge about the underlying decision problem. By contrast, the present paper is motivated by settings where human actors possess a naturally rich understanding of a complex, subjective objective. This is a novel consideration relative to the literature, and the paper arrives at a very different conclusion regarding algorithmic versus human evaluation.

There is an emerging literature on AI agents. This paper complements recent empirical work about the effectiveness of AI agents (such as Jabarian and Henkel [21] and Sarkar [41]) with a framework for studying the welfare implications of deploying AI agents in search.

Within economic theory, this work contributes to the rich literature on search, which has explored classic questions including how long to search for (McCall [31]; Stigler [43]; Weitzman [47]), what speed to search at (Urgun and Yariv [45]), and where to search (Callander [8]; Malladi [30]).5 The present paper focuses on a new question regarding the role of the complexity of the search space, as measured by the number of attributes (Klabjan et al. [26]). I show that dimensionality fundamentally alters optimal search behavior when search is conducted with error. This result differs from, for example, Bardhi [6] and Malladi [30], who show that their characterizations of optimal search extend but are not qualitatively changed in multiple dimensions.6

This paper’s analysis of the impact of complexity (as measured by the number of dimensions) relates to Ely [14] and Fudenberg and Levine [16]. Ely [14] shows that as an optimization space grows in complexity, decision-makers resort to incremental adaptations (“kludges”) which generate persistent inefficiencies. Fudenberg and Levine [16] examine how a partially-informed agent responds to systemic shocks, and show that the optimal intervention shrinks to zero as the number of dimensions grows large. My finding that higher dimensionality imposes fundamental limits on search aligns with these papers’ high-level insights.

Key to Theorem 1 is a comparison of asymptotics as the dimensionality of the search space grows large. Rate of convergence results have a long history in economic theory, but typically involve limits in the quantity of information [46, 35, 28, 15] or the size of a population [18, 12]. Iakovlev and Liang [20] consider a similar asymptotic to the present paper (namely, as the number of attributes grows large) but characterize limiting beliefs rather than the limiting value of search.

Finally, while this paper adopts the narrative of a dating platform, it diverges considerably both from the classic matching frameworks [17, 40], which examine how centralized mechanisms can achieve stable or efficient outcomes for populations, and from the search and matching literature [42, 9], which analyzes macroeconomic outcomes and equilibrium behavior in decentralized markets with many searchers. This paper instead focuses on the decision problem of a single agent navigating a high-dimensional search space.

Model

Setting

Individuals are represented by vectors in Rk\mathbb{R}^{k}, where each coordinate describes a characteristic.7 A focal individual, hereafter Subject, is searching for a match, which might be a spouse or an employee. Subject’s ideal match or target is normalized to the origin x0:=(0,0,…,0)x_0 := (0, 0, \ldots, 0). Potential matches i≥1i \ge1 have characteristic vectors Xi∈RkX_i \in\mathbb{R}^k drawn uniformly at random from the kk-dimensional unit ball Bk≡{x∈Rk:∥x∥≤1}B_k \equiv\{x \in\mathbb{R}^k : \lVert x\rVert\le1\}, which proxies for a first-round screen where individuals who are too far from Subject’s target are not considered. Subject’s utility from matching with individual ii is

u(Xi)=−∥Xi−x0∥=−∥Xi∥,u(X_i) = -\lVert X_i - x_0\rVert= -\lVert X_i\rVert,

or the negative of the Euclidean distance between individual ii and the target.

There are two search regimes. Under in-person search, Subject randomly meets mm individuals and matches to the most compatible among them.8 The expected distance between the target and this individual is

dkIP(m)≡E[min⁡1≤i≤m∥Xi∥].d_k^{\mathrm{IP}}(m) \equiv\mathbb{E}\left[\min_{1 \le i \le m} \lVert X_i\rVert\right].

Under AI-mediated search, individuals do not interact in person but instead participate on a platform that uses “generators” to stochastically represent the candidates and the target. Specifically, in the interaction between Subject and candidate ii, Subject’s target is represented by

Y0i=x0+ε0i,Y_{0i} = x_0 + \varepsilon_{0i},

and individual ii is represented by

Yi0=Xi+εi0,Y_{i0} = X_i + \varepsilon_{i0},

where ε0i,εi0∼N(0,σ2Ik)\varepsilon_{0i}, \varepsilon_{i0} \sim\mathcal{N}(0, \sigma^2 I_k) are multivariate Gaussian noise terms that are independent of each other and across ii. (The notation IkI_k denotes the identity matrix in kk dimensions, and I assume throughout that σ2>0\sigma^2 > 0.) Figure 2 depicts this model in three dimensions, i.e., k=3k = 3.

Geometric illustration of target and candidate representations

Figure 2. Rather than observing ∥X2−x0∥\lVert X_2-x_0\rVert, the true distance between individual i=2i=2 and the target, the platform observes ∥Y20−Y02∥\lVert Y_{20}-Y_{02}\rVert, the distance between their representations.

There are nn candidates on the platform. For each candidate ii, the platform observes ∥Y0i−Yi0∥\lVert Y_{0i}-Y_{i0}\rVert, the distance between the representation of the target and the representation of candidate ii. The platform recommends the candidate whose representation is closest to the target representation.9

in∗=arg min⁡1≤i≤n ∥Y0i−Yi0∥.i_n^*=\underset{1\leq i\leq n}{\operatorname{arg\,min}}\ \lVert Y_{0i}-Y_{i0}\rVert.

Crucially, although the selected candidate has the most compatible representation, Subject’s payoffs are determined by the actual distance between Subject’s target and this individual. For each sample size nn, define

dkAI(n)≡E[∥Xin∗∥]d_k^{\mathrm{AI}}(n)\equiv\mathbb{E}\left[\lVert X_{i_n^*}\rVert\right]

to be the expected distance between the match in∗i_n^* and the target, in analogy to the previous dkIP(m)d_k^{\mathrm{IP}}(m).

All results in this paper extend for two variations of this model. First, the representations may be fixed, so that Y0Y_0 and YiY_i are generated once and fixed in each pairwise interaction. Second, AI representations may be employed on only one side of the market, i.e., only by the searcher or by the candidates. In each of these cases, the analysis extends with minor modifications.10

Value of AI-Mediated Search

My main results characterize the following measure.

Definition 1. Fix any number of dimensions kk. The AI-equivalent sample size mk∗m_k^* is the smallest m∈Z+m \in\mathbb{Z}_+ satisfying

dkIP(m)<dkAI(n)∀n∈Z+.d_k^{\mathrm{IP}}(m) < d_k^{\mathrm{AI}}(n) \quad\forall n \in\mathbb{Z}_+.

If the inequality above is not satisfied for any finite mm, then mk∗=∞m_k^* = \infty.

The AI-equivalent sample size is the smallest integer mm such that searching over mm individuals in person leads (in expectation) to a better match than participating on the AI platform, no matter the size of the platform’s candidate pool. It serves as a quantitative benchmark for the value of AI-mediated search: An infinite AI-equivalent sample size means that for every mm, one can choose nn large enough such that search on an AI platform with nn candidates outperforms in-person search over mm candidates. A finite AI-equivalent sample size means that the advantage of the AI platform is fundamentally capped, and a sufficiently thorough in-person search will outperform it no matter its size. In this case, the practical value of the platform depends on the size of mk∗m_k^*.

Interpretation of Model

The proposed framework applies to environments in which AI representations are already beginning to be deployed (§ 1.1), such as the following.

Example 2.1 (Hiring). A firm seeks to hire a suitable candidate. Under in-person search, a hiring manager interviews a small number of applicants. Under AI-mediated search, an AI recruiter interviews the AI representations of a substantially larger number of candidates. As discussed earlier, all results extend if an AI representation is only used on one side of this market, i.e., if an AI recruiter interviews humans or if a human recruiter interviews AI representations of candidates.

Example 2.2 (Dating). An individual seeks a partner. Under in-person search, the individual chooses the (truly) most compatible from a limited number mm of candidates. Under AI-mediated search, a platform first seeks to understand the individual’s ideal match, and then evaluates a potentially much larger set of candidates n≫mn \gg m relative to its understanding of the searcher’s target. In the special case where the searcher prefers someone with similar characteristics, we can interpret the target representation as a representation of the searcher themselves.

The framework also extends naturally to settings where AI representations of individuals are not yet in use but plausibly will be, such as clients seeking lawyers, founders seeking investors, or families seeking care providers.

Finally, although this paper is motivated by emerging uses of AI representations, the framework applies more broadly to search problems with two key features. First, match quality depends on many attributes that are individually difficult to measure. Second, the human searcher does not need to explicitly identify or evaluate each relevant attribute, but instead has an intuitive grasp of what constitutes a better or worse outcome.11 For example, a hiring manager can judge that an interview went poorly without having explicitly evaluated the candidate’s communication style, problem-solving approach, or cultural fit. An AI platform, by contrast, lacks direct access to this holistic evaluation and must instead infer quality from noisy measurements of component attributes.12 These features appear in other search and delegation problems as well, such as the following.

Example 2.3 (Purchasing a Home). A family seeks to purchase a home. Under in-person search, the family visits mm houses directly. Alternatively, the family imperfectly communicates its preferences to a system that evaluates a larger set nn of homes relative to its understanding of the target.

Main Results

Section 3.1 presents two results about the limits of AI-mediated search. Section 3.2 outlines the proof of these results. Section 3.3 discusses possible extensions.

Limits of AI-Mediated Search

My first result says that the value of the AI representation regime is inherently capped: For every number of dimensions kk, there is some finite number of in-person encounters that yields a better expected match than the AI platform, no matter how large its candidate pool.

Proposition 1. For every number of dimensions k∈Z+k \in\mathbb{Z}_+, the AI-equivalent sample size mk∗m_k^* is finite.

Intuitively, imperfect representations place a fundamental limit on the quality of search: even with an arbitrarily large platform, the expected true distance between the Subject’s target and the platform’s best-ranked match remains bounded away from zero. In contrast, under in-person search, Subject’s expected distance to the best among mm individuals vanishes as mm grows large. Thus for sufficiently large mm, Subject’s expected match must be better in the in-person regime.

Proposition 1 holds for any number of attributes kk. The next result says that when kk is sufficiently large—so that many attributes are relevant to match quality—the value of AI-mediated search drops sharply.

Theorem 1. For all kk sufficiently large, the AI-equivalent sample size is mk∗=2m_k^* = 2.

Thus, in-person search yields a better expected outcome so long as Subject searches over at least two people. The threshold dimensionality kk at which this reversal occurs depends on the magnitude of AI measurement error σ2\sigma^2, but it remains finite whenever σ2>0\sigma^2 > 0.

The implication is not that AI-mediated search is inherently ineffective, but rather that its usefulness depends critically on the effective dimensionality of match quality. When compatibility depends on a small number of characteristics, large-scale automated search can perform well. By contrast, when many distinct attributes matter, Theorem 1 implies a fundamental limitation to AI-mediated search: Even a very small amount of direct, in-person evaluation yields a better expected outcome than participation on an arbitrarily large AI platform. (This conclusion relies on the relevant attributes not being well approximated by a lower-dimensional set; otherwise, despite many nominal attributes, the effective dimensionality of evaluation would be small.)

The number of relevant attributes kk likely varies across applications. In routine jobs, for instance, match quality may be predictable from a relatively small set of attributes. Other roles, such as selecting a CEO, plausibly involve a substantially larger set of traits. Long-term partner choice may involve an even richer set of relevant attributes [23, 22]—spanning values, personality, emotional compatibility, habits, and life goals—though the effective dimensionality may also vary across individuals. Theorem 1 thus implies that AI-mediated search can be valuable when alignment on a small number of attributes suffices, but that human judgment is indispensable—and ultimately the limiting constraint—when compatibility depends on a diverse, multidimensional set of attributes.

Proof Sketch

The proof of Theorem 1 proceeds in two main steps.

A uniform lower bound for dkAI(n)d_k^{\mathrm{AI}}(n)

Consider an idealized AI platform that continuously samples candidates until encountering one whose representation exactly matches the target representation, i.e., an individual ii for whom ∥Yi0−Y0i∥=0\lVert Y_{i0}-Y_{0i}\rVert=0. Let

dkAI(∞)≡E(∥Xi∥:∥Yi0−Y0i∥=0)d_k^{\mathrm{AI}}(\infty) \equiv\mathbb{E}\left(\lVert X_i\rVert:\lVert Y_{i0}-Y_{0i}\rVert=0\right)

be the expected true distance to such a perfectly matched individual. The following lemma shows that this quantity is a uniform lower bound on dkAI(n)d_k^{\mathrm{AI}}(n) across all nn.

Lemma 1. For all n∈Z+n \in\mathbb{Z}_{+},

dkAI(∞)≤dkAI(n).d_k^{\mathrm{AI}}(\infty) \le d_k^{\mathrm{AI}}(n).

To show this, I demonstrate that the pair of random variables

(∥Xi∥,∥Yi0−Y0i∥)\left(\lVert X_i\rVert,\lVert Y_{i0}-Y_{0i}\rVert\right)

satisfies the monotone likelihood ratio property. That is, conditioning on a smaller representation distance ∥Yi0−Y0i∥\lVert Y_{i0}-Y_{0i}\rVert shifts the conditional distribution of ∥Xi∥\lVert X_i\rVert toward smaller values.13 Consequently, the function

s↦E[∥Xi∥∣∥Yi0−Y0i∥=s]s \mapsto\mathbb{E}\left[\lVert X_i\rVert\mid\lVert Y_{i0}-Y_{0i}\rVert=s\right]

is increasing in ss. Since the selected individual must satisfy ∥Yi0−Y0i∥≥0\lVert Y_{i0}-Y_{0i}\rVert\ge0, the claim follows.

Benchmark comparison

Let dkIP(1)d_k^{\mathrm{IP}}(1) denote the expected distance to a single uniformly drawn individual from the unit ball. It is well known that

dkIP(1)=kk+1.d_k^{\mathrm{IP}}(1)=\frac{k}{k+1}.

I show that both dkIP(2)d_k^{\mathrm{IP}}(2) and dkAI(∞)d_k^{\mathrm{AI}}(\infty) converge to dkIP(1)d_k^{\mathrm{IP}}(1) as kk grows large, but crucially they do so at different rates. Specifically, the expected reduction in distance gained by picking the smaller of two draws rather than one is vanishing on the order of 1/k1/k,

References

  1. [2]Aher, G. V., R. I. Arriaga, and A. T. Kalai (2023): “Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies,” in Proceedings of the 40th International Conference on Machine Learning, ed. by A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, PMLR, vol. 202 of Proceedings of Machine Learning Research, 337–371.arxiv.org/abs/2208.10264
  2. [3]Amitabh, U. and A. Ansari (2025): “Hiring with AI doesn’t have to be so inhumane,” World Economic Forum, accessed 7 Jan 2026.
  3. [5]Athey, S. (2002): “Monotone Comparative Statics under Uncertainty,” The Quarterly Journal of Economics, 117, 187–223.
  4. [6]Bardhi, A. (2023): “Attributes: Selective Learning and Influence,” Working Paper.DOI
  5. [7]Bolton, P. and C. Harris (1999): “Strategic Experimentation,” Econometrica, 67, 349–374.DOI
  6. [8]Callander, S. (2011): “Searching for Good Policies,” American Political Science Review, 105, 643–662.
  7. [9]Chade, H., J. Eeckhout, and L. Smith (2017): “Sorting through Search and Matching Models in Economics,” Journal of Economic Literature, 55, 493–544.DOI
  8. [10]Christiano, P. F., J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei (2017): “Deep Reinforcement Learning from Human Preferences,” in Advances in Neural Information Processing Systems (NIPS), 4302–4310.arxiv.org/abs/1706.03741
  9. [11]Cornell, B. and I. Welch (1996): “Culture, Information, and Screening Discrimination,” Journal of Political Economy, 104, 542–571.
  10. [12]Dasaratha, K., B. Golub, and N. Hak (2023): “Learning from Neighbours about a Changing State,” Review of Economic Studies, 90, 2326–2369.DOI
  11. [13]Demopoulos, A. (2024): “The job applicants shut out by AI: ‘The interviewer sounded like Siri’,” The Guardian, accessed 7 Jan 2026.
  12. [14]Ely, J. C. (2011): “Kludged,” American Economic Journal: Microeconomics, 3, 210–231.DOI
  13. [15]Frick, M., R. Iijima, and Y. Ishii (2024): “Monitoring with Rich Data,” Working paper.
  14. [16]Fudenberg, D. and D. K. Levine (2022): “Intervention with limited information,” International Journal of Game Theory, 51, 379–389, published online 19 October 2021.DOI
  15. [17]Gale, D. and L. S. Shapley (1962): “College Admissions and the Stability of Marriage,” The American Mathematical Monthly, 69, 9–15.
  16. [18]Harel, M., E. Mossel, P. Strack, and O. Tamuz (2021): “Rational Group-think,” The Quarterly Journal of Economics, 136, 621–668, published: 08 July 2020.
  17. [19]Horton, J. J. (2023): “Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?” NBER Working Paper 31122, National Bureau of Economic Research.
  18. [20]Iakovlev, A. and A. Liang (2024): “The Value of Context: Human versus Black Box Evaluators,” Working Paper.arxiv.org/abs/2402.11157
  19. [21]Jabarian, B. and L. Henkel (2025): “Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews,” Working paper.
  20. [22]Joel, S., P. W. Eastwick, C. J. Allison, X. B. Arriaga, Z. G. Baker, E. Barr-Kalif, and et al. (2020): “Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies,” Proceedings of the National Academy of Sciences, 117, 19061–19071.
  21. [23]Joel, S., P. W. Eastwick, and E. J. Finkel (2017): “Is Romantic Desire Predictable? Machine Learning Applied to Initial Romantic Attraction,” Psychological Science, 28, 1478–1489.
  22. [25]Keller, G., S. Rady, and M. Cripps (2005): “Strategic Experimentation with Exponential Bandits,” Econometrica, 73, 39–68.
  23. [26]Klabjan, D., W. Olszewski, and A. Wolinsky (2014): “Attributes,” Games and Economic Behavior, 88, 190–206.DOI
  24. [27]Kleinberg, J., H. Lakkaraju, J. Leskovec, J. Ludwig, and S. Mullainathan (2017): “Human Decisions and Machine Predictions,” The Quarterly Journal of Economics, 133, 237–293.
  25. [28]Liang, A. and X. Mu (2020): “Complementary Information and Learning Traps,” The Quarterly Journal of Economics, 135, 389–448.
  26. [29]Lin, B. (2024): “Grindr Aims to Build the Dating World’s First AI ‘Wingman’,” The Wall Street Journal, accessed: 2025-02-15.
  27. [30]Malladi, S. (2023): “Searching in the Dark and Learning Where to Look,” Working paper, draft date: April 19, 2023.DOI
  28. [31]McCall, J. J. (1970): “Economics of Information and Job Search,” The Quarterly Journal of Economics, 84, 113–126.
  29. [32]Mei, Q., Y. Xie, W. Yuan, and M. O. Jackson (2024): “A Turing test of whether AI chatbots are behaviorally similar to humans,” Proceedings of the National Academy of Sciences, 121, e2313925121.DOI
  30. [33]Metz, C. (2024): “AI Is Changing Dating Apps, But Not in the Way You Think,” The New York Times.
  31. [34]Milgrom, P. R. (1981): “Good News and Bad News: Representation Theorems and Applications,” The Bell Journal of Economics, 12, 380–391.DOI
  32. [35]Moscarini, G. and L. Smith (2002): “The Law of Large Demand for Information,” Econometrica, 70, 2351–2366.
  33. [36]Park, J. S., C. Q. Zou, A. Shaw, B. M. Hill, C. Cai, M. R. Morris, R. Willer, P. Liang, and M. S. Bernstein (2024): “Generative Agent Simulations of 1,000 People,” arXiv, accessed 2025-01-01.
  34. [38]Replicant (2024): “About Replicant: Revolutionizing recruitment through AI-powered candidate replication,” https://replicant.im/about, accessed 2025-01-01.
  35. [39]Rocha, N. (2025): “Welcome to Your Job Interview. Your Interviewer Is A.I.” The New York Times, accessed 2025-01-06.
  36. [40]Roth, A. E. and M. Sotomayor (1992): “Two-Sided Matching,” in Handbook of Game Theory with Economic Applications, Elsevier, vol. 1, 485–541.
  37. [41]Sarkar, S. K. (2025): “AI Agents, Productivity, and Higher-Order Thinking: Early Evidence from Software Development,” Preprint.
  38. [42]Shimer, R. and L. Smith (2000): “Assortative Matching and Search,” Econometrica, 68, 343–369.
  39. [43]Stigler, G. J. (1961): “The Economics of Information,” Journal of Political Economy, 69, 213–225.
  40. [44]Temkin, M. (2025): “AI recruiter Alex raises $17M to automate initial job interviews,” Accessed January 7, 2026.
  41. [45]Urgun, C. and L. Yariv (2024): “Contiguous Search: Exploration and Ambition on Uncharted Terrain,” Journal of Political Economy, 132, published online December 18, 2024.DOI
  42. [46]Vives, X. (1992): “How Fast do Rational Agents Learn?” Review of Economic Studies, 60, 329–347.
  43. [47]Weitzman, M. L. (1979): “Optimal Search for the Best Alternative,” Econometrica, 47, 641–654.
  44. [48]Whittaker, R. (2024): “Would you let AI date for you? Bumble’s founder thinks that could be the future,” https://www.forbes.com.au/news/innovation/bumble-ai-dating-concierge/, forbes Australia, accessed January 2026.
  45. [49]Árpád Baricz (2010): “Bounds for Modified Bessel Functions of the First and Second Kinds,” Proceedings of the Edinburgh Mathematical Society, 53, 575–599.DOI

Paper details

vlnc:W_b5w5ndws
Version
v2 · latest
Published
Added
License
CC BY 4.0
Usage

views · 1

downloads · 5

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

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents