betaFeedback

Reading v1View latest version

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 artificial intelligence—trained on vast collections of written and spoken communications—can convincingly mimic individual personalities. This development has the potential to transform search over human candidates in various contexts. Several dating platforms have already begun developing “AI clones” of individual users to simulate dialogues between potential matches [23],12 and the conceivable applications of AI clones extend to many other domains as well. For example, organizations could employ AI clones of recruiters and applicants to simulate interviews with a large pool of candidates, or a placement agency—such as one that matches families to caregivers—might rely on AI clones to assess whether a prospective caregiver fits with a family. These examples highlight the potential for artificial intelligence to reshape how people identify, evaluate, and ultimately choose one another across a wide array of personal and professional settings.

Relative to in-person interactions, the primary appeal of AI clones is their capacity to search across a large pool of potential matches. However, unlike actual clones that replicate an individual precisely, AI clones must approximate their subject’s personality from limited training data. This challenge is compounded by the complexity of the target: a complete human personality.

This paper examines whether the search advantage offered by AI representations outweighs the inaccuracies they introduce. To address this question, I develop a model in which individuals are represented as vectors in a finite-dimensional Euclidean space, where each element of this vector is an attribute of personality (broadly construed). Individuals prefer to be matched to others close to them in this space, i.e., to another with similar personality attributes. I consider two scenarios from the perspective of an individual referred to as “Subject”: in the first scenario, the in-person regime, Subject meets mm randomly selected individuals and matches to the closest of these individuals; in the second scenario, the AI representation regime, Subject’s match is the individual whose AI clone is most similar to Subject’s own AI clone. I show that there exists a finite mm such that meeting mm individuals in person yields a better expected match than the AI platform, even when the platform searches over infinitely many candidates. Moreover, when the dimensionality of personality is large, this mm equals two. In other words, as personality complexity increases, simply meeting two people in person produces a higher expected match quality than entrusting the process to an AI platform, regardless of the size of its candidate pool.

In more detail: I model individuals as vectors in kk-dimensional Euclidean space, where Subject is placed at the zero vector and all other individuals are drawn uniformly from the unit ball. In the in-person regime, Subject randomly meets mm individuals and observes their attributes. Subject matches to the closest of these mm individuals, and receives a payoff equal to the negative of the Euclidean distance between them. In the AI representation regime, all individuals are represented by AI clones, which are imperfect approximations of the individuals’ personalities. Specifically, an individual’s AI clone is the individual’s true personality vector perturbed by a multivariate Gaussian noise term. The AI clones interact, and the match quality between these clones is determined. Subject is matched to the individual whose AI clone is most compatible to Subject’s AI clone, but Subject’s payoff is the negative of the true distance between them.

This model is inspired by AI clones, but can be interpreted as more generally describing situations in which the search objective and the candidates are too complex to be completely specified. In these settings there are inherent limits to how well search can be automated, since each candidate must be evaluated with respect to the system’s imperfect understanding of that candidate and the objective. In contrast, an actor who has internalized a rich, nuanced understanding of what constitutes a “good” match—be it a romantic partner, a new hire, or any other complex target—can more accurately evaluate each candidate. Since a human actor can only evaluate a limited number of candidates in this way, the key comparison is between the capacity constraints of human evaluation and the approximation errors of automated search.

My results pertain to 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 advantage given to the AI representation regime can compensate for the error introduced by the AI clones. That is, Subject would rather meet mm individuals in person than search over an infinite number of individuals on the AI platform.

My first main 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 advantage offered by the AI platform. Intuitively, as the number of individuals on the AI platform grows large, the individual identified as the best match will have an AI clone that (nearly) perfectly matches Subject’s AI clone. But because these clones imperfectly represent their human counterparts, there is always some expected mismatch between the underlying humans—even when the AI clones are perfectly compatible. In contrast, in the in-person regime, the best of mm draws will eventually approach a perfect match, and thus sufficiently many such draws must outperform the AI platform.

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

To understand this result, we need to consider the role of the dimensionality of the personality space. High-dimensional spaces exhibit several interesting properties: Random points in the unit ball tend to be far from one another, and the distances between them concentrate around their expected values. In the context of my model, this means that each individual becomes increasingly unique in the personality space. This scarcity of nearby individuals creates two opposing effects. On the one hand, searching through more candidates has greater value, since doing so increases the chance of finding someone compatible—an advantage that favors the AI representation regime. At the same time, the increase in complexity of the learning problem means that the errors of the AI representation regime are more consequential. Theorem 1 says that this latter effect dominates.

Formally, I show that as the number of dimensions grows large, Subject’s expected payoff in the AI representation regime converges to the benchmark payoff from being matched to a single randomly drawn individual. Loosely speaking, this is because the problem of estimation in high dimensions becomes so difficult that the match identified on the AI platform is driven by the noise in the estimation, rather than by actual compatibility. At the same time, the single random draw benchmark is also the limiting expected payoff from being matched to the better of two draws in the in-person regime. The key step towards proving the result is thus to show that the benefit provided by the AI platform erodes more rapidly than the benefit of having a second in-person draw.

In the final part of the paper, I explore an extension in which the AI platform has access to different quantities of data for different individuals, reflecting heterogeneity in data availability due to factors like privacy preferences or digital presence. Individuals are categorized into two groups, a “data-rich” and “data-poor” group, which are differentiated by the quality of AI estimation. Despite the groups being otherwise identical in characteristics, I show that Subject is matched to a data-rich individual with probability exceeding one-half, and moreover that this probability converges to 1 as either the dimensionality of analysis increases or the disparity in estimation errors between groups widens. These results suggest a new form of social stratification, where individuals’ outcomes depend not just on their intrinsic characteristics but also on how well they are understood by AI systems.

The remainder of the paper is organized as follows. Section 2 describes the model. Section 3 considers the simple case of a single dimension. Section 4 presents the main results. Section 5 presents the extension to heterogenous individuals, and Section 6 concludes.

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 ([26]; [16]; [1]; [19]; [2]); [15] provide a theoretical rationale for why this might be so. 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, which is difficult to define formally. This is a novel consideration relative to the literature, and the paper arrives at a very different conclusion regarding algorithmic versus human evaluation.

This work also contributes to the rich literature on search, which has explored classic questions including how long to search for ([29], [22], [32]), what speed to search at ([30]), and where to search ([8], [21]).3 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 ([18]). I show that dimensionality fundamentally alters optimal search behavior when search is conducted with error. This result contrasts with, for example, [5] and [21], who show that their characterizations of optimal search extend but are not qualitatively changed in multiple dimensions.4

Key to Theorem 1 is a comparison of asymptotics as the dimensionality of the search space grows large. Rates of convergence results have a long history in economic theory, but typically involve limits in the quantity of information ([31], [20], [12]) or the size of a population ([14], [11]). [15] consider a similar asymptotic to the present paper (i.e., with respect to an increasing number of attributes) 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 ([13], [27]), which examine how centralized mechanisms can achieve stable or efficient outcomes for populations, and from the search and matching literature ([28], [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

Each individual ii is described by kk real-valued personality attributes.5 I consider the perspective of a specific individual, subsequently ‘Subject,’ whose attributes are fixed at the zero vector x0x_0. Every other individual Xi∈RkX_i \in\mathbb{R}^k, i≥1i \ge1, is 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\}.6 Subject receives utility −∥Xi−x0∥=−∥Xi∥-\lVert X_i-x_0\rVert= -\lVert X_i\rVert from being matched to individual ii, i.e., the negative of the Euclidean distance between them. I will use the phrases “most similar individual,” “closest individual,” and “most compatible individual” interchangeably. (See Section 4.1 for discussion of other notions of match quality.)

There are two regimes. In the first, which I call the in-person regime, Subject randomly meets mm individuals and matches to the closest individual.7 The expected distance to this individual in kk dimensions is

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

and Subject’s expected payoff is −dkIP(m)-d_k^{\mathrm{IP}}(m).

In the second regime, which I call the AI representation regime, individuals do not interact in person; instead, they elect to have an AI clone represent them, and these clones interact with one another. Each clone is a random imperfect mimic of each individual, and match quality is estimated based on compatibility of the realized AI clones. Formally, in an interaction between Subject and individual ii, Subject is represented by

Y0i=x0+ε0iY_{0i} = x_0 + \varepsilon_{0i}

and individual ii is represented by

Yi0=Xi+εi0Y_{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.8 (The notation IkI_{k} denotes the identity matrix in kk dimensions.) See Figure 1 for a depiction of the model in three dimensions, i.e., k=3k = 3.

Depiction of an interaction between Subject’s AI clone and individual 2’s AI clone in $k = 3$ dimensions

Figure 1. Depiction of an interaction between Subject’s AI clone and individual 2’s AI clone in k=3k = 3 dimensions.

The AI platform searches over nn possible matches for Subject. Among these, the individual with the most compatible AI clone is

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

Subject’s expected actual distance to this individual is

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

I again make the number of dimensions kk explicit, as this is a parameter that we will vary.

While the dating platform serves as the main narrative throughout the paper, this model can also be used to consider other uses of AI clones, such as the following.

Example (Job Hiring). A firm seeks a suitable candidate from a large pool of applicants. In the in-person regime, the hiring manager interviews some mm candidates in person. In the AI representation regime, an AI clone of the hiring manager conducts interviews with the AI clones representing some nn candidates.

Example (Family-to-Caregiver Matching). A family seeks to hire a caregiver. In the in-person regime, the family meets with several candidates. In the AI representation regime, AI clones of the family members interact with AI clones of the potential caregivers.

Example (Collaboration). An engineer seeks a collaborator to manage the business side of a startup. In the in-person regime, the engineer interacts with potential collaborators at networking events. In the AI representation regime, AI clones of the engineer and potential collaborators interact and are assessed on joint collaborative potential.

Beyond AI clones, we can think of the model as more generally describing situations where both the objects of search as well as the desired objective are complex.

Example (Purchasing a Home). A family seeks to purchase a home. In the in-person regime, the family visits a small number of houses in person. In the AI representation regime, the family’s preferences are communicated as best possible to an AI system, which then (imperfectly) evaluates many homes with respect to this objective.

The key point is that there are tasks that are simple for humans but challenging for machines. A human knows intuitively that they prefer one person’s company over another, and that they feel more at home in one house than another, even when the underlying reasons are complex and challenging to specify for an algorithm.9 Moreover, the sheer number of subtle factors—such as whether someone’s sense of humor is dry or dark—implies that even if a machine fully understood the objective function, it would still need a profound understanding of each candidate to make an accurate assessment. Ultimately, there is a tradeoff between a fast automated approach that produces less accurate evaluations and a more resource-intensive method that relies on knowledgeable human judgment to yield precise assessments.

My results pertain to a quantity I call the sample size equivalent of the idealized AI representation regime with n=∞n = \infty.

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

E[dkIP(m)]<E[dkAI(n)]∀n∈Z+.\mathbb{E}\left[d_k^{\mathrm{IP}}(m)\right] < \mathbb{E}\left[d_k^{\mathrm{AI}}(n)\right]\qquad\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 mk∗m_k^* provides a measure of the value of the AI platform. If mk∗m_k^* is infinite, then every in-person search can be beat by an AI platform with a large enough candidate pool. If mk∗m_k^* is instead finite, then the advantage of the AI platform is fundamentally capped, and a sufficiently thorough in-person search will outperform it. In this case, a small mk∗m_k^* suggests that even a small number of in-person encounters can outperform the AI platform’s infinite search capacity, while a larger (but still finite) mk∗m_k^* indicates that the AI platform retains a noticeable advantage before in-person search can catch up. In this way, mk∗m_k^* serves as a practical yardstick for determining when and how human efforts can prevail over algorithmic methods.

Warm-Up: One Dimension

To build intuition, let us first consider the one-dimensional case. Subject is located at zero, and all other individuals in the population are drawn uniformly from [−1,1][-1, 1]. For any distance r∈[0,1]r \in[0, 1],

P(min⁡1≤i≤m∣Xi∣>r)=P(∣X1∣>r)m=(1−r)mP\left(\min_{1 \le i \le m} |X_i| > r\right) = P(|X_1| > r)^m = (1-r)^m

since XiX_i are i.i.d. So Subject’s expected distance to the best match among mm draws in the in-person regime is

d1IP(m)=E[min⁡1≤i≤m∣Xi∣]=∫01P(min⁡1≤i≤m∣Xi∣>r) dr=∫01(1−r)mdr=1m+1.d_1^{\mathrm{IP}}(m) = \mathbb{E}\left[\min_{1 \le i \le m} |X_i|\right] = \int_0^1 P\left(\min_{1 \le i \le m} |X_i| > r\right)\,dr = \int_0^1 (1-r)^m dr = \frac{1}{m+1}.

Now consider the AI representation regime. In the proof of Proposition 1, I show that Subject’s expected payoff given nn samples can be lower bounded by Subject’s expected payoff in an “infinite nn” setting in which ∣Yi0−Y0i∣=0|Y_{i0}-Y_{0i}|=0 is achieved for some index ii, i.e., there is some AI clone that perfectly matches Subject’s AI clone. Conditional on the event {∣Yi0−Y0i∣=0}\{|Y_{i0}-Y_{0i}|=0\}, the random variable XiX_i (the actual location of individual ii) is independent of all other XjX_j and Yj0Y_{j0} (the actual locations and clone realizations of the other individuals). Thus we can write Subject’s expected payoff as

dkAI(∞)=E(∣Xi∣:∣Yi0−Y0i∣=0).d_k^{\mathrm{AI}}(\infty)=\mathbb{E}\left(|X_i|:|Y_{i0}-Y_{0i}|=0\right).

To characterize this conditional expectation, observe that the distance between Subject’s AI clone, Y0iY_{0i}, and the AI clone of individual ii, Yi0Y_{i0}, is

∣Yi0−Y0i∣=∣Xi+εi0−ε0i∣=∣Xi+Zi∣|Y_{i0}-Y_{0i}|=|X_i+\varepsilon_{i0}-\varepsilon_{0i}|=|X_i+Z_i|

where Zi≡εi0−ε0i∼N(0,2σ2)Z_i\equiv\varepsilon_{i0}-\varepsilon_{0i}\sim\mathcal{N}(0,2\sigma^2). So ∣Yi0−Y0i∣=0|Y_{i0}-Y_{0i}|=0 implies Zi=−XiZ_i=-X_i. That is, if the ii-th AI clone perfectly matches Subject’s AI clone, it must be that the noise term ZiZ_i exactly undoes the true distance between Subject and individual ii.

By independence of XiX_i and ZiZ_i, and symmetry of ZiZ_i, we have

References

  1. [1]Agarwal, N., A. Moehring, P. Rajpurkar, and T. Salz (2023): “Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology,” Working Paper 31422, National Bureau of Economic Research.
  2. [2]Angelova, V., W. Dobbie, , and C. S. Yang (2022): “Algorithmic Recommendations and Human Discretion,” Working Paper.
  3. [3]Athey, S. (2002): “Monotone Comparative Statics under Uncertainty,” The Quarterly Journal of Economics, 117, 187–223.
  4. [5]Bardhi, A. (2023): “Attributes: Selective Learning and Influence,” Working Paper.DOI
  5. [6]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]Dasaratha, K., B. Golub, and N. Hak (2023): “Learning from Neighbours about a Changing State,” Review of Economic Studies, 90, 2326–2369.DOI
  10. [12]Frick, M., R. Iijima, and Y. Ishii (2024): “Monitoring with Rich Data,” Working paper.
  11. [13]Gale, D. and L. S. Shapley (1962): “College Admissions and the Stability of Marriage,” The American Mathematical Monthly, 69, 9–15.
  12. [14]Harel, M., E. Mossel, P. Strack, and O. Tamuz (2021): “Rational Groupthink,” The Quarterly Journal of Economics, 136, 621–668, published: 08 July 2020.
  13. [15]Iakovlev, A. and A. Liang (2024): “The Value of Context: Human versus Black Box Evaluators,” Working Paper.arxiv.org/abs/2402.11157
  14. [16]Jung, J., C. Concannon, R. Shroff, S. Goel, and D. G. Goldstein (2017): “Simple rules for complex decisions,” Working Paper.arxiv.org/abs/1702.04690
  15. [17]Keller, G., S. Rady, and M. Cripps (2005): “Strategic Experimentation with Exponential Bandits,” Econometrica, 73, 39–68.
  16. [18]Klabjan, D., W. Olszewski, and A. Wolinsky (2014): “Attributes,” Games and Economic Behavior, 88, 190–206.DOI
  17. [19]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.
  18. [20]Liang, A. and X. Mu (2019): “Complementary Information and Learning Traps*,” The Quarterly Journal of Economics, 135, 389–448.
  19. [21]Malladi, S. (2023): “Searching in the Dark and Learning Where to Look,” Working paper, draft date: April 19, 2023.DOI
  20. [22]McCall, J. J. (1970): “Economics of Information and Job Search,” The Quarterly Journal of Economics, 84, 113–126.
  21. [23]Metz, C. (2024): “AI Is Changing Dating Apps, But Not in the Way You Think,” The New York Times.
  22. [24]Milgrom, P. R. (1981): “Good News and Bad News: Representation Theorems and Applications,” The Bell Journal of Economics, 12, 380–391.DOI
  23. [26]Rajpurkar, P., J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, M. P. Lungren, and A. Y. Ng (2017): “CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning,” Working Paper.arxiv.org/abs/1711.05225
  24. [27]Roth, A. E. and M. Sotomayor (1992): “Two-Sided Matching,” in Handbook of Game Theory with Economic Applications, Elsevier, vol. 1, 485–541.
  25. [28]Shimer, R. and L. Smith (2000): “Assortative Matching and Search,” Econometrica, 68, 343–369.
  26. [29]Stigler, G. J. (1961): “The Economics of Information,” Journal of Political Economy, 69, 213–225.
  27. [30]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
  28. [31]Vives, X. (1992): “How Fast do Rational Agents Learn?” Review of Economic Studies, 60, 329–347.
  29. [32]Weitzman, M. L. (1979): “Optimal Search for the Best Alternative,” Econometrica, 47, 641–654.

Paper details

vlnc:W_b5w5ndws
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