Better Applications, Worse Matching: Artificial Intelligence and Talent Allocation

08 Wednesday Apr 2026

Posted by in Academic Research, Artificial Intelligence, Digital Economics, Labor, Learning, Organization, Social Dilemma

≈ Leave a comment

Tags

, , ,

The main message of my new paper “Better Applications, Worse Matching: Artificial Intelligence and Talent Allocation” is simple but important: generative AI can make everyone look better on paper, can, in fact, even make everyone more productive in everything their doing, yet the labor market may be worse off. AI can help applicants write cleaner resumes, sharper cover letters, and more polished work samples. But hiring does not happen after firms fully know who will perform best in the job. It happens earlier, based on imperfect screening materials. The paper shows that if AI improves those materials’ appearance while making them less informative about actual fit, then matching workers to jobs can deteriorate even as AI raises productivity inside every single job.

The core contribution is to separate two effects that are often blurred together in public discussion. One is what AI does inside a job once someone is hired, i.e., it may help them write faster, code better, or complete tasks more efficiently. The other effect is what AI does before hiring. It may change how informative applications, interviews or work samples are. The paper argues that these are not the same thing. A labor market can continue to sort efficiently on the basis of the information it sees, yet still produce worse matches if the visible ranking of applicants becomes a poorer guide to underlying talent and job fit.

That distinction leads to a n interesting theoretical result. Even if AI makes every possible worker–firm pairing more productive, total output in the economy can still fall if the screening stage gets worse enough. In other words, better applications do not necessarily mean better matching. The paper also argues that this creates an “arms race” on the applicant side. While using AI to polish applications may be privately attractive for each worker, it becomes socially excessive when everyone does it. Once firms can no longer rely as much on first-round materials, they are predicted to respond by using more verification, more work-sample tests, more probationary hiring, and more early post-hire evaluation.

For a broader audience, the paper’s contribution is to reframe the conversation about AI and work. Much of the debate asks whether AI makes workers more productive. This paper says that is only half the question. The other half is whether AI improves or degrades the information used to allocate people to opportunities in the first place. That is important for everyone trying to understand and test whether AI increases the performance of a specific labor market.

The Disruption of Attention Platforms by Generative AI

04 Saturday Apr 2026

Posted by in Academic Research, Artificial Intelligence, Digital Economics, Platforms, Strategy

≈ Leave a comment

Tags

, , , , ,

Generative AI is changing digital platforms in a very specific way, it explodes content supply but it is not creating more human attention. That matters because platforms do not make money from raw upload volume. Rather, they make money from keeping users engaged with content they actually value. My new paper “The Disruption of Attention Platforms by Generative AI,” with Chenyang Li, Xun Wu and Fei Xiao, argues that this creates a new management problem for attention platforms. The core challenge is no longer just how to attract more creators, but how to govern a flood of cheap AI-generated content when the truly scarce resource is informative user attention.

To study this problem, we introduce a simple but powerful framework. On the platform, both user-generated content and AI-generated content compete for recommendation traffic. Every upload first enters a “training pool,” where the platform uses early views to learn whether the content is good enough to deserve wider distribution. Only the content that proves itself is promoted into a “spotlight pool.” In the model, AI-generated content is cheaper to produce, but user-generated content is more likely to be high quality. Users stay on the platform only if their experience is good enough relative to what they could do elsewhere, so bad content does not just waste impressions today. It can also reduce future attention.

The main result is striking: cheaper AI-generated content can make a platform worse off even while it increases supply. The reason is congestion in quality discovery. When too many low-hit-rate uploads enter the system, the platform must spend more scarce human attention evaluating content. As a result, it has less attention left to amplify the winners. That lowers spotlight exposure, reduces realized average quality, and can even shrink the platform’s effective attention budget as users disengage. In other words, more content is not always better. If supply grows faster than the platform’s ability to identify value, abundance can become a liability rather than an asset.

The practical message is not at all that “AI is bad.” It is that AI changes what good platform governance looks like. The paper shows why tools such as provenance-based payouts, better disclosure, upload fees, and multi-stage screening can all matter. They help platforms allocate scarce attention more intelligently and keep low-signal content from overwhelming discovery. For managers, that shifts the strategic question from “Should we allow AI content?” to “How should we redesign incentives and screening when content is cheap but attention is not?” That is a useful lens not just for social media, but for any digital platform whose value depends on matching users with quality at scale.

Selling Synergies

24 Tuesday Mar 2026

Posted by in Academic Research, Antitrust, Innovation, Organization, Strategy

≈ Leave a comment

Tags

, , , , , ,

The Synergy Trap: Why Interconnected Tech Often Launches Half-Baked

We usually assume that when companies build complementary products—like electric vehicles and charging networks, or hardware platforms and specialized software—the natural synergies align their incentives and create a win-win for everyone. But what happens when independent firms are developing these interconnected ecosystems over time? In my new paper “Selling Synergies,” I formally explores this dynamic by modeling two profit-maximizing sellers developing complementary products. Each firm continuously invests in product development and decides exactly when to stop refining, launch to the market, and set a price for consumers who exhibit a synergy benefit from owning both items.

The analysis reveals an interesting mechanism that turns these positive synergies into a source of inefficiency: the firm that launches first (the “leader”) has a strong incentive to strategically underinvest in their own product. By halting development early and launching a lower-quality product with an exclusionary, high price, the leader forces the second firm (the “follower”) into a corner. The follower is nudged to target the broader mass market and must continue developing their product to a high standard. This allows the leader to essentially free-ride on the follower’s hard work, extracting a lucrative “synergy premium” from early, high-paying buyers without putting in the development time themselves.

Because securing this leader position is so incredibly profitable—yielding more than a standard standalone monopoly would—a destructive “preemption race” kicks off. Both firms aim to capture the first-mover synergy premium that they try to undercut each other’s launch dates. Ultimately, this intense competition completely unwinds the very premium they were chasing. The race only stops when synergy rents profits are totally dissipated, resulting in companies rushing underdeveloped products to the market long before they are actually ready. Crucially, note that this rush to the market is not driven by an arms race between competitors who try to beat each other to the punch, but rather by an urge to dominate a fledgling ecosystem.

The results challenge standard economic and regulatory playbooks. While we intuitively expect product synergies to enhance overall value, this uncoordinated race actually destroys product quality and reduces overall market welfare. Government interventions like standard R&D subsidies may ease the preemption race but fail to address the core issue. What is more, some results suggest that standard antitrust policies aimed at preventing corporate mergers might actually backfire in these specific markets. A merged, joint monopoly that internalizes these synergies actually delays its product launches to build higher quality, ultimately benefiting both producers and consumers.

Search platforms rewrite the rules of online shopping

18 Wednesday Feb 2026

Posted by in Digital Economics, Online Advertising, Platforms

≈ Leave a comment

Tags

, ,

The Cornell Chronicle published a nice piece called “Search platforms rewrite the rules of online shopping” breaking down the main message and policy relevance of my paper “Search Platforms: Big Data and Sponsored Positions,” joint with Maarten Janssen, Marcel Preuss and Cole Williams that recently published in the Economic Journal. The main message conveyed: It’s not necessarily advertiser-funded search results that harm consumers, but search platform market power.

Optimally Informative Rankings and Consumer Search

14 Friday Nov 2025

Posted by in Academic Research, Digital Economics, Learning, Platforms, Signaling, Strategy

≈ Leave a comment

Tags

, , , ,

Online platforms (and multi-product firms) generate lists of products (or services) in response to consumer search queries. In ranking these products, platforms draw on their vast amount of information about past consumer search behavior and their purchase history. In “Optimally Informative Rankings and Consumer Search” (joint with Maarten Janssen, Marcel Preuss and Cole Williams), we investigate how much of their pertinent information platforms convey through their rankings of products.

When consumers engage in costly search and expect to find some products they like better and others they like less, they are updating their expectations about the remaining alternatives whenever they inspect a product. As a result, three scenarios ensue: 1) If they like a product very much, they immediately buy. 2) If they strongly dislike a product, they become more optimistic about the remaining alternatives and continue searching. 3) If however, the consumer is fairly indifferent about a product, they may neither buy it nor continue to search as their experience does not induce sufficient optimism about the remaining products. In turn, they abort search altogether without buying a product.

Scenario 3) is a novel finding about consumer search with learning and has meaningful implications for platforms and their regulators. Understanding optimal consumer search behavior, we derive the optimal platform ranking of two products, one of which promises a higher value for the consumer (in expectation) and one a lower value. The more informative the platform’s ranking, i.e., the higher the probability it puts the higher-value product in the first position of their ranking, the higher the likelihood the consumer (inspects and) buys the first product. However, if the consumer does not like the first product, they are less likely to inspect the second the more informative the ranking, i.e., the less likely the product in the second position is the higher-value product.

It turns out that the platform optimally chooses to obfuscate their ranking to increase the probability of the consumer to inspect both products when consumer search cost (i.e., their implicit cost of inspecting another item) is low, but to provide a fully informative ranking to increase the probability the consumer buys the first product if the consumer’s search cost is high. Interestingly, the platform provides either less information than would be necessary to induce search of the second product, or more information than necessary to ensure the consumer’s participation in search in the first place. An intriguing result of our findings is that platform and consumer welfare are aligned only if search cost is high (in which case the platform maximizes social welfare) but at odds with each other when search cost is low.

Education Signaling and Employer Learning Heterogeneity

12 Wednesday Nov 2025

Posted by in Academic Research, Education, Labor, Learning, Organization, Signaling

≈ Leave a comment

Tags

, , , , , , , , ,

It is well known that individuals choose to obtain higher education degrees to signal their ability to potential employers (Education Signaling), and that employers learn about their workers’ abilities from observing their output over time (Employer Learning). The quicker and the more accurate employers learn about their workers, the less important education is as a signal of their ability.

In “Education Signaling and Employer Learning Heterogeneity,” (joint with Yuhan Chen and Michael Waldman), we investigate the consequences of combining these fundamental concepts of Labor Economics. In particular, we exploit the fact that the importance of teamwork and other determinants of the observability of individual output not under the worker’s control vary across industries (and occupations).

When industries differ in their speed (or accuracy) of employer learning, higher-ability workers tend to prefer a faster-learning environment. This is because they prefer their compensation depends on their own output, a function of their ability, rather than on the ability of others choosing the same education signal (e.g., degree). This is not the case for lower-ability workers, however, who benefit when the average ability of those choosing the same education is a determinant of their compensation. This finding has several important consequences.

First, as higher-ability workers face a strong incentive to choose a faster- rather than a slower-learning industry, they even join the faster-learning industry if they are more productive in the slower-learning industry. In other words, a sorting distortion regarding industry choice arises and lowers social welfare. Second, for any given education level, higher-ability workers choose a faster-learning industry. As a result, industry choice itself is a signal of worker ability on which employers condition their learning. Third, the sorting distortion across industries lowers education investment for signaling purposes, increasing social welfare.

We show that the logic of our results persists across industries with symmetric and asymmetric employer learning components, i.e., whether the worker’s current employer is better informed about their ability than their competitors. Likewise, our results are robust to different bidding specifications among firms, i.e., whether firms engage in simultaneous bidding for workers, or the worker’s current employer can submit counteroffers.

A variant of our model in which workers learn their productivity difference across industries before they choose their education level offers a potential explanation for a heretofore neglected labor market “puzzle:” why do few of the economically most successful individuals in real-world labor markets hold higher education degrees? We show that fewer workers that join a faster-learning industry choose higher education levels. In particular, if industries or occupations differ greatly in their speed or asymmetry of learning, it may be the the highest ability individuals in a faster-learning industry or occupation that achieve the highest lifetime wages but choose a low education level. The non-monotonicity of education levels in ability in our model is novel, and lends itself well to explain the success of entrepreneurs choosing low education levels or to drop out of school. Finally, we discuss the testable implications of our theory, and how it connects to existing empirical work.

Strategic Referrals among Experts

02 Sunday Mar 2025

Posted by in Academic Research, Antitrust, Labor, Organization

≈ Leave a comment

Tags

, , , , ,

Matching problems with the right expertise within and across firms is essential for the functioning of the economy. When problems are complex, however, their owners are often unaware of their exact nature, and costly diagnosis by an expert is required. This is why the initial allocation of problems is often independent of an expert’s specialization, but is affected by publicly available information of the expert’s past success rate in solving problems. Prospective clients, for example, consult medical providers with higher satisfaction scores or lawyers with better track records, while firms allocate problems to managers who got the job done in the past. The services of experts in these scenarios can be described as reputation goods. Reputation goods are (1) differentiated between sellers, (2) their product quality is consumer specific (3) and initially uncertain, (4) important, and thus (5) worth considerable information acquisition effort. In order to strengthen their reputation, thereby increasing demand for their services, experts sometimes find it worthwhile to forego compensation and sometimes refer problems to other experts who may be a better fit.

In Strategic Referrals among Experts (joint with Yi Chen and Mark Satterthwaite), we build a dynamic model to examine how reputational concerns of specialized experts shape their referral behavior and, as a consequence, market outcomes. We find that competition for problems among individual experts does in general not result in an efficient allocation of problems across experts. What is more, instead of sharpening their expertise, experts may want to invest in their abilities outside of their specialty—to the detriment of the economy. This issue may be exacerbated by the adoption of Artificial Intelligence (AI) in the workplace, often increasing professionals’ productivity outside their main specialty. Also, market efficiency increases in the experts’ compensation for diagnosing problems relative to treatment, potentially discouraging them from specializing in the first place. We show that referral alliances that set referral prices may alleviate these issues, though possibly at the expense of some experts. Pareto-efficient market outcomes generally require revenue sharing beyond referrals, i.e., some form of partnership contract. It is therefore imperative to account for referral incentives when regulating markets for professionals.

Our model features an economy with a mass of experts of two different types, who compete against each other for problems in continuous time. Each problem belongs to one of two categories, matching the experts’ types. That is to say, an expert has an absolute advantage (higher success rate) in treating problems of their category, i.e., those that match their type. Since problems require costly diagnosis, the demand for an expert’s services, i.e., the problems they receive, is independent of the expert’s type. An expert initially diagnoses problems, and then, for each problem, decides whether to to treat it or refer it to a more apt expert.

The demand for an expert’s services at any point in time depends on their own and other experts’ record of solving problems. An expert can solve a problem by either treating it successfully or by referring it to another expert who then treats it successfully. An expert’s record is characterized by its thickness—the decaying number of problems they diagnosed over time—and the expert’s reputation—their decaying historic success rate in solving problems. Reputation is directly payoff-relevant in that demand for the expert’s services increases in their reputation but decreases in an aggregate index of the their competitors’ reputation, perhaps the market average. The record’s thickness, on the other hand, does not directly affect demand but the inertia of the expert’s reputation. The thicker an expert’s record, the less today’s success rate matter in determining reputation.

We describe experts’ referral behavior in the unique steady state Markov equilibrium of our model. Uniqueness of the steady state follows from a feedback loop induced by the laws of motion that determine the expert’s record. As an expert refers more mismatched problems, their reputation increases, and so does their demand leading to a thicker record. Then, ultimately a tipping point is reached, at which reputation becomes sufficiently inert to the current success rate. As a result, the expert ceases to refer, and, ultimately, their record’ thickness decreases. This feedback loop is illustrated in the phase diagram below:

Poaching, Raids, and Managerial Compensation

13 Monday Jan 2025

Posted by in Academic Research, Labor, Organization

≈ Leave a comment

Tags

, , , , , , ,

In Poaching, Raids, and Managerial Compensation—joint work with my friends and colleagues Yi Chen, Fabiano Dal-Ri and Daniela Scur—we develop a theory of managerial poaching and subsequent worker raids. We postulate that firms do not only poach managers for their skills but also for their personnel-specific knowledge about workers they supervise. In the equilibrium of our model with asymmetric employer learning, substantially more productive firms poach managers for the option value of identifying high-ability workers. When poaching firms pay thus pay twice for superior worker ability: once indirectly through the manager’s salary and once through the worker’s wage. The ensuing information rent for the manager is a result of firms attempting to protect themselves from raids of their high-ability workers by retaining their manager, leading to adverse selection of workers in the labor market.

We then derive seven theoretical predictions to test our model against data from Brazil’s formal sector: 1) Managers are poached by more productive firms. 2) When a firm poaches a manager from another firm, the firm is more likely to raid their workers. 3) Poached managers earn higher salaries. 4) The salary of a poached manager increases in the poaching firm’s demand for information, and 5) in the poaching firm’s supply of information. 6) The salary of a poached manager increases in the raided workers’ abilities. 7) Raided workers are of higher ability than non-raided workers.

Our empirical analysis provides strong support for our theoretical model and differentiates managerial poaching for personnel-specific information from potential other explanations of manager-worker com-movement such as workers referring others in their network or non-performance based favoritism in hiring. While we find limited evidence for both these phenomena, they are not consistent with the weight of our empirical evidence. On the hand, the effect of workers following managers is substantially larger than a worker-worker effect,

while on the other raided workers that follow a manager are of significantly higher ability than similar hires into the same firms, i.e.,

The intuition behind our results is potentially meaningful for recruiting and promotion strategy, investment in human capital, knowledge flow design within organizations, and the regulation of labor markets.

Strategic Wage Posting, Market Power and Mismatch

26 Thursday Sep 2024

Posted by in Academic Research, Labor, Organization

≈ Leave a comment

Tags

, , , ,

Nine years after starting work on this paper it finally published in the Journal of the Labor Economics. My gratitude goes to a fantastic editor, Kevin Lang at Boston University, and three anonymous referees whose suggestions made the paper a stronger and more robust contribution to labor market theory.

The main idea conveyed in this paper is as follows: When firms compete for heterogeneous workers in a well-defined labor market, larger firms have an incentive to lower wages more than their small competitors. If firms pay all their workers equally, the result is intuitive: it is more expensive for a large firm to raise wages than for its smaller competitors.

Strikingly, however, the result remains true even if firms post different wages for their positions. The decisive force behind this result is that when a firm hires a worker of high ability, the set of workers its other open positions compete for becomes less attractive. This externality is stronger for firms with more open positions, and therefore more market power. As such, firms that enjoy higher levels of market power may end up recruiting lower ability workers.

In this paper, I show that the resulting inefficiency, i.e., mismatch when assortative matching is desirable, may be quite substantial, and that the result is robust to sequential wage posting.

Actions and Signals

24 Sunday Dec 2023

Posted by in Education, Signaling

≈ Leave a comment

Tags

, , , , ,

In the canonical signaling model, a sender holding private information about her type chooses an action to change the beliefs of uninformed receivers about her type, e.g., a (future) worker chooses a higher education level to signal her ability. In other words, the chosen action serves as a signal of her privately known type. There are numerous examples of real-world settings that feature signaling, however, that work differently.

Suppose that, instead, the sender privately observes her type, and choses unobservable actions that together with her type determine an observable outcome, which acts then as a signal of both the sender’s type and her chosen actions. For example, consider the college attended by a worker. A more prestigious institution serves as a signal of both the innate ability or skill of the worker as well as of her effort in high-school and in studying for standardized tests. The worker, however, clearly does not choose her college, but her effort in high-school, volunteering, test prep, etc.

In “Actions and Signals,” co-authored with my colleague Mike Waldman, we introduce a generalized model of signaling that captures strategic incentives in these environments. We show that in equilibrium, a fundamentally different behavior than in the canonical signaling model, a special case of our framework, can arise.

If an action and a signal are one and the same, receivers know the sender’s action, and infer her type from the signal. In this scenario, over-investment, i.e. the sender choosing an action (=signal) that exceeds its efficient level, ensues (independent of whether the action is productive or not). This is because the sender gets rewarded for her action (as observed by receivers) but chooses an even higher level of the action to communicate that she is a higher type sender.

On the other hand, if an unobservable action and the sender’s type combine to generate an observable outcome that serves as a signal of both the action and her type, a different logic applies. When the sender chooses a higher level of the action, the signal increases. Receivers attribute some of this increase to a higher action and some of the increase to a higher type. As a result, equilibrium behavior of the sender depends on how the action and the sender’s type contribute to the observable signal vs. how they contribute to the sender’s output (for which receivers pay). In fact, if the action chosen by the sender is less important (relative to her type) for the signal than her output, under-investment results.

In terms of education signaling, this means that in cases where effort in high-school is of importance for future job-performance, the worker may choose an inefficiently low level of effort if it does not equally increase the quality of the college she will attend. We discuss education as well as third-party reviews such as rating systems or certification as two of the (many) applications of our model, and argue why the specific determinants in a signaling environment will be the drivers of efficiency in terms of over- or under-investment by the sender (or potentially both in the case of multiple actions).

We furthermore introduce incomplete sender information that allows us to span an environment with signaling (sender perfectly knows her type) and signal jamming (sender is uninformed) as the extreme cases, and show that uninformed sender behavior may be more efficient than that of a well-informed one. Finally, we discuss the tradeoffs between multiple actions and deal with productive signals, i.e., when the signal realization itself increases sender output.