Markets increasingly accommodate large language models (LLMs) as autonomous decision-making agents. As this transition occurs, it becomes critical to evaluate how these agents behave relative to their human and task-specific statistical predecessors. In this work, we present results from an empirical study comparing humans (N=216), multiple frontier LLMs, and customized Bayesian agents in dynamic multi-player bargaining games under identical conditions. Bayesian agents extract the highest surplus with aggressive trade proposals that are frequently rejected. Humans and LLMs achieve comparable aggregate surplus within their groups, but exhibit different trading strategies. LLMs favor conservative, concessionary proposals that are usually accepted by other LLMs, while humans propose trades that are consistent with fairness norms but are more likely to be rejected. These findings highlight that performance parity—a common benchmark in agent evaluation—can mask substantive procedural differences in how LLMs behave in complex multi-agent interactions.

Large language models---because of how they are trained and designed---are implicit computational models of humans---a homo silicus. Social scientists can use LLMs like economists use homo economicus: LLMs can be given endowments, information, preferences, and so on, and then their behavior can be explored in scenarios via simulation. We replicate four experiments using this approach and find qualitatively similar results to the original. Benefits of this approach include trying new variations for fresh insights, piloting studies via simulation, and searching for novel social science insights to test in the real world. The full version of the paper can be accessed at https://apostolos-filippas.com/papers/hs.pdf.

Buyers pursuing unavailable sellers is a common source of inefficiency in matching markets. We report the results of a field experiment in a large online marketplace where sellers could pay to signal higher capacity via a badge that simply said "available now." All sellers could rent this signal but only randomly treated buyers could see it. We find that sellers who rented this signal were positively selected, treated buyers sought them out more, and as a result matching efficiency increased. We discuss why costly signaling is necessary to facilitate this coordination, and show that mere statements about seller capacity had become uninformative in this marketplace. Two years after the experiment, we show that sellers renting this signal continue receiving substantially higher buyer attention. The full version of the paper can be accessed at https://apostolos-filippas.com/papers/costly-capacity-signalling.pdf.

We study the introduction of advertising in an online labor market where applicants could bid to "boost" their applications. Applicants with the highest bids received a prominent position at the top of the employer's application list. The platform randomly varied whether the employer was exposed to boosted applications, and within the exposure group, whether the employer was shown a disclosure that the application was boosted. We find that boosted applications are positively selected, and that boosting an application increases the likelihood of an applicant being hired by about 40%. The experimental design allows us to estimate that about 80% of this increase was due to the ranking effect---boosted applications ranking higher---and 20.2% was due to the signaling effect---the disclosure that the application is boosted. The full version of the paper can be accessed at https://apostolos-filippas.com/papers/boosted.pdf.

We model social media as collections of users producing and consuming content. Users value consuming content but due to scarce attention they may not value all content from other users. Users also value receiving attention, creating the incentive to attract an audience by producing valuable content but also through attention bartering—users mutually becoming each others’ audience. Attention bartering shapes substantially the patterns of production and consumption on social media, explains key features of social media behavior and platform decision-making, and yields sharp predictions that are consistent with data we collect from #EconTwitter. We conduct Twitter and Instagram user surveys that yield additional direct evidence in support of attention bartering, and we discuss the implications of attention bartering for the design of social media platforms.

A solution to marketplace information asymmetries is to have trading partners publicly rate each other post-transaction. Many have shown these ratings are effective; we show that their effectiveness deteriorates over time. The problem is that ratings are prone to inflation, with raters feeling pressure to leave "above average" ratings, which in turn pushes the average higher. This pressure stems from raters' desire to not harm the rated seller. As the potential to harm is what makes ratings effective, reputation systems, as currently designed, sow the seeds of their own irrelevance.

Peer-to-peer rental markets for short-term accommodation enable "hosts" to rent out properties to "guests." A common critique of such markets is that hosts benefit at the expense of their unwitting neighbors on whom they impose costs, potentially creating a market failure that justifies policy intervention. In this paper we examine four policy regimes that differ in the allocation of the decision right to become a host. We consider the market outcomes when hosting decisions are made by (1) individual tenants that maximize their revenue, (2) building owners that consider only rents from long-term rentals, (3) cities that consider the surplus of all city residents, and (4) a social planner that takes into account the surplus of both city residents and guests. We find that the equilibrium where tenants are allowed to individually decide has too much hosting, whereas city-specific policies result in too little hosting. The efficient social planner's solution is equivalent to the equilibrium where building owners decide. We conclude with a discussion of our results and lay out directions for future research.

Buyers and sellers in markets often signal to inform the other side about their preferences. Both have a mutual incentive to reveal information with respect to horizontal differentiation, but the case of vertical differentiation is more complex: a buyer claiming they place a high value on quality may attract more sellers of the right ``type'' increasing efficiency, but they might also simply pay a higher price. Although an efficiency-minded social planner may not care about higher prices, if this fear prevents a buyer from stating his of her true preferences, then desirable sorting caused by information-revelation may be unattainable. In this paper, we consider the buyer's vertical differentiation disclosure problem through the lens of a large field experiment conducted in an online labor market. A new signaling mechanism was introduced into the market that allowed buyers to state their relative preferences over price and quality. We find that the buyer signal improved seller-side sorting, with more sellers going to buyers of the right ``type''; the total number of applications also fell. However, sellers also clearly tailored their wages bid to the type of buyer they faced. Despite this markup, buyers chose to honestly disclose their preferences, suggesting they found the sorting effect to dominate the bargaining power effect.

Using data from an online labor market, I show that buyers inefficiently pursue oversubscribed sellers. Although oversubscribed sellers are positively selected, this fact alone cannot account for the amount of attention they receive. "Excess" buyer attention is caused by an information asymmetry: buyers do not know seller capacities and cannot condition their search efforts accordingly. Sellers---having free disposal on buyer inquiries---have little incentive to disabuse searching buyers. This misdirected search effort is consequential: using an instrumental variables analysis, I show that a recruited seller rejecting a buyer's recruiting inquiry reduces the probability of match formation by as much as 67% points. Motivated by the empirical results, I show how the market-creating platform can optimally allocate buyer attention, given each seller's estimated per-encounter probability of forming a match.

Paid crowd work offers remarkable opportunities for improving productivity, social mobility, and the global economy by engaging a geographically distributed workforce to complete complex tasks on demand and at scale. But it is also possible that crowd work will fail to achieve its potential, focusing on assembly-line piecework. Can we foresee a future crowd workplace in which we would want our children to participate? This paper frames the major challenges that stand in the way of this goal. Drawing on theory from organizational behavior and distributed computing, as well as direct feedback from workers, we outline a framework that will enable crowd work that is complex, collaborative, and sustainable. The framework lays out research challenges in twelve major areas: workflow, task assignment, hierarchy, real-time response, synchronous collaboration, quality control, crowds guiding AIs, AIs guiding crowds, platforms, job design, reputation, and motivation.

The emergence of online labor markets makes it far easier to use individual human raters to evaluate materials for data collection and analysis in the social sciences. In this paper, we report the results of an experiment - conducted in an online labor market - that measured the effectiveness of a collection of social and financial incentive schemes for motivating workers to conduct a qualitative, content analysis task. Overall, workers performed better than chance, but results varied considerably depending on task difficulty. We find that treatment conditions which asked workers to prospectively think about the responses of their peers - when combined with financial incentives - produced more accurate performance. Other treatments generally had weak effects on quality. Workers in India performed significantly worse than US workers, regardless of treatment group.

We present a model of workers supplying labor to paid crowdsourcing projects. We also introduce a novel method for estimating a worker's reservation wage - the key parameter in our labor supply model. We tested our model by presenting experimental subjects with real-effort work scenarios that varied in the offered payment and difficulty. As predicted, subjects worked less when the pay was lower. However, they did not work less when the task was more time-consuming. Interestingly, at least some subjects appear to be "target earners," contrary to the assumptions of the rational model. The strongest evidence for target earning is an observed preference for earning total amounts evenly divisible by 5, presumably because these amounts make good targets. Despite its predictive failures, we calibrate our model with data pooled from both experiments. We find that the reservation wages of our sample are approximately log normally distributed, with a median wage of $1.38/hour. We discuss how to use our calibrated model in applications.