Making Talk Cheap: Generative AI and Labor Market Signaling

Large language models (LLMs) like ChatGPT have significantly lowered the cost of producing written content. This paper studies how LLMs, through lowering writing costs, disrupt markets that traditionally relied on writing as a costly signal of quality (e.g., job applications, college essays). Using data from Freelancer.com, a major digital labor platform, we explore the effects of LLMs’ disruption of labor market signaling on equilibrium market outcomes. We develop a novel LLM-based measure to quantify the extent to which an application is tailored to a given job posting. Taking the measure to the data, we find that employers had a high willingness to pay for workers with more customized applications in the period before LLMs were introduced, but not after. To isolate and quantify the effect of LLMs’ disruption of signaling on equilibrium outcomes, we develop and estimate a structural model of labor market signaling, in which workers invest costly effort to produce noisy signals that predict their ability in equilibrium. We use the estimated model to simulate a counterfactual equilibrium in which LLMs render written applications useless in signaling workers’ ability. Without costly signaling, employers are less able to identify high-ability workers, causing the market to become significantly less meritocratic: compared to the pre-LLM equilibrium, workers in the top quintile of the ability distribution are hired 19% less often, workers in the bottom quintile are hired 14% more often.

Introduction. This paper studies the equilibrium effects of large language models (LLMs) like ChatGPT disrupting the signaling value of written communication on hiring patterns and labor market efficiency.

To measure the signals workers send to employers, we develop a novel LLM-based approach to quantify how tailored an application is to a given job posting. Using data from Freelancer.com, a major digital labor platform, we provide new descriptive evidence that prior to the mass adoption of LLMs, workers used the customization of their applications to signal their ability, but that after LLMs, these signals became far less informative. To quantify the equilibrium effect of LLMs’ disruption of signaling, we build and estimate a structural model of the pre-LLM market that embeds signaling in the style of Spence (1973) into a model of labor supply and demand. Using the estimated model, we explore a counterfactual equilibrium in which LLMs reduce writing costs to zero in the pre-LLM market, thereby eliminating labor market signaling.

Written communication has long been used as a tool to facilitate matching in many markets, ranging from workers writing cover letters to employers to prospective students writing application essays to college admission committees. Because writing requires time and effort, the act of writing itself can send a signal (Spence, 1973). For instance, a worker interested in and well-suited to a specific job can send a customized cover letter instead of a generic one, thereby credibly signaling her value to the employer.

In recent years, however, the advent of generative artificial intelligence (AI) in the form of LLMs has dramatically lowered the cost of writing. LLMs can produce polished, human-sounding text in seconds at virtually no cost (Dathathri et al., 2024). Thus, this technology may threaten the signaling value of writing in the labor market. If all workers can cheaply produce highly customized, expertly written applications with the aid of an LLM, then employers may no longer be able to use written applications as credible signals of whom to hire. Given this potential threat to labor market signaling, it is important to empirically evaluate both the extent to which writing served as a costly signal before the mass adoption of LLMs and the extent to which LLMs now diminish the signaling value of writing and, in turn, impact matching in labor markets.

Our empirical context is the market for coding jobs on Freelancer.com, a major digital labor platform (DLP). On this platform, employers, seeking to outsource one-off digital tasks (e.g., website design), post jobs to find and hire freelance workers to complete those tasks. Workers apply to jobs with a short, written application—a “proposal”—as well as an asking wage—a bid.

This setting provides two distinct advantages for studying how LLMs disrupt labor market signaling. First, we can measure both the signals workers send and the effort they exert when sending those signals using the platform’s detailed text and click data on each application and job posting. In particular, we measure signals by quantifying the extent to which the text of each proposal is customized and relevant to the text of the corresponding job post, using an LLM to approximate human judgment at scale. Moreover, we measure signaling effort using click data to measure the time each worker spends on each application. Second, we can observe via click data which applications were written using an LLM-powered writing tool the platform introduced in April 2023 designed to automate proposal writing.

Related work. We contribute to four distinct literatures. First, there is a recent and growing empirical literature on how generative AI and LLMs affect labor markets (Acemoglu, 2025, Humlum and Vestergaard, 2025, Cui et al., 2025, Brynjolfsson, Li, and Raymond, 2025, Brynjolfsson, Chandar, and Chen, 2025, Dillon et al., 2025, Hartley et al., 2024, Stanton and Thomas, 2025, Stanton and Thomas, 2014, Teutloff et al., 2025, Eloundou et al., 2024). This literature has been primarily focused on studying how LLMs affect the nature of work and the productivity of workers via surveys and randomized experiments. The economics of AI in these studies is thus largely centered around how AI impacts the supply and demand for labor, though a few papers have begun to explore the implications of AI on signaling (Cowgill, Hernandez-Lagos, and Wright, 2024, Wiles and Horton, 2025, Cui, Dias, and Ye, 20257, Gans, 2024, Dhillon et al., 2025).

We contribute to this literature in multiple ways. Relative to existing papers that study LLMs’ effects on signaling via experiments, we move beyond partial equilibrium and provide empirical evidence on how LLMs have disrupted a market-wide signaling equilibrium. We also provide the first quantification of how this disruption of signaling affects equilibrium hiring patterns and welfare. Finally, to the best of our knowledge, our paper presents the first structural empirical model of LLMs’ effects on the labor market.

Second, there is a deep literature on the economics of labor market signaling itself. This literature has both been theoretical (Spence, 1973, Stiglitz, 1975, Waldman, 1984) and empirical (MacLeod et al., 2017, Tˆo, 2018, Layard and Psacharopoulos, 1974, Lang and Kropp, 1986, Bedard, 2001, Tyler, Murnane, and Willett, 2000, Clark and Martorell, 2014, Fang, 2006).8 This literature has largely focused, since the seminal work of Spence (1973), on educational choices as signals to the labor market. We contribute to this literature by providing the first empirical structural analysis of labor market signaling, where the signal is the information transmission in the application itself, i.e. the communication between workers and employers.

Third, there is a growing literature on the economics of digital labor platforms (Wiles, Munyikwa, and Horton, 2025, Brinatti et al., 2021, Krasnokutskaya, Song, and Tang, 2020, Autor, 2001, K ̈assi, Lehdonvirta, and Stephany, 2021, Horton, 2017, Agrawal et al., 2015, Barach, Golden, and Horton, 2020, Ren, Raghupathi, and Raghupathi, 2023, Stanton and Thomas, 2012, Pallais, 2014). This literature has studied a wide variety of topics from platform design to the role of communication in matching.

Method. While these findings provide evidence of LLMs disrupting labor market signaling, they do not reveal changes to the distribution of hired workers’ abilities, nor do they isolate the signaling channel from other potential effects of LLMs on labor supply and demand.2 To quantify how LLMs impact equilibrium hiring through their disruption of signaling alone, we therefore develop and estimate a structural model of labor market signaling. The model combines three typically distinct modeling approaches: we embed a (1) Spence signaling model— workers invest costly effort to produce noisy signals that positively correlate with their ability in equilibrium—into a (2) discrete choice demand model—employers form indirect expected utilities over application characteristics and their beliefs about ability—which, from the workers’ viewpoint, operates as a (3) scoring auction—workers submit applications competing on multiple dimensions to win a contract.

The core mechanism in the model is that higher-ability workers face lower costs of exerting effort to produce signals when applying to jobs than lower-ability workers do. As a result, in equilibrium, higher-ability workers exert more effort and thereby send higher signals on average.

Employers thus find these signals informative about worker ability, though they cannot perfectly infer ability since they do not observe effort and signals are noisy. The modeled signaling equilibrium is in the tradition of Spence (1973) in that while all workers could increase their chances of being hired by exerting more effort, only higher-ability workers find it worthwhile to do so.

Identifying the model poses two key challenges centered around identifying workers’ and employers’ equilibrium beliefs. First, we must disentangle workers’ costs of undertaking the job from their strategic bidding behavior as a function of their beliefs about their chances of being hired.3 Second, we need to separately identify employers’ beliefs about worker ability as functions of bids and signals from employers’ disutility from paying wages and willingness to pay for ability.4 Our identification argument addresses these two challenges by exploiting the information structure of our model. In equilibrium, a worker’s belief about her probability of being hired is independent of her private type (cost and ability), conditional on her choice of bid and effort.

As a result, worker beliefs equal empirical hiring probabilities as functions of bids and efforts, and are thus directly identified from data on equilibrium hiring decisions. Having identified worker beliefs, we can infer a worker’s cost of undertaking the job and her ability from her equilibrium bid and effort. With the recovered distribution of worker costs and abilities, we can identify employers’ beliefs about ability as functions of bids and signals, which allow us to identify employer disutility from paying wages and willingness to pay for worker ability.

We estimate the model on pre-LLM data, following the identification argument closely.5 We develop a novel simulation-based estimator to recover worker hiring probabilities as functions of bids and efforts, which we use to nonparametrically estimate the joint distribution of worker costs and abilities. Using this estimated distribution of worker types, we then estimate a flexible nonparametric model of employer beliefs about ability and, in turn, recover employer preferences.

Our estimates reveal that employers had a high willingness to pay for workers with high signals in the pre-LLM market because employers highly value worker ability, and the most reliable way to discern worker ability is through signals. We organize our findings into five results. First, employers directly value ability: we estimate that employers are willing to pay $52.16 on average for a one standard deviation increase in worker ability, which is about 79% of the standard deviation of bids in our sample. Second, there is significant variation in worker ability: employers value hiring workers at the 80th percentile of the ability distribution $97 more than they value hiring those at the 20th percentile.

Discussion. Leveraging this unique setting, we provide new descriptive evidence that LLMs have disrupted previously informative labor market signaling. We organize our descriptive evidence into three main findings. First, we show that before the mass adoption of LLMs, employers had a significantly higher willingness to pay for workers who sent more customized proposals. Estimating a reduced-form multinomial logit model of employer demand using our measure of signal, we find that, all else equal, workers with a one standard deviation higher signal have the same increased chance of being hired as workers with a $26 lower bid.1 Second, we provide evidence that before the adoption of LLMs, employers valued workers’ signals because signals were predictive of workers’ effort, which in turn predicted workers’ ability to complete the posted job successfully. Third, we find, however, that after the mass adoption of LLMs, these patterns weaken significantly or disappear completely: employer willingness to pay for workers sending higher signals falls sharply, proposals written with the platform’s native AI-writing tool exhibit a negative correlation between effort and signal, and signals no longer predict successful job completion conditional on being hired. Taken together, these findings suggest that workers’ job-specific proposals functioned as Spence-like signals of worker ability prior to the advent of generative AI, but that LLMs disrupted this signaling mechanism on which employers previously relied to make hiring decisions.

Having estimated the model, we explore how the elimination of signaling due to LLMs affects equilibrium hiring patterns and welfare. We simulate a counterfactual market equilibrium in which workers do not have access to any signaling technology and only choose bids when applying to jobs.6 Accordingly, employers only form beliefs about worker ability based on their observable characteristics.

Compared to the status quo pre-LLM equilibrium with signaling, our no-signaling counterfactual equilibrium is far less meritocratic. Workers in the bottom quintile of the ability distribution are hired 14% more often, while workers in the top quintile are hired 19% less often.

These effects are driven by three mechanisms. First, employers previously relied on signals to make hiring decisions, so losing access to them impinges on their ability to discern worker ability.

Second, more indirectly, the significant positive correlation between a worker’s ability and cost implies that, when employers lose access to signals and workers are forced to compete more intensely on wages, the prevailing workers with lower bids tend to have lower abilities. Third, since workers’ observable characteristics are poor predictors of their ability, employers have little to no information to distinguish between high and low-ability workers.

These changes to hiring patterns lead to a 5% reduction in average wages, a 1.5% reduction in overall hiring rate per posted job, a 4% reduction in worker surplus, and a small, less than 1%, increase in employer surplus. Worker welfare losses are driven by both the modest extensive margin decrease in hiring rates and the intensive margin decrease in wages. However, workers’ losses are mitigated by the fact that their writing costs are now zero, and by the fact that the counterfactually hired lower-ability workers tend to have lower costs. Employer surplus is virtually unaffected due to the highly competitive nature of the worker-side of the platform: the reduction in wages paid to workers roughly balances out the loss of hiring lower-ability workers.