Making Talk Cheap: Generative AI and Labor Market Signaling (2511.08785v1)

Abstract: 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 have a high willingness to pay for workers with more customized applications in the period before LLMs are 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.

• The paper shows that LLM adoption sharply diminishes the predictive power of written signals, with employer responsiveness dropping by over 60%.
• It employs large-scale Freelancer.com data and a structural equilibrium model to quantify changes in hiring outcomes and wage formation.
• Counterfactual simulations reveal that eroded signaling leads to a 19% reduction in high-ability hires and overall declines in market meritocracy.

Generative AI, Labor Market Signaling, and Meritocracy Disruption

Introduction

This paper examines the equilibrium effects of generative AI—specifically LLMs like ChatGPT—on traditional labor market signaling via written job applications. Written content, historically used as a costly signal of worker ability (Spence, 1973), is rendered inexpensive and nearly effortless by LLMs. The authors exploit detailed behavioral and application data from Freelancer.com to quantify the collapse of informative signaling and its implications for hiring efficiency, wage formation, and overall meritocracy, employing both empirical methods and a structural equilibrium model.

Empirical Setting and Signal Measurement

The authors utilize a large-scale dataset from Freelancer.com, comprising over 2.7 million applications to 61,000 coding jobs. The platform’s introduction of a native LLM-powered proposal generator offered a unique natural experiment, enabling isolation of the effects of reduced writing costs on signaling.

Signal measurement leverages LLM-based scoring via Meta’s Llama 4 Maverick model to assess proposal customization and relevance across nine well-defined criteria, split into “custom” (job-specific) and “generic” (general writing quality) dimensions. Signals are penalized for content re-use (Levenshtein edit-distance below 4% triggers a zero customization score), and the resulting signal metrics are designed to be invariant to the equilibrium environment. This approach avoids endogenous outcome-driven scoring and eschews conventional NLP metrics (which are shown to be insufficient for equilibrium signaling analysis).

Signaling effort is proxied by the interval between a worker viewing a job post and submitting the application, filtered for plausible bounds and corrected for worker-specific writing efficiency. Consideration sets—applications actually reviewed by employers—are algorithmically reconstructed from click and interaction data.

Descriptive Evidence: Pre- and Post-LLM Dynamics

Pre-LLM Equilibrium

Before adoption of generative AI, signals in proposals are highly predictive of hiring outcomes, both in unconditional win rates (upper quantiles of signal correspond to ~4x increased hiring probability) and in employer’s estimated willingness to pay: a one standard deviation increase in signal equates to a $25–$26 increase in bid attractiveness (relative to a $66 bid standard deviation).

Signals are strongly correlated with signaling effort; within-worker variation confirms a monotonic, concave relationship between time spent and signal produced (fixed-effects regression coefficient ≈ 0.87). Signals in turn predict job completion success, but their explanatory power collapses on conditioning for effort—suggesting that signals function entirely as noisy, costly correlates of application-specific effort rather than conveying intrinsic information.

Post-LLM Equilibrium

Following mass adoption of LLMs, most notably after the introduction of the on-platform AI-writing tool in April 2023, the distribution of signals shifts sharply towards higher “customization,” with AI-generated proposals dominating the right tail of the metric. However, the marginal value of signals to employers collapses: hiring probability responses to signal diminish by over 60%. Signals cease to predict effort, with the relationship actually turning negative in AI-generated applications. Crucially, signals no longer predict successful job completion.

Event-paper and regression analyses demonstrate that the equilibrium shift coincides precisely with LLM tool uptake, not with exogenous compositional changes in the applicant pool or endogenous changes in worker-side quality.

Structural Model: Spence Signaling in a Platform Market

To quantitatively isolate and project counterfactual effects, the authors develop a static equilibrium model integrating:

• Spence signaling: workers select effort to produce noisy signals, with higher ability workers facing lower marginal costs of effort.
• Discrete choice demand: employers rank and select candidates based on a utility composed of observable characteristics, bid, and estimated ability (inferred from signals).
• Scoring auction: workers compete on multiple dimensions for a single contract.

Supply-side identification exploits the independence of a worker’s private type (ability, cost) and hiring outcome, conditional on observable actions; worker equilibrium beliefs are mapped to empirical hiring probabilities. Demand-side identification leverages the recovered joint distribution of types and signals to estimate employer beliefs about worker ability as functions of signal and observables.

Model estimation utilizes simulation-based strategies, empirical Bayes corrections for effort variance, and nonparametric employer belief fitting (monotonic regression and PCHIP interpolation).

Main Results and Counterfactuals

Estimates

• Employers’ willingness to pay for ability: $52.16 per one standard deviation in signaled ability (≈ 79% of bid SD).
• Variation in ability across observables explains only ~3% of total ability variance.
• Signal-ability correlation: 0.55 in equilibrium.
• Ability-cost correlation: 0.19—higher ability workers also tend to have higher costs.

Counterfactual: No-Signaling Equilibrium

Simulating a market in which written communication no longer signals ability (writing cost zero, no separation on costly effort), the following results are identified:

• Market composition: high-ability workers (top quintile) are hired 19% less, low-ability workers (bottom quintile) are hired 14% more.
• Wages: average winning bids fall by 5%.
• Hiring rate: per job posting, hiring rates fall by 1.5%.
• Welfare: worker surplus drops by 4%, employer surplus increases by <1%, net market surplus falls by 1%. Welfare loss is concentrated on high-ability workers who cannot compete purely on wages due to higher opportunity costs; employer surplus is buffered via wage compression.
• Meritocracy: market becomes substantially less meritocratic, as selection shifts away from ability and towards cost.
• Observable group composition: hiring rates by observable group (reputation, arrival, country) change minimally, further confirming that LLM-induced disruption is mediated by unobservable ability/cost heterogeneity.

These results are robust to alternative NLP-based signal measures.

Theoretical Implications

The findings empirically validate a classic signaling model: when the cost structure enabling credible signaling collapses, separation fails, and noisy signals become uninformative. In the Freelancer market, high-ability workers previously distinguished themselves via costly customization; with LLM-written proposals, these distinctions evaporate.

The positive ability-cost correlation induces a welfare-reducing shift when the market moves to price-only competition. If ability and cost were negatively correlated, the disruption of signaling would have less impact.

Observable proxies (e.g., platform reputation, arrival time, country) are inadequate replacements for application-specific signals. The role of “cheap talk" becomes prevalent when writing is functionally costless.

Practical Implications and Future Directions

LLMs can disrupt not only labor markets but any matching market that relies on costly written communication for sorting (e.g., college admissions, grant applications, digital gig platforms). Without alternative costly signals or richer screening mechanisms, generative AI erodes the informational value of communication, potentially undermining meritocracy and efficient matching.

Markets might respond by:

• Designing new screening technologies resilient to AI-based gaming (e.g., live trials, technical assessments, proofs of prior work).
• Redesigning contracts to promote post-hire learning via trial periods or sequential project stages.
• Shifting to more horizontal matching, where LLMs enhance signaling of preferences and job fit, rather than vertical ability-based sorting.

Theoretically, the results highlight the importance of maintaining credible separation mechanisms in environments subject to technological disruption of signal production costs.

Conclusion

The collapse of costly written communication as a credible signal, due to generative AI, creates significant efficiency and welfare losses in congested labor markets. High-ability workers lose their edge, meritocracy is undermined, and embedding new forms of screening or contract experimentation becomes essential. The core economic insight—when talk becomes cheap, sorting by costly signals fails—carries far-reaching implications for labor market design, evaluation, and policy in the age of machine-generated text.