When Words Lose Meaning: AI and the Future of Job Applications

Author: Denis Avetisyan

New research suggests that the rise of generative AI is undermining the value of written applications, creating challenges for employers and workers alike.

The analysis of hiring probabilities before and after the implementation of Large Language Models demonstrates a discernible shift in recruitment patterns, suggesting these models exert a measurable influence on candidate selection processes.

A new equilibrium model demonstrates that Large Language Models erode the signaling value of applications, leading to less efficient job matching and decreased overall welfare.

Traditional labor market signals, like carefully crafted resumes and cover letters, presume costly effort distinguishes capable workers—but this assumption is challenged by the rise of generative AI. In ‘Making Talk Cheap: Generative AI and Labor Market Signaling’, we investigate how large language models disrupt this signaling dynamic, finding that the reduced cost of written communication erodes the value employers place on customized applications. Our analysis reveals a shift towards less meritocratic outcomes, with higher-ability workers facing diminished hiring prospects and lower-ability workers gaining ground. Will these trends necessitate a fundamental rethinking of how employers assess candidate quality in the age of readily available AI-generated content?

## The Architecture of Trust in Digital Labor

Digital labor platforms, such as Freelancer.com, function as markets where workers communicate their capabilities to potential employers. This creates inherent information asymmetry, as employers lack complete knowledge of worker quality. Workers therefore signal their suitability for projects. The Spence Signaling Model provides a theoretical framework: workers invest in costly signals – actions expensive to perform but reliably indicating high ability – to credibly differentiate themselves. These signals include detailed applications, portfolio development, and skill verification. This process unfolds within a ‘scoring auction’ framework, where workers compete on multiple dimensions.

Comparing the distributions of hired workers reveals that ability and cost quantiles differ between the no-signaling and status-quo conditions.

However, emerging technologies challenge these signaling mechanisms. Automated skill assessments and standardized profiles may reduce the value of traditional investments. Understanding these evolving dynamics is crucial for both workers and platforms, for every adjustment reverberates throughout the entire network.

## The Erosion of Credibility: LLMs and the Signal

Large Language Models (LLMs) disrupt the signaling equilibrium by substantially reducing the cost of generating persuasive application content. Historically, a well-crafted application signaled underlying ability. The proliferation of LLMs diminishes the reliability of written communication as an indicator of skill, as individuals can leverage these tools to produce high-quality materials irrespective of their inherent abilities.

This research demonstrates the impact of LLMs on signaling value through an LLM-Based Signal Measurement, assessing the degree to which an application is tailored to a job posting. Analysis reveals a significant shift in the distribution of signals following the integration of LLMs, indicating a change in the characteristics of AI-written communication.

The distribution of signals shifted following the integration of large language models, indicating a change in the characteristics of AI-written communication.

Quantification reveals that the core assumption of the Spence Signaling Model is challenged. A shift towards a ‘no-signaling’ equilibrium results in a 0.63 percentage point decrease in the hiring rate, suggesting employers struggle to differentiate candidates based on application quality, leading to potential inefficiencies.

## Quantifying the Cost of Opacity

A Discrete Choice Demand Model was utilized to estimate employer preferences and responsiveness to signals in worker applications. This allowed quantification of how employers evaluate candidates based on observable characteristics and perceived abilities. The model’s parameters were calibrated using a comprehensive dataset of application and hiring decisions.

Isotonic Regression and the Piecewise Cubic Hermite Interpolating Polynomial accurately modeled employer beliefs concerning worker abilities, facilitating the construction of a belief function that maps application signals to inferred skill levels.

A kernel density estimate of worker abilities demonstrates the underlying distribution of skill levels within the workforce.

Findings reveal a significant Welfare Loss – a 4% reduction in worker surplus – resulting from the diminished credibility of application signals. This decrease is a direct consequence of LLMs enabling workers to generate superficially persuasive applications with minimal effort, contributing to a 1% reduction in overall market surplus.

## Design Implications: Rebuilding Trust in the Network

Through Counterfactual Analysis, the effects of removing signaling costs were simulated, effectively decoupling application quality from worker ability. This demonstrated a clear deterioration in matching outcomes when signaling is absent.

Using a Reduced-Form Multinomial Logit Model, the robustness of these findings was validated. Results indicate a 14.9 percentage point decrease in hiring rates for high-ability workers, and a 38.6 percentage point increase for low-ability workers when signaling costs are removed.

Hiring rates experienced percentage changes across ability and cost quintiles when comparing the no-signaling and status-quo conditions, highlighting disparities in recruitment patterns.

These findings have significant implications for digital labor platform design, suggesting the need for alternative mechanisms to assess worker abilities and mitigate the consequences of disrupted signaling. The pursuit of frictionless platforms, while appealing, risks obscuring the very signals that ensure a functional allocation of talent – a reminder that sometimes, the most elegant solutions are not about removing constraints, but about understanding what they protect.

The study reveals a concerning paradox: attempts to optimize job market signaling through readily available generative AI ultimately diminish its efficacy. This erosion of meaningful signals creates new tensions, shifting the equilibrium towards less efficient matching and potentially lowering the overall quality of hires. As the paper demonstrates, optimizing one aspect of the system – the ease of application – creates unforeseen consequences elsewhere. Friedrich Nietzsche observed, “There are no facts, only interpretations.” This resonates with the findings, as the perceived value of written applications – the ‘fact’ of their signaling ability – is revealed to be a constructed interpretation, easily manipulated and thus devalued by technological advancements. The system’s behavior over time demonstrates that a focus on surface-level optimization can obscure deeper systemic vulnerabilities.

The Signal and the Noise

The observed erosion of signaling value in written applications, while troubling, merely highlights a perennial problem: information asymmetry. This work correctly identifies the symptoms – a shift towards less-skilled hires and diminished welfare – but the underlying disease is not the technology itself. Rather, it is the continued reliance on easily-spoofed credentials as proxies for genuine capability. The digital labor platforms, once touted as efficient matchmakers, are revealed to be vulnerable to the same informational deficiencies as their analog predecessors – they optimize for convenience, not accuracy.

Future research should move beyond quantifying the damage and focus on designing mechanisms that incentivize honest signaling. The current trajectory, optimizing for ‘cheap talk’, suggests a need to explore alternative assessment methods – dynamic, skill-based evaluations that are resistant to manipulation. It is worth noting that every abstraction leaks; a writing sample, a degree, even a skills assessment, are all imperfect signals. The true cost of freedom – in this case, the freedom to cheaply apply – is the increased noise in the labor market.

Ultimately, the field needs to address the fundamental question: how does one reliably identify productive capacity in an increasingly complex world? The answer will likely not be found in more sophisticated algorithms, but in a deeper understanding of the trade-offs between signal fidelity, assessment cost, and the inherent limitations of any proxy measure. Good architecture, in this case, is invisible until it breaks – and the current system is showing clear signs of strain.

Original article: https://arxiv.org/pdf/2511.08785.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Signal and the Noise

See also: