Does cheap writing weaken hiring based on worker ability?

The paper argues that LLMs, by making writing cheap, erode the written application as a costly signal of worker quality, and that losing the signal makes hiring less meritocratic. The evidence comes from coding jobs on Freelancer.com, where each application is a short written "proposal" plus an asking wage. Before LLMs, the excerpt reports, "employers had a high willingness to pay for workers with more customized applications": a worker with a one standard deviation higher signal had the same chance of being hired as a worker with a $26 lower bid. After mass adoption, "these patterns weaken significantly or disappear completely." The meritocracy result is a simulation, not an observed market. In a counterfactual equilibrium where writing carries no signal, workers in the top quintile of the ability distribution are hired 19% less often, and workers in the bottom quintile 14% more often.

The model embeds Spence-style signaling in a discrete-choice demand model and, from the workers' side, a scoring auction. Its core mechanism is that higher-ability workers face lower costs of effort, so they send higher signals on average, and employers read those signals as evidence of ability without perfectly inferring it. The paper gives three channels for the simulated loss. Employers lose the means to discern ability. Because ability and cost are positively correlated, wage competition leaves lower-ability workers holding the winning low bids. And since observable characteristics predict ability poorly, employers have little else to go on. The simulated changes are a 5% fall in average wages, a 1.5% fall in the hiring rate per posted job, and a 4% fall in worker surplus, against an employer surplus gain of less than 1%. Workers' losses are partly offset because their writing costs are now zero.

The paper positions itself against experimental work on AI and signaling, which it describes as partial equilibrium; its contribution is a market-wide measure of the disruption and of its effect on hiring and welfare. The nearest library note, What makes accountable judgment scarce when AI cognition is cheap?, holds that outcomes depend on the institutions that sort people once cognition is cheap. This excerpt names one such device, the costly application, and shows how cheap generation erodes it. The mechanism behind the simulation is taken up in the sibling note Why did AI tools break the effort signal in hiring?. Does AI turn freelance work into validation instead of creation? treats paid freelance work as where skill is built; this excerpt treats the same work as where employers read ability.

The excerpt does not establish several things. It covers one platform, one job category and one writing tool, which the platform introduced in April 2023. It describes the tailoring measure only as an LLM approximating human judgment, with no validation figures, sample size or date range. The welfare numbers are model output whose identification rests on assumptions the excerpt summarizes but does not test, and the introduction's motivating cases, cover letters and college essays, are not studied. The defensible claim is therefore narrower than the title: on this platform, once written applications stopped signaling, sorting by ability weakened, and a simulation puts a number on the loss. Whether the loss appears in other hiring markets, or at the size shown here, is open.