Why did AI tools break the effort signal in hiring?

The second finding concerns why the pre-LLM signal carried information at all. The model's core mechanism is that "higher-ability workers face lower costs of exerting effort to produce signals," so signals rise with ability on average and employers can read them as evidence. Before LLMs the chain ran in one direction: signals predicted effort, and effort predicted "workers' ability to complete the posted job successfully," which is why employers paid for signals. The platform's click data let the paper measure effort as the time each worker spends on an application.

After mass adoption, the excerpt reports three patterns. Employer willingness to pay for higher-signal workers "falls sharply." Proposals written with the platform's native AI tool "exhibit a negative correlation between effort and signal." And signals "no longer predict successful job completion conditional on being hired." The tool's use is identified from platform data, and a before-and-after comparison dates the break. The excerpt gives no size for the effort-signal correlation, no completion estimates and no comparison group, so the reported sign is the only figure available for the mechanism.

This finding is the mechanism beneath the counterfactual in the sibling note Does cheap writing weaken hiring based on worker ability?: if the signal stops carrying effort, the sorting it supported goes with it. The nearest library note, What makes accountable judgment scarce when AI cognition is cheap?, treats institutions that sort people as the scarce thing when cognition is cheap. The costly written application is one such institution, and this note shows what made it work. Does AI turn freelance work into validation instead of creation? argues that freelancers build skill through paid client work. This excerpt concerns the same kind of paid work as evidence to employers, so the two describe one platform setting from different sides: formation and screening.

The excerpt does not establish whether AI-tool proposals differ because of the tool or because of the workers who chose it; selection into tool use is not discussed. The implication is that the weakened signal is a documented pattern in this platform's data, and that reading it as an LLM effect is the paper's inference from timing and model structure. It does not license a general claim that AI-written applications carry less effort.