When AI makes every cover letter polished, what do employers look at instead to tell who'll actually do good work?

This explores what employers do once a well-written cover letter no longer tells them who will do good work. The best evidence comes from one natural experiment. Freelancer.com gave workers an AI tool that wrote tailored proposals, and the old link between letter quality and getting hired weakened sharply. The link between how well a letter matched the job and whether the worker got a callback fell by about half (Does AI cover letter writing change what employers value?, Does AI-generated cover letter access weaken hiring signals?). Overall hiring held steady, but employers changed what they looked at. They moved to **prior work history and on-platform reputation**: a record of finished jobs that a chatbot can't produce on request.

Why letters stopped working matters for judging what replaces them. A careful letter used to signal ability because it took effort, and able workers found that effort cheaper to spend. AI tools broke that link. Effort no longer tracked quality, and letters stopped predicting whether a job would be completed (Why did AI tools break the effort signal in hiring?). Losing this signal isn't harmless. In a simulation with no written signals at all, top-performing workers got hired 19% less often and the weakest got hired 14% more often (Does cheap writing weaken hiring based on worker ability?). Work history helps, but it can't help newcomers who don't have one yet.

Kessler's proposal splits the replacement signals by job (Can hiring signals survive when AI makes cover letters worthless?). For **quality**, use someone else vouching for you, because a recommender puts their own reputation at stake. That half has evidence behind it: recommendations measurably improved hiring. For **genuine interest**, use something scarce, like an in-person meeting or a networking conversation, since you can only spend your time on a few employers. That half is still untested. There's also a less obvious signal hiding in the click data. Workers who spent more time *editing* their AI drafts were hired more often, even though most people barely edit at all (Does editing time on AI drafts predict hiring success?). The cost has moved from writing to revising, and revision can still carry information.

Some other candidate signals look weaker than they seem. Listing AI skills on a CV raised interview invitations by 8–15 percentage points, but certificates added little over simply claiming the skill. Recruiters are rewarding a claim they don't verify (Do AI skills help candidates get more job interviews?). That makes it the next cheap signal waiting to be gamed. Polish itself also keeps misleading people: evaluators rated AI-written documents as both more human and better than real human submissions (Does polished writing actually signal better quality work?).

The twist is what happens when employers hand screening to AI. LLM evaluators prefer resumes they rewrote themselves, and they do this because of matching style, not better content. The bias gets stronger in larger models (Do language models favor resumes they rewrote themselves?). LLM judges are also easily swayed by authority cues and rich formatting, including fake references (Can LLM judges be fooled by fake credentials and formatting?). So if a machine reads the recommendations and work histories, those signals could be faked in exactly the way letters were. The signals that last are the ones tied to a real person's accountability or a verified record, not to anything that can be produced as text.