A hiring loop once ran on human attention. Somebody read the resume, somebody ran the screen, somebody made the call. Two ends of that loop now run on the same model, and the middle has stopped carrying information.
The employer side runs a screener. It reads ten thousand applications in the time a recruiter reads ten, and it scores each against the job description before a person ever sees it. The candidate side runs a writer. It drafts a tailored resume and cover letter for every posting, tuned to whatever the screener is known to reward. Two machines optimise against each other, and the result is noise: a thousand near-identical applications written by models and read by models, where neither side can tell which came from a person.
The screen and the writer cancel out
A screener looks for signal: the right words, the right tenure, the right title, the right gaps. A writer produces exactly that signal on demand. The moment the signal is cheap to fake, it stops being signal. This is not a new problem. Recommendation letters stopped meaning anything when anyone could ask a friend to write one. The difference now is scale and symmetry.
Hiring was already a poor filter. What changed is that both sides mechanised at once, so the noise is now manufactured in bulk instead of by accident. A human screener at least occasionally spotted the person behind the words. A model tuned to past hires learns the shape of the hire, not the substance, and it reaches that shape faster than a human ever could.
Proof of work is the thing left standing
What survives a machine-written market is the thing a machine cannot generate on your behalf: evidence that you built something.
- A shipped product beats a bullet point. Anyone can claim they led a migration. Fewer can point at the repo, the release, the post-mortem.
- A public artifact beats a private claim. A talk, a commit history, a working app, a design system with your name on it. These carry timestamps and authorship a prompt cannot fabricate.
- A reputation beats a resume. People who have worked with you will answer a phone call. A model cannot produce that reference.
This is the "portfolio over polish" move. It responds to a specific, dated change: the marginal cost of a polished application dropped to roughly zero in the last two years. When polish is free, polish stops being the thing you pay for.
Both sides brought a machine. The person who brought proof of work wins.
Where the machines actually fail
The strongest argument against the machine-written loop is that it misses. Screener models are trained on who got hired, so they encode whatever bias was already in the hiring. A model that learns to reward the shape of a Stanford CS resume keeps rewarding that shape and calls it merit. The writer on the other side feeds the same shape back, which makes the screener more confident and less accurate.
The deeper failure is the target. Screening is automated against a category nobody can define. A good hire is a bet that changes with the team, the quarter, and the problem. A model can optimise for a proxy of that bet. It cannot hold the bet itself, which is why the loop degrades into an arms race over proxies: keywords, tenure, schools, titles.
What to ask for instead
If you are hiring, ask for the artifact. Ask for the repo, the diff, the shipped thing, the decision someone made and why. These are hard to fake in bulk and cheap for a builder to produce.
Recruiters describe the shift in plain terms. A resume that reads like every other resume tells them nothing, and a portfolio they can open tells them plenty. The ones who still hire well have learned to scroll past the prose and look for a link that loads, a commit that predates the posting, a thing that runs.
If you are the candidate, stop feeding the machine. Spend the hour building something public instead of drafting another tailored cover letter. The artifact will still be there after the screener changes, after the keywords rotate, after the model is retrained.
The two machines are not going away. Screening will keep screening and writing will keep writing. The people who win are the ones who put a thing in the middle that was built, that can be checked, and that no model wrote for them.