Read Past the Polish

Put three documents side by side: a cover letter from a candidate, a proposal from a vendor who wants your next project, and a business case from one of your own managers asking for budget. A few years ago you could learn something from how good each one was. Today all three probably read well, which is why they may tell you less than they used to.

A document works as evidence when it costs a capable person less to produce than it costs anyone else, which economists call a costly signal. A strong candidate could write a letter that answered your job ad point by point because they understood the job. For weaker candidates, the same letter was harder work, so fewer of them bothered. You didn’t need to read minds, because the letter did some of the sorting for you.

AI writing tools shrank that cost difference. A tool that reads your job ad, or your request for proposals, and drafts a tailored answer will do it for anyone who asks, whether or not they could do the work. The document still tells you what’s on offer. It probably tells you much less about who is offering it.

Seth Godin calls this the smush. “Most forms of human expression are smushed by a decent AI,” he wrote in July 2026, and he argues that it “harms the market for pretty-good or even very-good human work.”

What happened on one hiring platform

Two research teams studied the same marketplace, Freelancer.com, where businesses post short, mostly remote jobs and workers apply with a short written proposal and a price. In April 2023 the platform added a tool that read the employer’s job post and drafted a proposal tailored to it, which the worker could send as it was or edit first.

Anaïs Galdin of Dartmouth and Jesse Silbert of Princeton looked at roughly 61,000 coding jobs and 2.7 million applications posted between 2021 and mid-2024 (arXiv, November 2025). Before ChatGPT, a proposal scoring about 3 points higher on the researchers’ 18-point tailoring scale, roughly the typical spread between proposals, went with a 30% to 40% higher chance of being hired, among the applications an employer actually considered. After the platform’s tool arrived, that edge fell to about 10%. Proposals got more tailored on average even as workers spent less time on them.

Jingyi Cui, Gabriel Dias and Justin Ye at Yale came at it from another direction, with 5 million cover letters for PHP programming and internet marketing jobs on the same platform (arXiv, November 2025). After the tool launched, the link between how closely a letter matched the job and getting a callback fell 51%, and the link to getting an offer fell 79%. The tool also narrowed the gap between writers: those who had written the most tailored letters before gained about half as much tailoring from it as the weakest writers did.

What it costs you to keep reading polish

Cui and colleagues found no evidence of a change in overall hiring or job completion on the platform after the tool arrived, a result they call preliminary. As far as their data show, jobs still got filled and still got done, which looks reassuring at first.

But the employers weren’t standing still. They rewarded tailoring far less than before and leaned more on workers’ review scores, the platform’s summary of past jobs. In the Yale data, the link between those scores and callbacks rose 5%.

If you keep reading polish anyway, Galdin and Silbert’s data show what you’d be choosing on. After the tool arrived, how tailored a proposal was no longer predicted whether the job was completed with a five-star rating. The effort the worker put into the application, measured by the time they spent on it, still did.

So the main cost of reading polish isn’t fewer hires. It’s who you pick: you’d be choosing on something that no longer tracks how the work turns out, while the thing that still does is hard to see in the finished document. In the Yale data, the same worker was more likely to get an offer when they spent longer editing the tool’s draft, though most AI-drafted letters went out with little or no editing and that link is only a correlation.

Whether that costs you money depends on how you pay. Galdin and Silbert built a model of the platform from its pre-AI data, and in a version where writing carries no signal at all, employers came out about even: they hired weaker workers more often, but those workers also bid lower, so employers paid less and got less. When the price is set before you choose, as with a salary band or a fixed project budget, picking the weaker option doesn’t lower it, so that offset may not reach you.

The fallback employers on the platform reached for held up better than the proposal did. In Galdin and Silbert’s data, a worker’s reputation score still predicted a five-star job after the tool arrived, and the Yale team says leaning on review scores helped keep screening working in the short run.

That fallback has two limits. The Yale team notes it may favor established workers at the expense of newcomers trying to get in, who have no record yet. And in Galdin and Silbert’s model, a worker’s basic profile (where they were from, when they applied and their platform reputation score) explained only about 3% of the variation in ability. So on that platform the score was a real signal, but a weak one.

Re-price the evidence you decide on

Both studies are about short, mostly remote freelance jobs on one platform. The method below applies the same mechanism to your own hiring and to the other documents on your desk, and that step is our reasoning rather than a research finding: any document you trusted because it was hard to write well is likely in the same place now.

Three steps, in this order.

1. List the documents your decisions rest on. The applications for your next hire, the proposals for your next supplier, the business cases that arrive with a budget request. Write them down, because leaning on polish is invisible until you do.

2. Ask one question of each: does it still cost a weak sender more to produce than a strong one? If an AI tool can now make it just as good for both, it fails. Only the documents that fail need anything changed.

3. For each one that fails, decide what it’s for. If you still need what it says, keep it as information: the vendor’s scope and price, the business case’s numbers. If you were using it as proof that the sender can deliver, replace it as proof. Often it’s both, and you keep reading the document while asking for the proof somewhere else.

The replacement is evidence that still costs the capable less than everyone else. There are three kinds, and each has a price.

• Checked. A track record you verify yourself, by calling the reference or looking at work they delivered. It’s closest to the one fallback the research saw still working, the platform’s reputation score. The price is that it only exists for people and suppliers who’ve done the work before, and the research never tested calling a reference.
• Live. They respond to your specific situation in front of you, where the thinking has to be theirs. For a candidate, that’s a real problem from your business worked through with you, a live version of the test in What Stays Expensive. For a vendor, it’s a working session on your actual numbers instead of a deck. For a manager, it’s the business case defended in conversation, including the question they didn’t prepare for. The price is your time.
• Committed. They agree to be judged on delivery before you commit in full: a paid small first phase before the full contract, payment tied to a milestone, a manager naming the number they’re willing to be measured on. The price is a slower start.

Start with the record wherever a solid one exists, because it’s the cheapest proof you’ll get and the closest to what the research saw still working. It narrows the field, but it rarely settles the question. For a decision that matters, follow it with a live session or a committed first phase, and where the record is thin or missing, as with a newcomer or a first-time vendor, go straight there. Go live when you’re judging someone’s thinking, and committed when you’re judging delivery, whoever promises it. Galdin and Silbert close their paper pointing toward the live and committed kinds, screening “that cannot be gamed by generative AI” and contracts that allow “exploratory short-term hiring” so employers learn ability on the job.

Notice what isn’t in the method: any attempt to spot AI in the document. The Yale team points out that detection tools are error-prone and unreliable. Even a perfect detector would answer the wrong question. What matters is whether the evidence in front of you still costs the capable less than everyone else.

And the same polish problem turns up when the AI answers the strategic question itself rather than drafting a document for someone else, which is where confident first answers get accepted too easily.

Where this stops being certain

The evidence comes from two working papers on one freelance marketplace, where the jobs were short, mostly remote coding and marketing work. Full-time hiring, vendor selection and internal budgets involve more contact and more context, which may keep signals a platform proposal can’t. The editing result is a correlation rather than proof that editing causes success.

What the studies point to is the direction. Once a tool made a tailored application cheap for anyone to produce, Galdin and Silbert found it stopped predicting how the work would go, and the Yale team describes employers shifting toward “harder-to-fake indicators of quality.” The documents on your desk are likely heading the same way, and the evidence worth reading now is whatever is still expensive to fake: a record you can check, time spent with you on your problem, something put at stake.