OpenAI Published 722 Math Papers. Nobody's Checked Most of Them.

If an AI tool handed you 722 reports this week and told you they were all correct, how many would you actually open? That's the question OpenAI's latest release puts in front of anyone who reads AI output for a living — which, increasingly, is everyone.

What happened

On October 6, OpenAI published 722 AI-generated math manuscripts to GitHub, produced by an unreleased model working through roughly 4,000 open problems. The post climbed to the Hacker News front page with over 1,300 points and nearly 1,500 comments. The catch, flagged repeatedly in that thread: many of the results lack formal Lean proofs, and independent mathematicians hadn't reviewed most of the claims by the time the batch went public. The papers read as complete, cited work. Whether the underlying math holds is a separate question nobody had answered yet.

722 is not a typo and not an edge case. It's what happens when a model can generate publishable-looking output faster than any team of humans can check it — a problem that shows up in code review, market research, and legal summaries just as easily as it shows up in math.

The skill this breaks

The instinct that fails here is treating output volume as a quality signal. A stack of 722 papers feels more credible than one, because it looks like the work of a large, careful team. It isn't. It's the output of one model run many times, and each individual result carries the same unverified status it would if it were the only one.

The skill that holds up is deciding, before you look at the output, what "verified" actually means for this specific claim — a named source, a reproducible method, a check you can run yourself — and then checking the one piece you're about to act on, rather than skimming the pile and inferring the rest.

Monday morning: a research summary with footnotes

Priya runs market research for a 60-person B2B software company. Her VP asked for a competitive pricing summary ahead of the meeting, and she fed six competitors' pricing pages and two analyst reports into Claude, asking for a synthesis with sourced claims. What came back was thorough — eleven footnoted statistics, a clean table, a confident recommendation.

She picked three numbers that would directly shape the pricing decision and traced each one back to its cited source before the meeting. One checked out. One was a rounding of a stat that actually said something narrower than the summary implied. One footnote pointed to a page that no longer said what the AI quoted. She fixed the deck, not the AI's confidence problem — that's hers to manage every time.

A second account: the audit nobody asked for

Marcus leads a four-person data team at a regional healthcare analytics firm. When a vendor pitched an AI tool that could "validate" clinical coding rules against a regulation, he didn't ask whether it was accurate — he asked what counted as a pass before he ran a single test case. He wrote three criteria down first: the rule had to cite the specific regulation section, match the firm's existing audit log, and flag (not silently resolve) any conflict with a second source. Only then did he run the tool against twenty real cases.

Nine passed clean. Six needed a human to resolve a conflict the tool had flagged honestly. Five looked fine and weren't — they cited the right regulation but misread which clause applied. Writing the criteria first is what let him catch that last group; checking after the fact would have meant grading against a standard he hadn't actually pinned down.

The takeaway

722 unverified papers isn't a story about one AI lab moving fast — it's a preview of how much confident, unchecked output is about to land on your desk, and the only defense is deciding what counts as checked before you're holding the stack.