AI Can Create Faster Than You Can Review

Here is where we are. Your team can make more stuff than ever. Code, reports, research papers, whole homework assignments. The making got cheap and fast. But the checking did not. Nobody bought a subscription that made them better at telling good work from plausible junk. That gap is where your risk lives now. Let me catch you up.

The deep cut

• Production is bought. Judgment is earned. You can buy an API key in an afternoon; you cannot buy the eighteen months that teach a bad architecture from a good one.
• Fluent output beats non-adversarial review by default. ICML doubled to 6,352 accepted papers while reviewers stayed volunteers, and one spotlight reviewer admitted he never checked the proofs.
• Uncertainty hardens into fact when nothing checks it. A 0.51 classifier score became a durable "probable duplicate charge" record with nothing in between failing.

The flood already arrived

ICML 2026 took in 23,918 submissions and accepted 6,352 papers, roughly double the year before. Reviewing did not double. One accepted spotlight paper carried a reviewer's own words: "My low confidence score is because I did not check all the proofs carefully." When Hugging Face ran a hackathon to reproduce the conference, agents found the growth term in that paper's proof was wrong, confirmed at roughly nine sigma.

That pattern showed up across the audit. Of 2,226 papers attempted, 23% had at least one claim falsified or contested. In 242 papers, independent teams reached opposite verdicts on the same claims. The work looked done. Checking it was the hard part, and the checking got skipped.

Why judgment did not scale with output

Production capacity transfers instantly and improves every time a lab ships. Judgment does not. As one writer put it plainly, judgment is "compressed experience of being wrong, and there is no way to buy the compression." You learn a bad approach by living inside one, not by reading a spec.

This hits your juniors hardest. The old path to senior ran through years of producing mediocre work under review. Agents now do that tier competently and instantly, so the economic reason to pay a human for it is gone. But someone still has to review a volume of code no prior era produced. You cannot skip to the reviewing tier. Reviewing is the thing you learn by producing.

Plausible is not the same as checked

When a tool makes fluent work cheap, plausibility stops carrying signal. That is the real lesson from the 3M case. An expert witness asked ChatGPT to "create an exceptional expert witness report" showing the company was "0 percent" at fault, then uploaded a photo of the gas detector at the center of the case and asked, "what am I looking at?" The report read like expert work. It was not. The court subpoenaed 350 pages of his prompts, and the whole thing fell apart.

Same shape in classrooms. Students now point agents at Canvas and Blackboard to sit entire courses, and none of the three top chatbots would refuse to write a full paper. Any review process calibrated on human effort is now measuring nothing.

The 0.51 that became a fact

Here is the quiet version of the same problem inside your own systems. A support agent reads "I was charged again after I changed my plan." The classifier splits 0.51 duplicate charge, 0.46 plan change. The system picks duplicate charge and moves on. Retrieval pulls duplicate-charge policies. The prompt states it as fact. A tool compares two charges. Memory stores "probable duplicate charge" as case history.

Five steps later a human reviewer sees a pre-labeled case and trusts it. Nothing in the chain failed. Each part did its job. The five-point gap got laundered into a durable record. The fix is not a better model. It is deciding how far an uncertain output travels before something checks it, and refusing to let a narrow margin reach anything irreversible.

What the teams getting it right actually do

The Princeton reproduction study is the encouraging part. Agents ran hundreds of experiments and reviewed the literature, but their papers got rejected from NeurIPS. Kapoor's read: they were "unambiguously bad" at the research itself, committing to weak approaches and unable to backtrack. The engineering scaled. The judgment did not.

So the teams running agents at the highest volume are the ones rebuilding friction on purpose. Detailed specs that force judgment before generation, not after. Adversarial review, where a named person's job is to find what is wrong, not to nod at what looks fine. In the Hugging Face hackathon, the most reliable results came from a human steering the agent, re-pointing it and killing bad premises before a week of compute burned. Manage the intelligence like a PI manages grad students.

Three questions for your team

• Can you name the person whose job on each shipped artifact was to find what was wrong with it, not to check that it looked fine?
• Where does an uncertain model output get written into durable state or an irreversible action with nothing in between, like the 0.51 that became a refund record?
• Which problem classes are you keeping human-only on purpose, so your juniors still produce work under review instead of only checking agent output?