- Greg Brockman described AI results that can be too hard for people to check, and debate between AIs as one way to keep humans in charge of oversight.
- Our reading: this shifts human skill from making to judging, and a board-member role only works if people can tell whether output is what they wanted.
Remarks on a No Priors episode, published 15 October 2026, point to a shift in human work. When AI produces proofs and answers people cannot easily check, the skill that counts moves from making to judging. Greg Brockman, OpenAI's president, described oversight methods such as debate. A speaker near the end of the show, whom the transcript does not name, added a picture of people as board members.
What was said
The hosts noted that few people can judge an advanced theorem, even one written by a human. Brockman replied that mathematics already has proofs nobody finds illuminating. His example was the four color theorem, settled when a machine worked through a large number of cases. He said of that result, "there's no beauty in that."
He also explained why checking is hard. An arithmetic answer is easy to compare with the right one, and so is a formalized proof. The natural-language proofs his models produce are far harder to verify. He said OpenAI had published almost 400 important open problems now resolved.
He then described debate, an idea discussed since 2017. He framed the general problem as AIs solving tasks too hard for humans to check. Take an AI-written book summary. You cannot judge it without reading the book. But a second AI can try to show the summary is bad, for instance by pointing to a major plot point it left out. You can judge that argument without reading the whole book.
Near the end, a speaker on the show said that in some fields "we become more like CEOs or board members." The test, that speaker said, is whether you are happy and whether this is what you wanted.
Why it matters
Our reading: the effort in AI work moves from producing it to verifying it. A team that cannot check what a tool returns is guessing when it approves. For managers and buyers, a plain question follows. How does this tool's work get checked, and by whom?
It also makes judging a skill that needs training. A reviewer who cannot follow the subject cannot tell a good answer from a plausible one. The board-member picture only works if the person in the seat can still tell whether the result is what they wanted.
The other side
Brockman did not say the published results are unverifiable. He said the output is not incomprehensible theorems. By his account, the mathematical community is reviewing the solved problems, and domain experts who posted about them describe surprising insights. The "too hard to check" case was the general problem that motivates debate, not a finding about those results.
He was also hopeful about the gap. He said he has some optimism that models better able to judge what humans know will help people follow their reasoning, and he saw promising signs. The excerpts give no evidence on whether that holds at scale.
Debate was offered as an illustration, not as proven practice. In the book example, the challenger is itself an AI, and the conversation did not say how people should decide whether to trust it. The board-member remark concerned human agency, not verification, and did not explain how a board member checks work that others understand better.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- No Priors: The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (2026-10-15)
