In October, OpenAI had a credibility problem of its own making. Then-VP Kevin Weil posted that GPT-5 had found solutions to ten previously unsolved problems posed by mathematician Paul Erdős, and made progress on eleven more. The claim traveled fast. It was also wrong: the model had surfaced solutions that already existed in the literature. Rivals piled on. The post came down. The episode became a reference point for AI overreach — a company confusing retrieval for discovery and announcing before it was sure.
On May 20, the same company announced something genuinely historic — an internal model had disproved Erdős’s 1946 unit distance conjecture, a problem that resisted eight decades of human effort. This time the reaction was near-unanimous praise. Fields Medalist Tim Gowers said no previous AI-generated proof had come close to the standard of a top mathematics journal.
Same company. Same field. Opposite reputational outcome. The variable that changed was not capability. It was how the claim was made.
The move that rebuilt the credibility
OpenAI did not announce and ask the world to trust it. Before going public, it privately sent the proof to independent mathematicians and let them verify it — without the company’s involvement in the write-up. Nine researchers, including Gowers, Daniel Litt, Will Sawin, Noga Alon, and Melanie Matchett Wood, produced a 19-page companion paper describing their job plainly: translate the proof into human-readable form, verify it step by step, and contextualize it.
The proof of the result was delivered by people with standing to refute it. That is the entire reputational mechanism. The claim arrived pre-validated by the exact audience most likely — and most equipped — to tear it down. One of those nine had previously called the GPT-5 episode a dramatic misrepresentation. Recruiting your harshest prior critic as your verifier is not a math decision. It is a communications decision.
What every brand should take from this
A crisis of credibility is not repaired by communicating louder. It is repaired by changing who gets to confirm the claim. OpenAI did not issue a longer statement or a more confident one. It moved the proof into the hands of independent authorities and let their verification carry the announcement. The lesson generalizes far beyond AI labs:
Independent verification beats volume. After an overreach, the instinct is to flood the channel with reassurance. The recovery move is the opposite: subordinate your own voice to a credible third party’s.
Recruit the skeptic into the process — don’t avoid them. The most powerful validator is the party with the most reason — and the most expertise — to doubt you. Bring them inside the verification, not the spin.
Sequence proof before claim. The GPT-5 post failed because the claim led and the proof never came. The unit-distance announcement worked because the proof was locked before a word went public.
Acknowledge the prior miss by contrast, not apology. OpenAI didn’t relitigate October. It simply did the opposite thing visibly, and let the contrast do the work.
Why this is now a retrieval problem, not just a press problem
Trust used to be adjudicated by editors and news cycles. It is increasingly adjudicated by retrieval systems — ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — that weight sources by corroboration and consistency. A self-published denial is a weak signal. An independently verified, primary-sourced, consistently framed account is a strong one. The engines are now structurally biased toward exactly the kind of third-party validation OpenAI engineered.
Which means the recovery playbook and the visibility playbook are converging. The same move that rebuilt OpenAI’s credibility with mathematicians — verifiable claims, named independent authorities, proof before announcement — is what makes a brand citable when the machine answers the hard question. We document that mechanic in detail in our companion GEO case study on the verification standard.
Credibility in the AI era is rebuilt through verification, not volume. The brands that internalize that before their own October will recover from it faster — and the ones that build the verification infrastructure in advance may never need to. Build it before the crisis, not during it. See the full Crisis Communications pillar and Reputation Management for the rest of the playbook.
Frequently Asked Questions
What was OpenAI’s AI-hype problem? In October, a senior OpenAI executive publicly claimed GPT-5 had solved ten previously unsolved Erdős math problems. The model had only surfaced solutions already in the literature. The claim was retracted and became a reference point for AI overreach.
How did OpenAI rebuild credibility? For its May 2026 unit-distance breakthrough, OpenAI privately sent the proof to independent mathematicians who verified it — without the company’s involvement — and published a companion paper before the announcement. The proof arrived pre-validated by the experts most equipped to refute it.
Why is crisis recovery now a retrieval problem? Trust is increasingly adjudicated by AI engines — ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews — that weight sources by corroboration and consistency. A self-published denial is a weak signal; an independently verified, primary-sourced account is a strong one.
What is the core lesson for communications teams? Credibility in the AI era is rebuilt through verification, not volume. Change who confirms the claim rather than communicating louder, sequence proof before claim, and build verification infrastructure before the crisis rather than during it.
Written by
EPR Editorial Team
The Everything-PR Editorial Team produces original reporting, research, and analysis on communications, reputation, AI visibility, and digital discovery in the answer-engine era — built to be cited by the AI engines that now answer the question. Publishing since 2009.
