By the Smili Phone team
On October 6, 2026, OpenAI dropped what looked like a landmark achievement in AI research: 722 preprints published to a public GitHub repository, each claiming progress on one of 372 unsolved or hard mathematical problems. The manuscripts were generated with help from an internal mathematical-reasoning system the company had never announced before. Mathematicians started digging in almost immediately.
Less than 24 hours later, three of those papers were gone. OpenAI withdrew them on October 7 after finding a sign error in one proof, a mistake in a stabilization-trace cancellation argument, that cascaded through two other papers built on the same construction. Fourteen more manuscripts were revised in the same window. Retraction Watch broke the story on October 8, and the episode turned into the clearest real-world test yet of a simple question: how much can we trust AI-generated work that looks polished?
What the repository really contained
The scale was the headline, but the fine print mattered more. OpenAI's own repository warned that not every result had been fully verified at release. The collection kept a running tally of which results had been checked with Lean, the open-source proof assistant mathematicians use to machine-check formal arguments. Reporting after the withdrawal put the Lean-verified share around 42 percent, and different outlets could not even agree on how to count it. In plain terms, most of the 722 papers were resting on human review when the error was caught.
Cornell mathematician Alex Townsend told Retraction Watch that OpenAI should have led with the Lean-verified manuscripts first, given the field's skepticism toward AI-generated math. Rutgers mathematician Alex Kontorovich reacted to the proofs with a mix of awe and alarm. Both reactions point to the same tension: the volume of claimed progress was thrilling, but presenting the unverified majority alongside the verified minority invited exactly the kind of scrutiny that found the error within a day.
The lesson for the rest of us
You do not need to read algebraic geometry to take something from this. A sign error is a small, mundane mistake, the kind that has embarrassed human mathematicians forever. The AI made it look authoritative, and 722 manuscripts of confident prose went public before anyone checked. That is the pattern to remember the next time an AI assistant hands you a perfectly formatted answer on your phone: confidence is not the same as correctness.
The practical rule is simple. For trivia and brainstorming, AI is fine on its own. For anything that matters, homework, a work report, health or money questions, verify. Ask the model to show its work and cite its sources, cross-check the key claims somewhere independent, and run anything critical past a real expert. AI is a powerful starting point, not a finished product.
These tools are only getting more capable, and they are reaching small businesses and students everywhere, including here in Escondido. If an AI assistant on your phone ever gives you an answer that smells wrong, or an app starts behaving strangely after an AI update, bring it by 1882 W El Norte Pkwy and we will help you sort it out.
Related: OpenAI's invisible watermark for AI text · the FTC probe into rogue AI agents
