Image: Meta

Meta says its AI chatbot helped mathematicians answer five math questions that had no known answer when the work started. The work turned into six research papers, which Meta shared on its research blog.

Meta frames this as a step up from school-style tests. Earlier this year, its models hit gold-medal-level scores at five high school contests in math, physics and chemistry. But contest problems have an answer key. These problems didn't.

Following gold-medal-level performance from our AI models across five competitions in mathematics, physics, and chemistry, we asked a harder question: can AI contribute when a problem is genuinely open and without an existing solution path?

Over the past several months,…

— AI at Meta (@AIatMeta) October 2, 2026

The AI is Muse Spark, the "brain" that powers the Meta AI chatbot. It's not the same thing as Muse, Meta's separate assistant that does tasks for you, which we covered in our story on Meta's new gadget kit. The mathematicians used two older versions of it in Thinking mode, a setting that makes the bot take extra time to reason before it answers. Meta says they used the regular chat window at meta.ai, with no special research tools.

So what did the AI actually do? Here are three examples, in plain terms:

• The black swan. A mathematician had guessed that a certain kind of math structure always has a second property. One exception is enough to prove a guess like that wrong, the way one black swan proves not all swans are white. Muse Spark wrote a search program that found the exception.
• Dots and an oval. Picture random dots scattered on a page, and you try to draw one oval that touches every dot. One paper pins down the exact tipping point where that stops being possible as you add more dots.
• A collapsing wave. In a math model inspired by lasers, one paper proves that a certain kind of wave must collapse in a limited amount of time. That question had been open since 2015.

The humans stayed in charge

This is the part worth noticing. Mathematicians picked the problems and steered the work. A second group checked it. Each paper marks which parts the AI drafted and which parts people wrote. In one paper, Meta says the researchers "corrected gaps" in the argument. In other words, the AI made mistakes, and people caught them.

There's some fine print, too. Meta describes the reviewers on one paper as mathematicians "on our team." It doesn't say whether outside journals have reviewed any of the papers yet. And Meta wasn't alone. A Yale team posted its own proof of the oval problem in August. An AI system called Nilradical reported a different black swan in September. Meta credits both.

That caution makes sense. In 2025, OpenAI staff said GPT-5 had solved 10 famous unsolved math problems. The mathematician who keeps that list called it "a dramatic misrepresentation." The AI had actually found old papers that already had the answers, TechCrunch reported.

What it means for you

Our take: there are two lessons here for regular people. First, the chatbot on your phone is getting better at hard, step-by-step thinking. Meta says Thinking mode is in the Meta AI app and on meta.ai, not locked away in a lab.

Second, trained mathematicians still had to check the AI's work. So if you ask a chatbot to check a loan payment, plan a monthly budget or help with your kid's algebra, treat the answer like a smart friend's first draft. It's useful, but check it before you trust it. And Meta's habit of labeling which parts the AI wrote is one other AI companies should copy.