Humans outperform AI in 'First proof' test published in Nature
Technology
Turns out, humans still have the edge over AI when it comes to tricky math problems.
In the "First Proof" test published in Nature on June 12, 2026, four advanced AI systems tackled 10 research-level questions they'd never seen before.
The top AI managed to solve six, but every human mathematician solved their assigned problem without a hitch.
AI systems erred and omitted citations
The test was set up to really challenge AIs: no shortcuts or old data allowed.
It revealed that while humans handle new puzzles pretty well, AIs often got stuck or made mistakes with fresh challenges.
Plus, some ethical concerns popped up since the models didn't always cite their sources.
Another round of testing is coming this fall as researchers keep pushing to see how far AI can go in math.