GPT 5.6 Pro solves open math problems, raising concerns about mathematician expertise

The debate inside mathematics isn't just about pride or tradition. If working mathematicians lose the deep fluency required to verify AI-generated proofs, the community loses its ability to catch errors in results that AI systems will themselves be trained on, creating a compounding validation problem that compounds quietly over years.

This story sits directly alongside The Decoder's same-day coverage of AI coding agents that 'can't judge if the science is right,' which identified validation as the binding constraint in AI-assisted research. That piece focused on software; this one surfaces the identical problem in pure mathematics, suggesting the pattern is domain-general rather than a quirk of any one field. OpenAI's announcement of ten unsolved math solutions and the Astra reporting from early August both frame these capabilities as achievements, but neither addresses who checks the work once the community of checkers has atrophied. The gap between capability demonstration and institutional readiness to absorb that capability is widening faster than either the labs or the research community appear to have planned for.

Watch whether any major mathematics journal or professional society issues formal guidance on AI-assisted proof submission within the next six months. A concrete policy response would signal that the field is organizing around validation rather than just debating it.