DeepMind Proposes Taxonomy for AI-Assisted Math Proofs

DeepMind proposed a taxonomy for classifying AI-assisted mathematics by significance and degree of AI contribution, following consultation with the mathematics community, per Google DeepMind. The credit question matters: if a model does 80% of a proof, who gets the paper, and how do you measure that 80%? Nobody has a clean answer yet.

The taxonomy arrived alongside two new papers built on Gemini Deep Think. That's DeepMind's effort to push its models past benchmark scores and into open problems that don't come with a known answer.

Benchmarks just check whether a model lands on the right number. Open problems need someone, human or machine, to convince other mathematicians the proof actually holds. That's a different kind of credit than acing a test.

Software rarely fights over who wrote a function once a model drafts it and a person edits it. Math can't shrug that off the same way, because a proof's value rides on who's vouching for it. I'd watch whether the next Deep Think paper forces the question: a model-authored proof nobody had to touch, credited how?