Mentatcurated
Artificial Intelligence medium · first-party

Proof indigestion

Machines now write and check proofs faster than mathematicians can understand them. Tao argues the obvious fix, training AI to smooth out the hard parts, is the wrong one.

Terence Tao splits mathematical work into three acts: generating a proof, verifying it, and digesting it — the slow communal process of actually understanding what a result means and why. AI is racing through the first two and leaving the third behind. His name for the result is proof indigestion: theorems that are produced, sometimes even machine-checked, but never absorbed. The concrete symptom is a queue. After GPT-5.5 helped crack one of the open Erdős problems, the site tracking those problems piled up nearly twenty claimed solutions marked 'pending assessment' — where before it saw one or two at a time.

The tempting fix is to point AI at the backlog: train it to rewrite proofs against a readability rubric until they score well on digestibility. Tao argues this backfires. Some difficulty in a proof is what he calls natural difficulty — the chewing that is the point, the part a reader has to work through to actually learn something — as opposed to artificial difficulty worth smoothing away. Optimizing a digestibility metric can't tell the two apart, so it planes off both. Follow the logic to its end, he writes, and you arrive at pureeing every meal into a feeding tube: technically no more indigestion, and no more nourishment either.

This is an argument, not a measured result — it rests on Tao's standing, not a study showing that readability tuning degrades proofs. But the frame travels. Swap 'proof' for 'code' and it describes a familiar unease: a machine that produces more correct output than you can keep pace with, optimized for passing checks rather than for the human who has to hold the thing in their head. The mismatch Tao is pointing at isn't that the machines are wrong. It's that being right, fast, and unreadable is its own kind of failure.

The lenses

Novelty 3
Impact · breadth 2
Impact · depth 3
Actionable 1
Delight 3
Substance 3
Hype 2

The facts

Open mathstodon.xyz →

How this connects

Tap a node to open it

How this connects