Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-10 · high
The proposal targets incentives, not capability: a bare AI answer with no shown reasoning would count for little.
Why now: Tao raised the idea in a thread posted as of September 10.
Terence Tao wants unexplained AI answers to open problems to carry a warning label. In a Mathstodon thread, he proposes marking certain problems "analysis-required," so an AI-produced answer with no shown reasoning earns little credit even if it's correct.
The reasoning is about incentives, not accuracy. Tao says even a rumor that someone is working on a problem can now trigger a race to solve it with AI first. He warns this pressure may push researchers to stop sharing promising directions with the broader community at all, reversing centuries of open science norms in math.
There's a research-quality argument underneath the policy one. Mathematicians working on open problems have long paired goal-directed effort with curiosity-driven exploration, the tangents and dead ends that often turn into separate discoveries. Tao's concern is that AI turned loose on a problem without expert supervision optimizes straight for the stated goal. It captures the flag, but skips the detours where a lot of real math actually gets found.
The label itself is a soft mechanism. It doesn't block AI from working on a problem or verify how an answer was produced, it just changes how much a bare result is worth once submitted. That leaves the hard part to whoever runs the venue, a journal, a competition, an open-problem list, deciding whether to require an explained derivation and how to check one. Tao's thread doesn't lay out enforcement, and it's the piece that determines whether "analysis-required" changes behavior or just becomes a tag nobody checks.
Each link below shares sources, entities, or timing with this story.
In a four-post Mathstodon thread, he says problems are infinite but fruitful ones are not, the way a country can lack drinking water while surrounded by ocean. His mechanism: every new tool flattens a field's difficulty landscape, and the AI era is unusual in having no visible...
In a Mathstodon post, Tao argues pre-AI open problems are a finite supply of uncontaminated benchmarks: once a solution is published you cannot tell whether a later AI solved it independently or absorbed the answer in training. He adds that open problems have value for trainin...
The number that reframes everything isn't ten. It's two thousand. OpenAI published "Ten advances in mathematics and theoretical computer science" on August 1, claiming an internal version of Astra produced new results on ten problems that had seen no progress on the main resul...
The same man whose framework a model regression destroyed also published the most aggressive prediction of the week, and the tension between those two facts is the whole argument. "The Shape of Things to Come, Part 1: The Continuous Thunderdome" argues traditional CI/CD collap...
Deputy President Michael Easton ruled in Sadnan Khan's unfair-dismissal claim against ALDI that Khan's AI-generated legal advice was "plain wrong" and that he used ChatGPT as a quasi-legal advisor despite repeated warnings his case had no substantial prospects. Easton wrote th...
Quanta's August 3 piece tallies the assault: OpenAI found a counterexample to Erdős's 1946 unit distance conjecture on May 20, then Astra produced 10 further advances. Google DeepMind evaluated 700 open conjectures in January, solving four and recovering nine forgotten solutio...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.