News · 2026-07-13

AI is now solving hard math and physics problems faster than humans can formally check them

A distinct pattern is emerging from the recent run of AI-cracks-a-hard-problem stories: artificial intelligence is now producing candidate solutions to difficult math and physics problems faster than human experts can formally verify them. The result is a 'verification lag' - a temporary truth vacuum where compelling headlines outpace peer-reviewed proofs, and the right response is neither dismissal nor hype but patient checking.

Key facts

• Theoretical physicist Yuji Tachikawa reported that Anthropic's Fable provided a mathematical bridge that unblocked a six-month research problem.
• A GPT-5.6 solution to a 50-year-old Erdos problem is circulating without a formal peer-reviewed proof.
• The cadence of such 'solves' is accelerating, but formal verification remains slow.
• The gap between AI output and human verification creates a temporary period where claims lead confirmation.

Start with the strongest case, because it clarifies what is genuinely new. Yuji Tachikawa, a theoretical physicist, reported that Fable solved a problem that had blocked his research for six months. What makes this high-signal is not that an AI produced an answer - it is that a named domain expert, someone fully capable of judging the work, publicly vouched that the model supplied a specific mathematical bridge his human collaborators had missed. That is a different kind of evidence than a leaderboard score. The value was not a lookup or a regurgitation; it was a genuine step in a real research problem, validated by the person best positioned to know.

Contrast that with the GPT-5.6 Erdos claim. Reports circulating in AI communities say the model solved a 50-year-old problem posed by the legendary mathematician Paul Erdos. That would be remarkable - but as of now it is circulating as a success story, not as a formally published, peer-reviewed proof from the mathematicians involved. (See our earlier coverage of OpenAI's math-conjecture claim.) The distinction is not pedantic. In mathematics, a proof is not 'true' because it looks convincing or because a smart system produced it; it is true because it has been checked, line by line, and survived. Until that happens, an AI-generated proof is a candidate, however plausible.

The mechanism driving the lag is straightforward. Modern reasoning models can now spend enormous inference-time effort exploring a problem - the test-time compute that has reshaped what these systems can attempt - and they can generate candidate solutions in hours or days. Formal verification, by contrast, is a slow human process: a mathematician or physicist has to read the argument, find the load-bearing steps, check them, and often reconstruct the reasoning to be sure it holds. Generation has sped up by orders of magnitude; verification has not. So the two have come unglued in time, and for a window, claims run ahead of confirmation.

The analogy: it is like a prospector who can now dig a hundred promising holes a day, while the assay office that certifies whether any hole actually contains gold still processes one sample at a time. The pile of 'possible gold' grows faster than anyone can confirm it, and in that gap, rumor and excitement fill in for verified fact.

Why it matters: this is a real epistemic shift in how science absorbs machine-generated results, and it demands new habits. The correct posture toward an AI-produced solution is the one a good referee takes: log the candidate, withhold the verdict, and wait for the check. The danger is a field that gets so used to impressive-looking AI output that it starts treating candidates as conclusions - exactly the 'fluency lulls the reviewer' failure that is haunting AI-assisted software this week too. The honest caveat runs both ways: some of these solves will hold up under scrutiny and represent genuine acceleration of research, and some will not survive review. Tachikawa's expert-vouched result sits at the credible end; the un-refereed Erdos claim sits at the wait-and-see end. Distinguishing them is precisely the discipline the verification lag now requires.

Primary source, verified: read the paper →

Ground Truth. (2026, July 13). AI is now solving hard math and physics problems faster than humans can formally check them. Ground Truth. https://groundtruth.day/news/ai-solutions-are-outpacing-human-verification.html

@misc{groundtruth:ai-solutions-are-outpacing-human-verification,
  title  = {AI is now solving hard math and physics problems faster than humans can formally check them},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/ai-solutions-are-outpacing-human-verification.html}
}

Topics: ai-for-science · mathematics · physics · verification · reasoning