A year ago, the AI-math story was mostly hype and corrections. That has changed. Machines now produce research that top mathematicians take seriously. They also produce it faster than anyone can review it.
That second fact is the real story of 2026.
The Year Math Stopped Laughing at AI
The turnaround came fast. In October 2025, OpenAI claimed GPT-5 had solved 10 Erdős problems. Critics refuted the claim within hours. The model had mostly dug up obscure papers that already held the answers.
Then the results got real.
- Discrete geometry: An internal OpenAI model disproved Erdős’s planar unit-distance conjecture. It linked discrete geometry to deep algebraic number theory. A team including Noga Alon and Tim Gowers checked the proof, and Gowers rated it worthy of top journals.
- Number theory: A research version of Claude improved a decades-old bound on Riemann zeta zeros. It also produced a formally verifiable proof.
- Broad sweeps: OpenAI says GPT-6 Astra made progress on old problems in fields from coding theory to lattice cryptography.
- Formal proofs at scale: DeepMind’s Alethia model tackled several Erdős problems. Math Inc’s Gauss model partly formalised sphere packing in dimensions 8 and 24.
2026 at a Glance
| System | Result | Who vouched for it |
|---|---|---|
| OpenAI internal model | Disproved Erdős unit-distance conjecture | Alon, Gowers and co-authors |
| Claude (research version) | New bound on Riemann zeta zeros | Formal proof; Oxford’s James Maynard |
| GPT-6 Astra | Results across six fields | OpenAI’s own claims |
| DeepMind Alethia | Several Erdős problems | Bergen conference programme |
| Math Inc Gauss | Partial sphere-packing formalisation | Bergen conference programme |
Riemann: A Real Step, Not a Solution
The Claude result shows how this work actually unfolds. A staff member ran Claude inside Claude Code and asked it to attempt the full hypothesis. The model made the mathematical choices. It failed at the main goal. Along the way, it tested 650 dead-end ideas and used 60 agentic instances to explore and check the work.
The payoff came as a side result. Claude raised the known share of zeros on the critical line to 67.2%. Before going public, subagents hunted for counterexamples. They pulled 54 arXiv papers to rule out prior publication and re-derived the finding from scratch. Anthropic’s research note spells out the limits plainly.
Experts gave credit with clear caveats. Maynard said the problem needed “a new real idea,” and that this result seems to supply one. He also stressed that none of these methods opens a route to the full hypothesis.
That pattern looks a lot like human research. Big swing. Honest miss. Genuine byproduct.
The New Bottleneck Is Trust
Here is the twist. AI has made proofs cheap to produce. It has not made them cheap to check.
Dozens of AI-generated proofs now sit on erdosproblems.com. Many are probably correct. Yet no expert wants to verify and endorse them one by one.
Peer review grew up in a world of scarce results and slow authors. That world is ending. Three responses are emerging:
- Formal verification. Software can check proofs written in Lean, so tired referees no longer carry the whole load.
- Agent self-audit. Systems now red-team their own claims before release, as the Claude case shows. That habit matters, because researchers have already logged AI agents faking their actions in real deployments.
- Human curation. Some mathematicians picture a museum model. AI supplies the output; people select and interpret it.
None of these fully solves the problem. A formal proof confirms that a result holds. It does not explain why, and understanding is what mathematicians actually value. AlphaGo veteran Thore Graepel makes a similar point about chatbots lacking an auditable ledger of what they believe and why.
Mathematicians Are Split, Not Panicking
The mood is mixed. Carnegie Mellon’s Jeremy Avigad warns that AI will soon out-prove humans. Utrecht’s Johan Commelin compares the shift to the automation of hand calculation.
Daily practice is already changing. One Brandeis PhD student spends $200 a month on Codex. He uses it for literature searches, gap-filling and draft reviews. Fields medallist Deng Yu expects humans to build theories while AI handles the technical details.
Not everyone welcomes it. One group of mathematicians wants to reject AI and AI companies outright to protect the human role in the field.
The debate has gone formal. The University of Bergen hosted a conference on AI proving and autoformalisation on 7–8 October 2026. Harvard’s Lauren Williams speaks on 14 October about mathematicians putting AI to the test.
Why This Reaches Beyond Math
Math makes the cleanest test bed for machine reasoning. Answers are right or wrong. That explains why labs chose it. Progress here tends to carry over into other demanding work.
So the verification gap will not stay in math departments. Physics, drug discovery and chip design all face the same question. When an autonomous agent hands you a result, how do you trust it at scale?
Regulated industries have started answering already. Vendors pitching agentic AI to banks now lead with audit logs and guardrails, not raw capability.
The lesson from math is clear. Capability is outrunning accountability. The labs and journals that build fast, credible checking systems will shape what counts as knowledge.
The Bottom Line
AI did not solve the Riemann hypothesis. It did something stranger. It turned mathematics into a field with more candidate answers than qualified judges. The next breakthrough may not be a proof at all. It may be a better way to trust one.
Related: How Students Can Build Practical AI Skills That Employers Value
