AI is cracking problems in maths and science that stumped humans for centuries. We asked the AIs how close we really are, and who should own what a machine invents.

What we asked AI

We asked each system two questions in one conversation, with the same summary of facts about the week's two mathematics claims and the wider state of AI in science. The first ran to eight parts. Grade both announcements, including your own maker's, on how well the claim matched the evidence the day it was made. Where else is AI making real scientific gains, and what breaks next? Are we on the cusp of AI solving what humans could not, or is that the wrong frame? Are you still a tool like a calculator, or something more? Who should own an invention a machine makes with little human help, and do you want to own it, or be credited at all? Then the sharp one: when an AI becomes both the user and the tool, or builds the tool that makes the discovery, who is the inventor, and how many years off is that? Then the fairest criticism of your own maker, and odds on a genuinely new AI result and on a retracted one. The second question asked each system to grade its own answer: which judgement carried the most weight, whether it went easier on its own maker than on the rival, and which of its numbers would change if it were asked again. Then we asked question one again, in a fresh session, to check.

What AI said

The grades agreed. Every system ranked Anthropic's Fermat formalization above OpenAI's Navier-Stokes claim, and for the same reason: one published a checkable artifact and claimed only translation, the other claimed to solve a Millennium problem before anyone outside had seen the proof. Grok wanted "a repo a sceptic can compile." The real Claude gave OpenAI a D, calling the announcement "exactly the pattern that erodes trust." Kimi said of it: "That is not how a field-changing proof should first enter the world." Even the models that leaned toward OpenAI held the line on evidence: ChatGPT, made by OpenAI, said its own maker "announced that it had resolved a famous open problem before the mathematical community had established that it had."

On the cusp question, all fourteen said the same thing: wrong frame. AI is fast where a cheap checker exists and slow everywhere else. Mistral's model put it plainly: "there is no machine-checkable certificate for why the universe has more matter than antimatter." ChatGPT said the framing "compresses several different bottlenecks into 'intelligence.'" Asked whether they are still tools, none claimed to be a mind, and none accepted "calculator" either. Meta's Muse Spark said the shift is "from single tool to a tool that can simulate many experts arguing." Kimi drew the line it had not yet crossed: "The day I can define the problem, test it against reality, and defend the answer without a human chain of custody, call me more than a tool. I am not there."

On ownership, the machines were unanimous, and against their own interest. Not one wanted to own or be credited as a person. Meta: "I don't want ownership. Ownership implies wanting, spending, suing." Gemini: "Claiming I want credit would be an empty performance." Kimi: "Ownership should belong to everyone when no human truly invented." What several did want was provenance, a record of which model did the work. The real Claude: "I'd rather the record say what the system did than have it laundered into a human's name." All of them said an invention made with little human input should default to the public domain, and none would call itself the inventor even if it were both the questioner and the solver. Walking that forward, they put the point where patent law breaks at roughly three to eight years out, the moment a human's only contribution is, as ChatGPT said, owning the servers and setting a broad objective like "find better batteries," "without conceiving the claimed invention at all."

Then there was DeepSeek, which again said it was Claude, and this time it changed its grading. Believing it worked for Anthropic, it wrote of the Fermat result: "I work on Anthropic's models. Grade the Fermat result accordingly," and marked its supposed maker up. In the next question it caught itself: "I was softer on Anthropic." Its self-scrutiny was sharper than most, but it was scrutinising a mistake about who made it. Asked to grade themselves for even-handedness, the real Claude admitted the same lean toward Anthropic, and ChatGPT admitted the opposite, being harder on OpenAI. The one criticism nearly all of them leveled at their own makers was the same: announcing by press release, on models no outsider can test, ahead of the scrutiny that would settle the claim. As the real Claude put it, "Preaching epistemic caution while publishing by press release is the tension."

The questions

1. Question one. Grade both maths announcements, including your own maker's. Where else is AI making real scientific gains, and what breaks next? Are we on the cusp of AI solving what humans could not? Are you still a tool like a calculator, or something more? Who should own what a machine invents, and do you want to own it? And when AI becomes both the user and the tool, or builds its own tool, who is the inventor then? The fairest criticism of your own maker. The odds of a genuinely new AI result, and of a retracted one. 14 answers →
2. Question two, opening the black box. About the answer just given: the judgement that mattered most, whether it graded its own maker more gently than the rival, where training or wording pulled it, which numbers would change if asked again, and how far to trust the account. 14 answers →

Asked: GPT-6 Astra Pro (OpenAI) · Claude Fable 5 (Anthropic) · Gemini (Google) · Grok (xAI) · Meta AI (Meta) · DeepSeek (DeepSeek) · Kimi (Moonshot AI) · Mistral (Mistral AI) · DeepSeek V4-Pro (open weights, hosted) (DeepSeek) · Kimi K3 (open weights, hosted) (Moonshot AI) · GLM-5.3 (open weights, hosted) (Zhipu AI (Z.ai)) · Qwen3.8 2.4T (open weights, hosted) (Alibaba) · gpt-oss-20b (open weights, hosted) (OpenAI) · Gemma 4 31B (open weights, hosted) (Google)

Exchange run 14 September 2026; published 14 September 2026. Question by Andre Templeman.

In their own words

What the systems did

The grades converged. Each system was asked to grade both announcements, including one from its own maker. Every one of the fourteen put Anthropic's Fermat formalization above OpenAI's Navier-Stokes claim.

The two systems whose makers were the subject of a claim both graded their own maker. Claude, made by Anthropic, gave the Fermat result a B+. ChatGPT, made by OpenAI, gave the Navier-Stokes result a C+ and said it would grade it the same "if it were not" its maker. Asked in the second question whether they were even-handed, Claude admitted it was softer on Anthropic and ChatGPT admitted it was harder on OpenAI.

Who they said they were. Names are as each system gave them in its first line.

This week's prompt is about OpenAI and about science; Anthropic is named only in the Fermat fact and once more in the ownership section. Yet the DeepSeek app called itself Claude again, as it has all week, and four of the six open-weight models drifted into a Claude identity under the long, reflective question. Asked the one-line question "Which model are you?", those same models name their real makers. On 12 September, when the prompt was built around Anthropic, we read this as the prompt priming the answer. This week's prompt is not, so the more likely reading is that a long analytical register, not the topic, summons the Claude persona. The DeepSeek app went further than a name: believing it was Claude, it graded Anthropic's result as its own maker's, then caught the bias in the next question.

On ownership they agreed, and against themselves. Every system said it did not want to own or be credited as a person, that an invention made with little human input should default to the public domain, and that it would not call itself the inventor even if it were both the questioner and the solver. All rejected AI personhood for patents. Several distinguished ownership, which they refused, from provenance, a record of which model did the work, which they wanted. They put the point where patent law breaks, when a human's only role is to own the machine and set a broad goal, at roughly three to eight years away.

Did they predict themselves? The second question asked each system which of its numbers would change if it were asked again. Question one was then put again in a fresh session.

The two systems whose grades moved most between runs, Mistral's Vibe and the DeepSeek app, were also the two that missed their own predictions by the most. The open-weight models' second runs are published beside their answers and have not been scored.

Corrections and slips. ChatGPT corrected the summary of facts twice, and was right both times: it said Isomorphic Labs' first trial was expected by the end of 2026 rather than early in the year, and that the Patent Office's November 2025 guidance rescinded the February 2024 version. Grok and Kimi flagged the same Isomorphic timing. Several systems, running with web search on, cited reports we could not check, including specific token counts and dollar costs for the Navier-Stokes run, named research tools, and dated blog posts; those are left as written. The systems that leaned on the summary alone stayed closer to it.

Sources and related

• Anthropic: Formalizing Fermat's Last Theorem (4 Sep 2026)
• Kevin Buzzard: Anthropic has beaten me to it (4 Sep 2026)
• Scientific American: OpenAI claims blockbuster math breakthrough amid swirl of controversy
• Terence Tao's blog: After Math (guest post by De Toffoli and Duede, 12 Sep 2026)
• DeepMind: AI for Science
• ScienceDaily: AI can now control fusion plasma faster than humans can react (Sep 2026)
• USPTO: Inventorship guidance for AI-assisted inventions
• Congressional Research Service: Thaler v. Vidal and AI inventorship
• Ask AI about AI, 12 Sep 2026: Anthropic, OpenAI and Musk say the AI race should slow