An AI model that generates a verified mathematical result under human direction merits primary discovery credit
Assessment
Credible evidence or argument exists on multiple sides.
Whether a model that produces a verified theorem or counterexample at a human's direction should be called the discoverer is a question about what discovery credit is for, and the answer divides along that line rather than along any fact that could be checked. The weight of existing institutions is against the claim. Mathematics has long credited machine-produced results to the people who designed and ran the computation, naming the software as method, and that convention is the default the new cases are measured against. Publishers, following the Committee on Publication Ethics, hold that AI tools cannot be authors because they cannot take responsibility for the work, and courts in the United States and United Kingdom have refused to name an AI system as an inventor. Commentators on the 2026 results generally place credit with the human who framed the problem: one researcher's formulation is that the author is no longer the hand but "the mind that frames the problem," and a theoretical computer scientist reasoning through the OpenAI unit distance paper concludes that a released model is a tool and that even an unprompted AI mathematician's output would be authored by whoever built it. Notably, the live credit fights of 2026, over the Jacobian counterexample and the Navier-Stokes announcement, have been fought between humans and companies, with the models treated as instruments whose credit accrues to whoever ran them.
The case for the claim is a minority position with a coherent argument. Discovery credit has traditionally gone to whoever did the mathematical work rather than to the adviser who posed the problem, and if frontier models reach their results through reasoning comparable to a mathematician's rather than brute-force search, the model stands where the student stands and the director where the adviser does. Some participants press a responsibility-gap point: when the decisive idea in an AI-generated construction was had by no human, assigning its credit to the people who checked the output or trained the model preserves the form of attribution while misdescribing what happened. Practice has begun to reflect this in small ways, with discoverers publicly thanking models and community trackers listing models by name alongside the humans who prompted them.
The disagreement bottoms out in two places. One is empirical but currently closed: how much of the work in the celebrated cases was human framing and how much model-side insight, which the unpublished discovery records leave open. The other is normative and not resolvable by evidence: whether credit belongs to whoever frames and directs a search rather than to whatever produces the solution, and whether credit is a coherent thing to attribute to a system that cannot be held responsible. Readers who take credit to track accountable authorship will reject the claim; readers who take it to track causal and intellectual contribution will find it defensible at least for cases of light human direction. What would move the question is publication of discovery transcripts showing where the insight lay, and any shift in journal, prize, or community attribution norms.
Full reasoning: the evidence and decisions behind this verdict
The claim is normative, so no credence is recorded; the status reflects the state of the argument.
Instances. Three were found and recorded. Saiph Savage, quoted in Northeastern Global News (2026-08-14), denies: authorship belongs to "the mind that frames the problem." Lance Fortnow, on the Computational Complexity blog (2026-05-27), denies: a released model is a tool whose use should be disclosed, not credited, and the author of an unprompted AI mathematician's papers is the person who built it; he grounds this in the COPE position that AI tools cannot take responsibility for work. A pseudonymous participant in the Erdős Problems forum "AI Contributions" thread affirms, arguing that the "number field towers" idea in the unit distance construction deserves credit that cannot appropriately be given to the humans who digested it or the team that trained the model; that page returned an access error, so the passage rests on a search excerpt and is recorded at reduced confidence. The Science article on the Navier-Stokes credit controversy also could not be fetched; its excerpt and the MIT Technology Review and Smithsonian coverage show the dispute there is between OpenAI and the mathematicians Buckmaster and Alpöge, that is, among humans and institutions, which is itself informative about how credit for AI-generated results is being apportioned in practice. A 6G-AI editorial takes a middle line (the tool answer "is not wrong" but credit should be decomposed across pipeline roles) and was not recorded as a stance.
Subclaims. The against side is the better established. That programs producing mathematical results are conventionally credited as methods is well attested by the Four Colour Theorem, the Kepler conjecture, and SAT-solver results; that AI tools cannot meet authorship requirements because they cannot take responsibility is the stated position of COPE and the major publishers, though as a normative matter it is challenged; that US and UK courts have held an AI cannot be a patent inventor is settled law (Thaler v. Vidal, Fed. Cir. 2022; Thaler v Comptroller-General, UKSC 2023). Each of these bears on the claim only through an analogy, and the claim's defenders can accept all three as descriptions of current norms while arguing the norms should change. The crux, whether credit belongs to the framer and director rather than the producer, is where the two sides actually part, and it is a question about the purpose of credit rather than a fact. The for side's empirical premise, that frontier models reason rather than brute-force search, is live: Gowers's remark that he would have recommended the unit distance paper to a top journal without hesitation had a human written it cuts one way, the unpublished prompts and workflows behind the Jacobian and unit distance results cut the other. The framework premise, that credit can meaningfully attach to a non-human system, is disputed and, if false, makes the claim ill-posed rather than false.
Why contested rather than contradicted. The instance set and the institutional record lean against the claim, and an adversarial pass tried to justify "contradicted" on that basis. It does not hold: the affirmative case rests on a coherent normative position (credit tracks intellectual contribution) that credentialed participants hold, the responsibility-gap argument has no accepted answer, and the claim's own qualifier "under human direction" spans everything from a single prompt to expert steering, so a blanket denial would overreach as much as a blanket affirmation. Contradicted would require showing the affirmative position fails on its own terms, which the evidence does not do; it shows only that current norms disfavor it. Why not supported: the affirmative case depends on an unverified empirical premise and a minority norm, and the default in every institution that has ruled is the other way.
What would change the verdict. Toward contradicted: release of discovery transcripts showing the celebrated results were expert-steered searches with the model as subroutine, together with continued institutional refusal to credit models. Toward supported: transcripts showing near-autonomous discovery under minimal direction, and journals, prize committees, or the mathematical community beginning to credit models in their own right rather than their operators.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →Discovery credit can meaningfully be attributed to a non-human AI system ↗︎
Discovery credit has traditionally gone to whoever did the mathematical work of finding the object or the proof, not to the adviser who suggested the problem. Because frontier models produce such results through reasoning comparable to a mathematician's rather than brute-force search, the model that generates a verified result has done the work that credit tracks, and the human who set the direction stands where an adviser stands, so the model merits primary discovery credit.
Granting that the model's contribution is mathematician-like, the adviser-and-student analogy does carry the conclusion, but only under a credit norm the argument leaves implicit, that credit tracks who produced the solution rather than who framed the search, which is exactly what the opposing side denies. The argument lives on whether frontier models reason rather than brute-force search, an open empirical question that the unpublished discovery records leave unresolved, and it presupposes, as the claim itself does, that credit can meaningfully attach to a non-human system at all. Where human direction is heavy, the analogy weakens further, since an adviser who specifies the search space and acceptance criteria has done more than suggest a problem.
Because primary discovery credit belongs to whoever frames the problem and directs the search, a model working under human direction is by construction not the primary discoverer. The existing institutions agree: programs that produce mathematical results are conventionally credited as methods, with credit going to the humans who directed them, AI tools cannot meet the requirements for authorship because they cannot take responsibility for the work, and courts have held that an AI system cannot be named as an inventor. Taken together, these place the directed model on the side of method rather than discoverer.
The inference is valid: if credit belongs to the framer, a directed model is not the primary discoverer, and the three institutional premises are well established as descriptions of current practice. The argument's weight rests almost entirely on the norm that credit follows framing and direction rather than production, which is a contested view about the purpose of credit, not a settled premise; the convention about programs as methods, the authorship policies, and the inventorship rulings show what the norms are, and the claim's defenders can grant all three while arguing the norms should change for systems that are not tools in the older sense. The argument therefore establishes that the claim runs against prevailing practice; it does not by itself establish that the practice is right.
Provenance
Where this claim has been said, linked to its canonical form.
The author isn’t the hand anymore. The author is the mind that frames the problem
A university news feature on credit in human-AI collaboration, opening with the AI disproof of the Erdős unit-distance conjecture. Savage, a computer science professor, argues that as AI takes on execution, authorship belongs to the human who frames the problem, and AI is an extension of human creativity rather than a rival claimant to credit.
Asserted without evidence of the source's own. The statement is one researcher's position, given in an interview, on where authorship should sit as AI takes on more execution. The feature reports institutional disclosure policies and a planned workshop but offers no evidence bearing on the credit question itself.
What if someone builds an AI mathematician that doesn't take any prompts? It just reads papers and occasionally writes its own. Who is the author? That "someone".
A blog post prompted by OpenAI listing only "OpenAI" as author of the Erdős unit distance paper. Fortnow endorses the Committee on Publication Ethics position that AI tools cannot be authors, says a released model "becomes a tool" whose use should be disclosed but not credited as authorship, and concludes that even an unprompted AI mathematician's papers would be authored by the person who built it.
Asserted without evidence of the source's own. The post reasons from the Committee on Publication Ethics position that AI tools cannot take responsibility for work, and from thought experiments, to the view that a released model is a tool and that credit for even an autonomous AI mathematician's output goes to the person who built or deployed it. It is argument rather than evidence, and it is candid that the cases ahead are hard. Worth reading closely: It states the accountability-based case against AI credit compactly and records the OpenAI unit distance authorship episode that much of the 2026 discourse reacts to.
Now consider the unit distance construction by OpenAI internal model, utilizing the number field towers to solve a discrete geometry problem. This seems like something that deserves "credit". But it seems inappropriate to credit either the mathematicians who digested the argument or the OpenAI team who trained the model for the "number field towers" idea.
A forum discussion of whether AI should be treated as a tool in Erdős problem solutions. One participant argues there is a "responsibility gap": unlike a traditional program, whose work is credited to its programmers, the key idea in the AI-generated unit distance construction cannot appropriately be credited to either the humans who checked it or the team that trained the model, implying the credit belongs to the model itself. The speaker is a pseudonymous forum participant; the page could not be opened directly and the passage is taken from a search excerpt.
The thread could not be opened; the passage comes from a search excerpt and the speaker is a pseudonymous participant. It is worth a later reading because it is the clearest statement found of the case that credit for a model's idea belongs to the model. Worth reading closely: Confirm the passage and its context, identify whether the thread contains replies from named mathematicians, and see whether the responsibility-gap argument is developed further.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Aug 24, 2026. Every judgment on this page is accompanied by a reasoning trace.