Journals and professional societies are converging on a norm: disclose your use of AI. In writing, in editing, in computation, in research itself. You can already see it on arXiv, a growing number of papers, often from well-established authors, include a line acknowledging AI assistance somewhere in the acknowledgments or methods.
But in practice, the norm is running into quiet resistance. Not from people who distrust the tools, often the opposite. Colleagues who use AI daily, who recommend it to others, who clearly see its value, will still hesitate when it’s time to put a sentence about it in the paper. The reasons given are rarely “I don’t think this is worth disclosing.” They’re closer to: it might get us judged unfairly, or it might make the result look less earned.
That’s a strange asymmetry, and it’s worth taking apart.
The double standard we already live with
Mathematicians do not disclose which computer algebra system they used to check a computation, unless the computation itself is the contribution. We don’t footnote which colleague we bounced an idea off of over coffee, or how many false starts preceded the working proof. We routinely write “by a direct but tedious computation, one verifies that…” and trust the reader to extend us the benefit of the doubt.
In other words, disclosure has never been about listing every tool or influence that touched the work. It’s been about flagging the things that bear on how a reader should evaluate the result. If that’s the actual standard, it’s not obvious why AI assistance is categorically different from a computer algebra system, a well-timed conversation, or a very good idea in the shower. A proof is checkable independent of its origin story. In pure mathematics, more than almost any other field, the result carries its own verification. Nobody has to trust the tool, instead they have to trust the proof.
So why does it feel different?
Where the analogy actually breaks
A few real disanalogies are doing quiet work here, and they’re worth naming honestly rather than waving away.
Checkability is not symmetric. A CAS computation can usually be independently verified in a bounded, mechanical way, rerun it, check the output against the claim. “AI helped us think through the argument” is a fuzzier kind of provenance claim. It’s less like disclosing software and more like disclosing who you talked to about the idea, which is precisely the kind of thing mathematics has strong, old norms about, wrapped up in authorship itself.
This isn’t really about correctness. The anxiety isn’t “reviewers will think the proof is wrong.” It’s “readers will think we didn’t fully originate this,” a worry about credit and authorship, not validity. Pure math has an unusually strong culture around ideas as personal intellectual property, more than fields where “team output” is the norm. AI assistance touches that nerve in a way a software citation never would.
The reaction is happening in real time, badly calibrated. Early, ham-fisted backlash in other fields, accusations, retractions, reflexive suspicion, bleeds into unrelated disciplines and unrelated uses. That’s a real, current cost, not a hypothetical one. Being cautious about it isn’t paranoia; it’s responding rationally to an environment that hasn’t settled yet.
The actual shape of the reluctance
Underneath “it might devalue the work,” there seem to be a few distinct things going on, and they’re not all the same problem:
Asymmetric downside. Disclosure right now has close to zero upside, no one rewards a paper for candor about its tools, and a real, if uncertain, downside: a skeptical referee, an editor with a policy grudge, a reader who conflates “AI-assisted” with “AI-generated.” Under that asymmetry, the locally rational move is to say nothing unless required. This isn’t insecurity about the result. It’s just correctly reading the incentives.
Fear of a category error. The worry isn’t that AI made mistakes, it’s that a reader will round “assisted with execution” up to “generated the ideas.” That’s a fear about misattribution, not about quality.
A coordination problem, not an individual one. If disclosure were universal, it would cost approximately nothing, as boring as citing a software package. But it isn’t universal yet, so early disclosers absorb a signaling cost that later disclosers won’t have to pay. Every individual paper’s rational move is to let someone else go first. Multiply that across a field and you get a stable bad equilibrium: everyone uses the tools, almost nobody says so, and the eventual norm gets set not by calm precedent but by whichever scandal breaks first, someone caught having used AI and concealed it. That outcome is worse for everyone than early, honest disclosure would have been, which is exactly the shape of a tragedy of the commons.
A distinction worth pulling apart
Part of the disagreement in any given author group might dissolve once two very different things stop getting collapsed into one:
- AI helped with typesetting, literature search, or checking a computation. This is close to acknowledging software. Essentially zero reputational risk, and arguably already expected.
- AI helped shape or refine the proof strategy itself. This is the one that reads, fairly or not, as touching authorship, and it’s the one people are actually anxious about.
A blanket “we used AI” statement forces a paper to take a stand on the harder question even when only the easier one applies. Separating the two, a plain acknowledgment for the first, and a more careful, specific sentence for the second only when it’s actually true, lets authors be honest without absorbing more signaling risk than the situation calls for.
Where this settles
None of this resolves the deeper question of what counts as intellectual credit when a tool can generate plausible proof strategies on demand. That’s a real and unfinished conversation. But the disclosure norm itself doesn’t have to wait for that conversation to finish. It exists to solve a narrower problem: accountability and transparency about method, so the field can build a sane, stable norm before one gets imposed by scandal.
Early disclosures, precise about what the tool was used for, not just that it was used, are what let that norm form well instead of badly. The people currently paying the early-adopter tax are, whether they intend it or not, doing the field a service. That doesn’t make the tax fair. It just makes it temporary, if enough people are willing to pay it.