Who Gets Credit When the Co-Author Is a Model?

Scholarship runs on an honor system older than any policy on your campus. AI just moved into the middle of it, and most institutions have not decided what they think.

A faculty member at a teaching-focused college, the kind where scholarship happens on evenings and summers between four-course semesters, finishes an article she has been chipping at for two years. AI helped: it tightened her literature review, caught a methodological inconsistency, and rewrote her conclusion twice before she rewrote it back. She is now staring at the journal’s new disclosure checkbox, genuinely unsure what the honest answer is. Did AI “contribute to the manuscript”? What does that even mean, when the ideas are hers, the judgment is hers, and the sentences are somewhere in between?

Her institution has an AI policy. It covers students. On the question she is actually facing, it is silent, which means she is deciding institutional research policy alone, at her kitchen table, at 11 p.m.

The institutional question

What are your institution’s positions on AI in research and scholarship, and does anyone doing the scholarship know them?

Positions, plural, because this is not one question. What must be disclosed, and to whom? What counts as authorship? What can AI touch in a thesis? How does AI-assisted work count for tenure and promotion? These questions land at every institution that expects faculty to publish anything at all, which is to say nearly every institution. This is not an R1 topic that teaching institutions can skip. If anything, the fewer resources an institution has for research support, the more its faculty are already using AI to stay in the scholarly game, and the more it owes them clarity.

What this looks like in practice

Start with what AI actually does in scholarship now, because the honest inventory is broader than most policies imagine. It searches and summarizes literature. It drafts and revises prose. It writes analysis code, cleans data, and suggests interpretations. It critiques arguments and finds holes. Some of that is unambiguously assistance, the digital descendant of a good librarian and a sharp colleague. Some of it edges toward contribution. The line is blurry, the tools keep moving it, and pretending the line is obvious is how institutions end up with faculty guessing.

Disclosure is where the outside world is forcing the issue. Journals, conferences, and funders increasingly require authors to state how AI was used, and the requirements differ from one venue to the next, sometimes sharply. An institution cannot control those rules, but it can do two things: keep its people informed about the terrain, and set a sensible internal default. The default worth considering is simple: when in doubt, disclose, and treat disclosure as evidence of integrity rather than a confession of weakness. A culture where honesty about methods feels dangerous will get less honesty, not less AI.

Authorship is the cleaner question, and it has a broadly emerging answer worth stating plainly: the model is not an author. Authorship carries accountability, the ability to stand behind the work, respond to critique, and own the errors, and a model can do none of that. The person who directed the tool, verified its output, and takes responsibility for the result is the author. That framing does useful work: it keeps the focus on accountability rather than on percentages of assistance, which nobody can measure anyway.

Student research deserves its own paragraph, because the thesis, the capstone, and the senior project are where teaching institutions actually live this domain. The same logic from the assessment issue applies: decide what the learning goal is, then decide what AI may touch. If the capstone exists to prove a student can conduct and defend an independent inquiry, the defense, the oral component, the documented process, are where that proof now lives, not in the polish of the final document. Advisors need shared expectations, students need them in writing, and both need them before the project starts, not during the misconduct hearing after.

Research integrity gains a new failure mode worth naming: fabrication by fluency. Models generate plausible citations to papers that do not exist, confident summaries of studies they misread, and statistical interpretations that sound right and are not. A researcher who trusts without verifying is not committing old-fashioned fraud, but the published record ends up polluted just the same. The verification burden is real, it belongs to the human, and institutions should say so in exactly those words.

Then tenure and promotion, the place where institutional values become career arithmetic. Evaluation committees are already receiving files containing AI-assisted work, mostly without any agreed way to think about it. The risk runs in both directions: a committee that quietly discounts AI-assisted scholarship punishes transparency and teaches the next candidate to hide, while a committee with no standards at all invites volume over substance. The fix is not complicated, but it is deliberate: updated written guidance, so that a candidate’s honesty about methods is never the thing that costs them their career.

For resource-constrained institutions, a closing reframe: this domain is also an opportunity, not just a compliance chore. AI meaningfully lowers the cost of staying scholarly active for faculty who have no grad assistants, no course releases, and no research office. An institution that provides clear rules and decent tools is, in effect, expanding its research capacity without hiring anyone. The schools that treat this domain only as a risk to manage will miss the part where it quietly levels a playing field that has been tilted for decades.

The Atlas connection

In Atlas, our AI operating map, this is the research domain, in the Mission group — Educate — and it completes a triangle with teaching and learning and with assessment and integrity: the same question, what does honest intellectual work look like with AI inside it, asked about faculty instead of students. An institution that holds students to disclosure standards its own faculty do not follow will be teaching a lesson about hypocrisy, not integrity. The domain also leans on policy and acceptable use, over in Foundation, for where the rules live, and on AI fluency and professional development, in Equip, for how anyone learns them. Next issue, this part of the series closes with the technology that stresses every one of these seams at once: AI that acts on its own, which Atlas treats not as a domain but as a lens on the whole map.

Questions worth putting on the agenda

Best owned jointly by the provost’s office and faculty governance, since legitimacy here cannot be imposed.

• Does our institution have a stated position on AI use in faculty scholarship, or only a student policy wearing a general title?
• What is our disclosure default, and do faculty know the major journal and funder rules in their fields are diverging?
• For theses and capstones: have advisors agreed on what AI may touch, and do students get that in writing before they start?
• Who carries the verification burden for AI-assisted work, and have we said so explicitly?
• Have our tenure and promotion guidelines been updated so that honesty about AI methods is never career-fatal?

The bottom line

Scholarship has always been a chain of accountability: someone stands behind every claim. AI does not break that chain unless institutions let people pretend it is not there. The faculty member at her kitchen table does not need surveillance. She needs a policy that was written with her in mind, so the honest answer to the checkbox is something she can know rather than guess.

The model can draft, critique, and polish. It cannot stand behind the work. That remains the human’s job, and the institution’s job is to say so clearly.