Radar Perene / Archive / science

AI in scientific research: what journals already regulate

◦ Index methodology v2.2 (working papers with DOI). See the methodology.

Science

In a few years, scientific publishing went from "not our problem" to a surprisingly convergent body of rules on artificial intelligence. The speed is striking because the system that produced those rules is the same one that takes years to evaluate an article. And their content matters to any reader of research — including market research — because it defines what, from here on, a signature under a study means. This article describes the state of the rules; it speaks about the subject and is not a usage manual for any tool.

Journal regulation of generative AI converges, today, on three points: an AI cannot be the author of a scientific article; relevant use of AI in preparing the work must be disclosed; and full responsibility for the content — including for what the tool got wrong — remains with the human authors.

The convergence was coordinated by no central authority, which makes it more notable. Competing publishers, ethics committees and associations arrived, in short order, at nearly identical formulations — a sign that the three points touch something structural in the definition of science, not a preference of style.

What the house's survey found

While preparing this trail, the house compared the public AI policies of a set of large international scientific publishers, of publication-ethics committees and of economics and finance journals — the policies are open documents, published on instructions-to-authors pages. Three layers stand out in what is required.

The first layer is the authorship ban, by now practically universal. The declared foundation is always the same: scientific authorship implies accountability — being able to answer for the work, approve the final version, be held to account for error or fraud — and a computational system does not answer, does not approve and cannot be held to account. The question yields more than it seems, and it is the subject of the next article on this trail.

The second layer is disclosure, where the real variation between policies lives. The consensus covers the center: AI use that affects content — drafting passages, generating or analyzing data, producing code relevant to the results — must be declared, typically in the methods section or in a dedicated note. The edges diverge: some policies explicitly exempt grammar and style revision, treating it as equivalent to the checkers word processors already embedded; others ask for declaration of any use. AI-generated images receive the harshest treatment — near-general prohibition in scientific figures, with narrow exceptions for work about the technology itself.

The third layer is the least known and perhaps the most revealing: the rules for referees. Several policies forbid a reviewer from submitting someone else's manuscript to AI tools, for a reason prior to the quality of the report — confidentiality. A manuscript under evaluation is a privileged document, and feeding it into a third-party tool is sharing it with parties who should not read it. The rule is a reminder that the system's integrity does not rest on authors alone.

Why the ruler is responsibility, not technology

Read together, the policies reveal a common criterion that is rarely spelled out: none of them regulates the technology itself — they regulate the traceability of responsibility. The question implicit in every clause is always "who answers for this sentence, this datum, this figure?". It is the same principle that organizes the rest of publication ethics, from plagiarism to data fabrication; AI did not create a new problem in kind, it created a new scale for an old problem — plausible text with no clear owner.

For the reader of market research, the practical consequence is twofold. On one side, a study published under these rules carries an implicit declaration: the authors affirm they verified what they sign, whatever tool was used along the way. On the other, the literature circulating outside journals — reports, commercial material, posts — is subject to none of these rules, and the asymmetry between the two worlds has widened: peer-reviewed text and text loose on the internet have never stood further apart in guarantees of provenance.

Let the limit of the rules be recorded honestly: disclosure depends on who declares, and enforcement is imperfect — a matter that flows into the question of detectors, also treated in the next article. Publication rules do not prevent bad conduct; they make it costlier and give grounds for sanction when discovered. That is what integrity rules have always done.

Frequently asked questions

Do journals forbid the use of AI in research?

In general, no — they forbid AI authorship and require declaration of uses that affect content. The distinction between banning the tool and demanding responsibility for the result is the axis of practically every current policy.

Where is AI use declared in an article?

It depends on the journal: methods section, dedicated note or submission letter. Each journal's instructions to authors — public documents — specify the format.

Does fixing grammar with AI require declaration?

Policies diverge exactly there. Some exempt language editing; others ask for declaration of any use. Prudent authors treat the destination journal's policy as the ruler, not general custom.

Do these rules apply to working papers?

Repositories tend to demand less formally, but the logic of responsibility is the same — and journal rules reach the work when it proceeds to review. Integrity practices are born better at the working paper stage than grafted on afterwards.

---

The authorship ban looks like the settled point of the policies — until one asks what exactly defines an author, and how anyone would verify a violation. That is the subject that closes this trail: AI authorship and detection in science.

House reading: today's reading is in the Diário.