2026-09-16
Hokkaido's Hayate Shimizu argues that hiding AI involvement that would change an editor or reviewer's decision treats their labor as a mere means; honest disclosure is not banned.
The ethics of generative AI in writing has split along two Kantian tracks. Aylsworth and Castro (2024), in the same journal, argued that students have a duty to themselves not to let ChatGPT write certain humanities papers, because that writing is how one cultivates the capacity to set and pursue ends. Sticker (2026) shifted the lens to duties owed to others: Kant's Formula of Humanity forbids treating people as mere means. They have since traded replies. Hayate Shimizu, at Hokkaido University's CHAIN centre, leaves the self-regarding dispute in a footnote and takes the other-regarding route into a setting neither paper treated: journal submission.
What is spent in that setting is evaluative labor. Editors and reviewers decide whether to take a manuscript, how long to spend, and on what terms to judge it. Publishers already have generative-AI policies; Elsevier, Springer Nature, and the ICMJE all appear in the bibliography. Shimizu's claim is that the wrong, if there is one, is not "you violated a house rule." The question is whether the other party got to decide, with the information that would actually move that decision, whether to supply the labor.
The article sits behind Springer's paywall. The reconstruction below uses the abstract, notes, reference list, and the author's own generative-AI declaration. Casework and replies in the body were not available.
Kant's Formula of Humanity says: treat humanity, in yourself and others, always as an end and never merely as a means. His false-promise example makes the structure concrete. You borrow money and promise to repay while intending not to. Had the lender known that, she would likely have refused. The cooperation you got was purchased by blocking the information she needed to decide.
Shimizu maps that structure onto submission. An editor or reviewer supplies evaluative labor. If an author knowingly misrepresents AI involvement in a way that bears on that person's decision whether, and on what terms, to provide the labor, the author uses the participant's rational activity while bypassing her judgment about supplying it. Deception is the route, not the whole offence. That is why the wrong does not collapse into a policy breach: the measure is what would enter the participant's judgment, not what a given journal happens to print. Policies go missing, run too wide, and change. The interpersonal treatment does not track the clause.
Two boundary cases do the remaining work. Anonymous review is also a form of nondisclosure, but it is an openly acknowledged institutional arrangement. Participants already consented to that informational setup. AI assistance itself is not banned by this argument, provided it is honestly represented. What is banned is buying labor the other party would not have offered on those terms, had they been told.
This is a Commentary in Philosophy & Technology volume 39, received 9 August 2026, accepted 11 September, published online 16 September. There are no experiments and no baselines. The normative yield, from the abstract, is three lines.
| Practice | Kantian verdict (from the abstract) |
| Knowingly misstating AI involvement that would change whether or how someone takes the work | Treats the editor or reviewer as a mere means |
| Openly acknowledged nondisclosure, such as anonymous review | Not the same wrong |
| AI assistance that is honestly represented | Not prohibited by this argument |
The author's ethics note is itself an instance of the permitted horn. ChatGPT and Claude were used for language editing, critical feedback on argument structure, finding and checking recent literature, and drafting candidate wording for particular passages, which the author revised before adopting. The argument, the philosophical claims, and the reading of Kant are the author's. Cited sources were checked against the originals.
The bibliography also includes current Elsevier and Springer AI policies, ICMJE authorship rules, Hosseini and colleagues on disclosing AI writing tools, and systems such as The AI Scientist that can already submit papers. The abstract does not walk through those extreme cases. It does show the commentary was written after "a machine handed in the manuscript" stopped being a thought experiment.
For people who actually submit papers, the line is both stricter and narrower than a journal checklist. Stricter: even if a journal has no AI-disclosure clause yet, hiding facts that would change an editor's or reviewer's decision still counts, on this view, as using them as a mere means. Narrower: after honest disclosure, copy-editing, literature checks, and even drafting candidate sentences do not automatically fall under the prohibition.
The practical test is one question. If the editor knew how this text was made, would she still review it on the present terms? If yes, disclose and proceed. If no, hiding that fact is purchasing labor she would not have volunteered.
This is incremental normative cleanup, not a detector and not a percentage threshold. The abstract is enough to get the claim. The cuts between cases remain behind the paywall.
The paywall blocks the body. The abstract never answers the operational question that practitioners need: how much AI involvement "bears on" the other party's judgment? Fixing grammar, rearranging paragraphs, generating the thesis, and fabricating experiments are not the same act. Where the body draws that line is unknown. Korsgaard on lying and O'Neill on consent sit in the references; how they are used is unknown too.
The argument is only as strong as the Kantian frame. Anyone who rejects the Formula of Humanity is not compelled. Most journals already require disclosure, so the practical delta may be small. What the account still adds is a verdict for the gaps: silence in the policy does not clear you, and an over-broad policy does not by itself create the wrong.
There is no evidence here about how reviewers actually behave once AI use is flagged. Metcalf (2026) on AI-produced humanities as misconduct and Hurshman and colleagues on authorship without writing are cited; the abstract gives no ruling on either. Invited commentary, single author, no data. The paper stands or falls on the Kant reading, not on a new fact.