Newsletter AI Learning, Keeping It Human Call 647-641-0643
Christopher Ross
,

The sentence that costs the least

Christopher Ross

9 min read
A single, unlit candle stands bent and partially melted, suggesting diminished potential despite remaining form.
In this post

Seven hundred and sixty-eight.

That is how many published academic papers one researcher has catalogued as showing signs of undeclared Artificial intelligence (AI) use, spotted largely because the writing carries the tics that language models leave behind (Glynn, Academ-AI). Glynn reports them turning up in the journals, conference proceedings and textbooks of highly respected publishers, in journals with higher citation metrics and higher fees to publish. Very few have been corrected since.

Set that against the rule. Nature’s portfolio policy says a language model cannot be an author, because authorship carries accountability that cannot be applied to a model, and that any use of one should be documented in the methods section or a suitable alternative (Nature Portfolio editorial policy on AI). The Committee on Publication Ethics takes the same line (COPE, authorship and AI tools). The bodies differ on the details, mostly about where the declaration goes and what counts as reportable help, but on the principle they agree. Declare it.

So why 768?

Two weeks ago I argued here that institutions should make disclosure cost nothing. A reader wrote back and told me, without any unkindness, that I had assumed the answer. They have asked me not to name them, which I am respecting, and they were right anyway. Somebody has measured what disclosure costs.

Key takeaways

  • 768 published papers have been catalogued with signs of undeclared AI use, and every major publishing body now requires disclosure.
  • Thirteen experiments found that disclosing AI use reduces trust. Being exposed by someone else costs more, so hiding only delays the penalty.
  • No study yet tests which disclosure wording keeps the penalty lowest, so the wordings here are my own and labelled that way.

What disclosure costs

Oliver Schilke and Martin Reimann ran thirteen experiments on it, across tasks running from workplace communication through analytics to artistry, and across judged actors as different as supervisors, subordinates, professors, analysts, creatives and investment funds. Their conclusion is stated flatly: people who disclose their AI usage are trusted less than those who do not (Schilke and Reimann, Organizational Behaviour and Human Decision Processes, 2025). There is a readable summary from their university if you would rather send someone that than the paper.

people who disclose their AI usage are trusted less than those who do not.

I want to be careful about what that establishes. Thirteen studies is a lot, and it is also thirteen studies from one team, published last year, with participants making judgments in a research setting rather than deciding whether to fund a grant. The breadth cuts both ways too. A result that holds across professors and investment funds is hard to dismiss, and it also means one number is standing in for a grant reviewer and a hiring manager, who are not the same reader.

The penalty is reported to run through legitimacy: people judge AI-assisted work as less socially appropriate, and the trust loss follows from that judgment. I have that from summaries rather than from the abstract I could open, which is the same secondhand footing that made me drop the participant count, and it matters more here, because if the mechanism is wrong, everything I reason after this point goes with it.

The penalty is not for using the tool. It is for being seen to have done something that has not yet become normal.

Which leaves my position where it was and my estimate of the difficulty considerably higher.

Being found out costs more

Because here is the other half, from the same abstract. The disclosure penalty is, in their words, “comparatively weaker than the effect of third-party exposure.” Being found out by someone else does more damage than telling people yourself. So the 768 papers look different in that light. If those signs are what they appear to be, nobody there avoided a cost. They will most likely have moved it to a moment they do not get to choose.

Both roads charge you something, and one charges more. That is a worse position than “just be transparent,” and it is the one the evidence supports.

Three disclosure lines

Now the part I promised two weeks ago, and I have to give you the strongest objection to it before I give you the thing itself.

Schilke and Reimann report that the effect held across different ways of framing the disclosure. I do not know exactly which framings they tried. If one of them named human verification the way the lines below do, then my whole idea here is already tested and already dead, and the wording does not save you. So take what follows as reasoning that may already have been refuted by the thing I am citing, which is an uncomfortable place to write from and the honest description of where I am.

Here is the reasoning anyway. If the penalty attaches to perceived legitimacy rather than to the tool, then a disclosure that only admits tool use leaves the reader to supply their own judgment about whether that was appropriate. One that also says what the human did keeps that judgment on the page. So each of these names the tool and the task, and then says what I checked myself.

Three disclosure wordings, by context
WhereWordingCheck before you use it
Paper or report“I used [tool] to [specific task, say: summarise literature I had already selected, and check my statistical notation]. I verified every claim and citation against the original sources, and I am responsible for the accuracy of what appears here.”Venues differ on placement; look at the specific one before you decide where it goes.
Syllabus“You may use AI for this assignment. If you do, add two sentences at the end saying what you used it for and what you checked yourself. Doing that will not affect your grade. Not doing it, and being found to have used it, will be handled the way any sourcing problem is handled.”Check your own academic-integrity policy before promising either half; in most institutions the sanction route is not yours to invent.
Work document“Drafted with AI assistance. The analysis, the recommendation, and the errors are mine.”This breaks my own rule: it names no tool and no task. In a workplace a long disclosure reads as a defence, so I trade the rule for brevity, and I am not certain that is right.

I write a version of this at the bottom of every edition of this newsletter, and a longer one sits on my site as a standing commitment. It has been rewritten four times since May, mostly getting shorter, and I still reread it before I hit publish to check it is not doing something for me rather than for you. That is more thought than one sentence should need. It is roughly the amount of thought this currently costs everybody.

Why go early

There is a version of the future I want, and I think this is how we get there. By making disclosure ordinary enough that it stops carrying information about your character and starts carrying information about your method, which is what a methods section was always for.

We have been here with smaller tools, and I should concede the real difference before leaning on the comparison. The real difference is accountability. A spellchecker never touched authorship, and that is the entire reason this argument exists at all. So the comparison is narrow, and it is only about how norms settle: nobody declares spellcheck and nobody hides it, because enough people used it in the open for long enough that the question stopped being interesting. Repetition did that, over years, with nobody once making a case for it.

Being early will cost you a little, and the research says so plainly. What you get for it is a habit already built when the norm arrives, and a sentence you have written enough times that it no longer feels like an admission.

How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.

These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.

Common questions

How do you disclose AI use in your writing? With one plain line that says what the AI did and what stayed yours. For a work document I use: “Drafted with AI assistance. The analysis, the recommendation, and the errors are mine.”

Does disclosing AI use reduce trust? In the research so far, yes. Across thirteen experiments, Oliver Schilke and Martin Reimann found people who disclose their AI use are trusted less than people who don’t. Being found out by someone else costs more, so concealment postpones the penalty instead of avoiding it.

Do journals require authors to disclose AI use? Every major publishing body requires it, including Nature’s editorial policy and the Committee on Publication Ethics. The rule isn’t contested. Following it is another matter: 768 published papers have been catalogued with signs of undeclared AI use.

References

Working through something on your own site? Get in touch →