Thinking in Public

Between Spell check and fabrication: what AI does to authorship

The Scientific Advice Mechanism (SAM) of the European Union has published guidelines on its own use of AI.  This is relevant because it shows how an advisory body proposes to apply to itself the kind of standards it is recommending to others. This is rarer than it should be.

The guidelines cover anyone involved in SAM processes, from the Chief Scientific Advisors, the secretariat, to the experts participating in working groups or acting as peer reviewers. They map the actual tasks, from literature searching and screening, peer review and policy mapping, to drafting, translation and expert identification. Anchoring the principles in actual tasks rather than hovering over it.

The seven principles are personal responsibility and accountability, human-centred use of AI, review and human control, transparency and reporting, data handling, intellectual property compliance, and environmental footprint. All relevant issues, but here I want to focus on a particular aspect that these guidelines make clear, and which is also very relevant for authorship: "AI-assisted" is not one single category.

The current debate about AI and authorship is focusing mostly on the technology. Journals ask whether AI was used; funders ask for disclosure; universities write policies about AI-generated content. All assuming that AI use is a act that can be declared or not, prohibited or not. It is not that simple: AI is used to correct spelling, to improve grammar, to translate from your own language to one you are not fully fluent (as English for many of us), to ask feedback on formulation, to test arguments, to screen abstracts, to summarise a body of literature, to generate a section from a prompt and editing the output, to fabricate results, references or data. And so on.

These are not points on a scale of increasing AI involvement. They are fundamentally different acts, and their distinction is not how much the machine was involved, but where the thinking happened and who can answer for it. The first examples above leave the intelectual work with the author, others describe delegation of mechanical operation under human defined criteria, which is something instruments are meant to do and entirely defensible if the criteria and the verification are stated. The next examples substitute human activities to machines without checking fidelity, and finally, the last refers to pure fabrication, and is not an AI problem at all: it is research misconduct done using a new instrument.

Treating all of these as the same category 'AI use' comes with two costs: it stigmatises uses that actually improve the work, particularly for researchers writing in a second or third language, and obscures the uses that genuinely damage the record, because a disclosure line saying "AI was used in the preparation of this manuscript" is compatible with all the cases above.

Four things that follow:

1. Accountability does not move. Authorship is not just a credit claim, it is a commitment that someone can be asked to defend what is written. AI systems cannot be asked, cannot answer and cannot be sanctioned. Whatever tools are used, the responsibility stays exactly where it always was. The SAM guidelines put this first, where it rightly belongs: users, and authors, are accountable for understanding the capabilities and risks of the tools, and for the integrity of what those tools help produce. This is also the practical reason why AI systems cannot be authors, apart from philosophical arguments. The point is not whether AI systems can be creative, is that they cannot hold the account.
2. Quality is about reasoning, not prose. LLMs produce fluent, well-structured, confident text. Fluency may have been historically a weak but usable proxy for quality, because writing clearly is hard and the effort correlated with thinking. Now this correlation is broken and quality needs to move from surface to structure. This is hard for many reasons, mostly practical rather than conceptual: checking prose is quick, checking that arguments, conclusions or citations are what they claim to be is harder and takes expertise and time that review systems don't account for.
3. Involvement is part of what a text is. Writing is not just the production of text, it is the process by which authors come to hold a position. The hardship of writing is where the thinking happens. Behind a good paper lies a lot of reading, many abandoned arguments, experience that gives the author a sense of which findings are solid and which are weak. None of this is visible in the sentences, but is what makes a text worth reading. Text produced without this process does not loose style, it looses the link between the author and the claim.
4. Caring is the thing at stake. Scientific texts carry authority because someone took the trouble. They read the difficult papers, chased the inconsistent findings, argued with colleagues, stayed with the problem until it made sense. What erodes trust is not that the scientist used AI, it is the suspiction that no one actually did the work, that the text was assembled rather than reasoned, and that if you pushed at any part of it, there would be no one behind. This suspiction is the real risk and it is not solved by disclosure rules.

The question is not how much AI is used. Is about being an author. It is not a credit to receive but is to put yourself behind something. What is at risk is the experience of authorship itself: the work of coming to know something and to stand behind saying it.

None of this is new. Codes of research integrity have long required authors be answerable for the work. AI made clear how weakly this requirement was enforced. The remedy is not complicated. The SAM guidelines point the way when they ask that the tasks in which AI was applied be recorded, not merely the fact that a tool was used. This is the shift that authorship needs. It takes judgment, and judgment cannot be delegated to the tool being declared.