Adapting the AI Act's transparency logic for scientific disclosure

Three design decisions under Article 50

Article 50 of the AI Act applies from 2 August 2026, and it places four duties on two sets of people [1]. Providers must design systems that interact directly with people so the person is informed they are dealing with an AI, unless that is obvious to a “reasonably well-informed, observant and circumspect” person; and providers of systems generating synthetic audio, image, video or text must mark the outputs in a machine-readable format and make them detectable as artificially generated or manipulated [1]. Deployers must give notice when an emotion-recognition or biometric-categorisation system is in use, disclose deep fakes covering image, audio or video, and disclose AI-generated text published to inform the public on matters of public interest — in each case “clear[ly] and distinguishab[ly]”, at the latest at the time of first interaction or exposure, and conforming to accessibility requirements [1].

Strip away the subject matter and three design decisions remain. Disclosure is clear and arrives at first exposure, not buried where a diligent reader might eventually find it. The obligation is proportionate to what the system actually did. And marking is machine-readable, so verification does not depend on a reader noticing anything [1]. Even the hedge is instructive: technical solutions must be effective, interoperable, robust and reliable “as far as this is technically feasible”, taking into account the state of the art and the cost of implementation [1]. None of this was written as a research-integrity rule, and it should not be read as one. But those three decisions are a better starting point for disclosing AI in scientific work than the yes-or-no checkbox most research guidance currently offers.

Proportionality is the part worth borrowing

The exemptions are where the proportionality lives. The marking duty “shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof” [1]. For work that is “evidently artistic, creative, satirical, fictional or analogous”, the duty shrinks to disclosing that generated content exists, “in an appropriate manner that does not hamper the display or enjoyment of the work” [1]. And the public-interest text duty lifts entirely “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content” [1].

Read together, that is a graded scheme rather than a set of loopholes, and the instinct behind it is sound. An LLM tightening a sentence is not the same act as an LLM producing a figure, and a rule that treats them identically will be complied with in form and ignored in substance. Much research guidance still asks one flat question, did you use AI, which is exactly the flattening Article 50 took care to avoid.

The transplant needs one correction, and it is the whole argument. Article 50 asks a recognition question: can the audience tell that this artefact is synthetic? Science needs a consequence question: where did the model touch the research process, and does that change what a reader should believe? Those two come apart immediately. A methods section is not deceptive, and no reader is fooled by it. It can still be irreproducible because nobody recorded which model classified the behaviour, or on what prompt. The question is not hypothetical. Kobak and colleagues, measuring excess vocabulary across 15.1 million PubMed abstracts, put a lower bound of 13.5% of 2024 abstracts processed with a large language model: roughly 200,000 papers a year, and at least twice the size of the entire COVID literature at its peak [2].

What a scientific version would have to add

If disclosure scales with consequence rather than with tool use, the rungs look roughly like this, and journal policy has already climbed the first three [8].

• Language editing. A brief acknowledgement. ICMJE already places writing assistance in the acknowledgments [8].
• Substantive drafting. Name the model and record that a human reviewed it. ICMJE requires disclosure in both the cover letter and the paper, forbids AI authorship, and holds humans responsible for everything submitted [8].
• Analysis or figure generation. The methods section. ICMJE puts data collection, analysis and figure generation there, and where AI was used to conduct the study it must be described “in sufficient detail to enable replication of the approach, including the tool used, version, and prompts where applicable” [8].
• Agentic workflows. Versioned, machine-readable provenance. Once a system plans, calls tools and writes across many steps, a sentence in the acknowledgments cannot carry the information. What ran, in what order, on which inputs, is a record, not a remark.
• Anything that affects a scientific claim. Detailed enough to assess or reproduce. This is the rung that decides the others, because it names the trigger: consequence, not tool use.

One place the analogy should stop. Article 50 lifts the text-disclosure duty where a named person holds editorial responsibility [1], which is the right call when the harm being prevented is deception, since an edited and attributed piece is not passing itself off as anything. In science the same fact points the other way. A peer-reviewed paper with accountable authors is precisely where a reader still needs to know what the model did, because they are being asked to build on it. Human review should raise confidence in a disclosure. It should not be a reason to skip one.

Machine-readable, or it is only a statement

The fourth rung needs a format, and the record of research-transparency formats is not encouraging for anything that is merely asserted. A randomised trial asked authors to submit a completed ARRIVE checklist at submission: compliance did not improve, and no manuscript in either arm reported every subitem [3]. Where a journal made a checklist mandatory rather than requested, improvement in reporting randomisation, blinding and sample-size estimation was about threefold relative to a comparator journal [4]. Verified open-practice badges accompanied a rise in reported open data at one journal from under 3% to 39%, with badged datasets more often actually available [5]. And a data-sharing statement requirement, ICMJE's own, produced 68.6% of trials declaring they would share and 0.6% with data actually available [6].

The pattern is narrow but useful. A declaration that is only asserted behaves like the statement; one that something can check behaves like the badge. Two of the four are observational rather than randomised [4][5], and the one properly randomised test is the one where nothing happened [3]. None of them studied AI disclosure. This is a prior, not a prediction, but it is the prior a new disclosure standard should be designed against. Formats for exactly this already exist and cost one file each. AI-DECLARATION.md declares a level from none through auto, resolved per process and per directory [7]. trust.md separates an evidence record from assessments about it, lets conflicting assessments coexist rather than averaging them into one reassuring number, and states that conformance is not certification [9]. fair.md works through the FAIR sub-principles with inspectable evidence attached to every result [10].

This site runs all three, and the useful part was writing them rather than publishing them. Attaching inspectable evidence to every FAIR result, which fair.md v0.3 requires, exposed that this repository has no root LICENSE file while its manifests were asserting Apache-2.0 for the code. The licence claim was retracted rather than defended. Nobody audited that; the format did it, because a machine could check the answer. All three are self-reported and unaudited, which is the weakness the evidence above predicts and the reason to keep them mechanically checkable. Article 50 will never ask a laboratory for any of this. What it has done is show the shape of the ask: say it clearly, say it at first exposure, scale it to what the machine actually did, and write it where something other than a reader can check.