Research Solutions | Blog

What The EU AI Act's Watermarking Rules Mean For Researchers

Written by Research Solutions | Marketing Team | Nov 20, 2026, 2:17:33 PM

Since August 2, 2026, the EU AI Act's Article 50 transparency rules have applied to generative AI systems serving the European market (systems already on the market before then have until December 2, 2026, to add machine-readable marking). Providers must disclose when people are interacting with AI, and AI-generated text must carry machine-readable marking that software can detect after the fact. Around 190 signatories, including the major model developers, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, and Anthropic has said Claude's outputs carry a text watermark.

If you use AI tools in your research, the first thing to know is that most of these obligations don't land on you. They land on the companies that build and operate the AI systems. But the rules do change a few things about how AI-assisted work gets published, and the watermarking piece has generated more anxiety than it deserves. They also arrive while the research community is separately working out what good AI disclosure looks like.

 

The Rules Bind Providers First

Article 50 draws a line between providers and deployers. The provider is the company placing a generative AI system on the market under its own name. That company owes the disclosure ("you're talking to an AI") and the machine-readable marking of generated text. A researcher using an AI assistant is a deployer, and deployer obligations are narrower. They mostly attach at one moment: publication.

So, the compliance clock ticking toward late 2026 belongs to the AI providers, not to the people using their tools. Your part starts when AI-assisted text leaves your desk and informs the public.

How A Text Watermark Works

The word "watermark" suggests something stamped onto the text. Text watermarking generally works differently. When a language model generates text, it constantly chooses among near-equivalent words, the kind of choice where "overcast" and "grey" would both do. A watermark uses a cryptographic key to settle those low-stakes choices instead of an arbitrary random number. The words remain natural, and the text reads identically. But anyone holding the key can later check whether the sequence of choices matches the pattern the key would produce and assign a probability that the model was involved.

This method isn't specific to any one provider; the signatories to the Code of Practice are each implementing their own versions, with their own keys. Anthropic's explainer spells out three properties that follow from the design. Nothing is inserted into the text, so there are no invisible characters to find or strip. The watermark carries no information about the user, the organization, or the conversation. And detection is probabilistic. It works poorly on short passages, on heavily factual text where word choices are constrained, and on text a person has substantially rewritten. Running a check isn’t widely available yet, either: the detection tool is in private preview for regulators, researchers, media, and organizations with their own compliance obligations, so most publishers can’t verify a document today. A watermark can indicate that a model was likely involved somewhere. It cannot distinguish drafting from heavy editing, and it says nothing about authorship or ownership. A watermark can indicate that a model was likely involved somewhere. It cannot distinguish drafting from heavy editing, and it says nothing about authorship or ownership. That gap shows up on the editor’s desk. Publishers’ own research-integrity leads acknowledge that AI-detector false positives are common, and advise treating a flag as a hint, not a verdict. A watermark is a better-founded signal than a detector’s guess, but it still can’t tell an editor what the AI did or who stands behind the result. That’s what disclosure is for.

Publishing With AI Assistance

For researchers, the substantive provision is the deployer rule on published text. If you publish AI-generated text to inform the public on matters of public interest, you either disclose the AI's involvement, or the work needs to have undergone human review with a person or organization holding editorial responsibility for it.

That second path describes most scholarly and regulatory workflows already. A manuscript that went through drafting, revision, co-author review, and peer review has human editorial control at every stage. The same holds for regulatory submissions. The carve-out exists precisely for work where people review, verify, and take responsibility for the output. One exception: AI-generated or manipulated images and figures fall under separate, stricter disclosure rules, and the human-review carve-out doesn't cover them.

The Law Sets A Floor, Journals Set The Standard

Meeting the Article 50 carve-out settles your obligation under the AI Act. It doesn't settle what your journal expects. ICMJE, COPE and most major publishers already ask authors to disclose substantive AI use and to confirm that humans remain accountable for the work, whatever editorial review took place.

Those expectations are now converging. The Global Reporting Standard for AI Disclosure in Research, known as the Vancouver Standard, came out of the 2026 World Conference on Research Integrity and is being developed by the International Science Council, COPE, STM and the Global Young Academy. Modeled on contributor taxonomies like CRediT, it is working through which AI uses are significant enough to disclose, where disclosure belongs in a manuscript, a shared taxonomy of AI-assisted tasks, and how authors signal accountability. Consultation is ongoing and a refined standard is expected during 2027.

Biomedical publishing has been here before. Company-sponsored research spent two decades making non-author contributions visible: medical writing support, sponsor roles and funding are now routine disclosures under Good Publication Practice. The International Society for Medical Publication Professionals (ISMPP) is bringing that experience to the AI question, arguing for proportionate disclosure built on existing ICMJE and GPP workflows and focused on what the human is accountable for.

Keep humans in the loop, keep records of review, and disclose to your journal's standard. The watermark doesn't decide whether you met your obligations. Your editorial process and your disclosure do.

Retrieved Content Sits Outside The Marking Rules

Article 50's marking obligation attaches to synthetic content, meaning text the model generated. Search results, citation records, bibliographic metadata, and verbatim excerpts surfaced from the published literature are retrieved content, and the marking obligation doesn't attach to them.

That distinction is why grounding tools work the way they do. Within the context of Scite, the connector supplies citation context and metadata to an AI assistant, and the marking applies to what the assistant generates, not to the underlying records. The supporting evidence remains what it always was: the published record, traceable to its source. The same split runs through the disclosure debate, where proposed taxonomies treat literature search and text generation as different AI tasks.

For researchers, the retrieved half is the part that carries the trust. It was the published record before any model touched it, and it stays that way after.

What To Do Between Now & December

The near-term deadlines belong to AI providers, and the Code's signatory task forces will keep sharpening what compliant marking looks like. In the meantime, four habits cover the deployer side: know which of your tools generate text and which retrieve it, keep substantive human review in every publication workflow, check your target journal's AI disclosure policy (and expect policies to converge on the Vancouver Standard over the next year), and disclose AI involvement where no editorial process stands behind the text.

A watermark tells you a model was involved. A disclosure tells you how. A citation tells you where the claim came from.

The transparency rules reward research workflows that were defensible all along, the ones where every claim traces back to a source, and a person stands behind the result.