Two questions decide whether the EU AI Act's transparency rules already reach something you write. The marking and disclosure obligations were built with deepfake video in mind, so does a rule like that actually reach plain text? And if it does, does writing it as research get you out from under it? Both answers turn out to be less reassuring than the "this is about images" reading suggests, and Article 50, the part of the regulation aimed squarely at transparency, is where both get settled: providers of systems that generate synthetic content have to mark it in a machine-readable, detectable way, and deployers who generate or manipulate AI text on matters of public interest have to disclose that fact (Article 50, EU AI Act Service Desk; full regulation text, EUR-Lex). It has applied since August 2, 2026, and non-compliance carries fines up to €15 million or 3% of global turnover, whichever is higher.
| Applies since | August 2, 2026 |
| Maximum penalty | €15 million, or 3% of global annual turnover |
| Code of Practice on marking and labelling | Published June 10, 2026 |
| Grace period for systems already on the market | Runs until December 2, 2026 |
The clause that actually reaches long-form writing
Most Article 50 commentary is about images and video, for good reason: the machine-readable marking requirement was built for synthetic media.
"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."
Article 50(2), Regulation (EU) 2024/1689
The Commission's Code of Practice on marking and labelling, published June 10, 2026, spends most of its technical detail on embedding provenance metadata in file formats that support it (IPTC; Wilson Sonsini). Plain text has no equivalent embedding standard, which is part of why it keeps getting treated as an afterthought.
But Article 50(4) is not only about deep fakes:
"Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated."
Article 50(4), Regulation (EU) 2024/1689
There are narrow carve-outs where the content has gone through human editorial review and a natural or legal person holds editorial responsibility (Article 50 practical guide, artificialintelligenceact.eu; Orrick). But scientific and policy writing that reaches a public audience sits close enough to "informing the public on matters of public interest" that the "we mostly deal with text, this is really about images" reading looks increasingly optimistic (Herbert Smith Freehills Kramer).
The research exemption is narrower than it sounds
That answers the first question. The second is the exemption researchers reach for by instinct: surely writing something as research gets you out from under a rule aimed at deepfakes and disinformation? The Act does carve out research:
"This Regulation does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development."
Article 2(6), Regulation (EU) 2024/1689
"This Regulation does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service. Such activities shall be conducted in accordance with applicable Union law. Testing in real world conditions shall not be covered by that exclusion."
Article 2(8), Regulation (EU) 2024/1689
Read closely, both exemptions are about the AI system, not about the output once it leaves the lab. Legal scholarship on the point has been blunt about the gap: the exemption "is without prejudice to the obligation to comply... where an AI system is placed on the market or put into service as a result of such research and development activity," and applying it to how contemporary AI research actually gets published and reused turns out to be genuinely hard (GRUR International; npj Digital Medicine). A research group experimenting internally with an AI writing tool is plausibly exempt. The paper that experimentation produces, once it is submitted for publication and starts informing public claims, is a different object, and the exemption's rationale, protecting freedom of scientific inquiry, was never really about that object.
Disclosure statements again
This is where Article 50 runs into the same wall as every other transparency mandate: nothing in the regulation specifies how a deployer proves what they disclose is accurate. In practice, for text, that collapses back into a disclosure statement, an unverifiable claim made by the person with the least incentive to overstate their AI use. We have written about why that specific failure mode matters for research writing already, and Article 50 does not change the underlying problem; it just adds a legal deadline and a penalty schedule to it.
What actually closes the gap
Both threads above turn on the same detail: whether a given piece of text was AI-generated, and when, which is exactly what nobody can reconstruct accurately once the writing is finished. A TeXposit project keeps that answer as it happens instead of leaving anyone to remember it later. Every save is attributed, so the provenance report it produces is a signed, per-passage record of what was typed directly versus accepted from an AI suggestion, something a reviewer, an editor, or eventually a regulator can check rather than take on faith.
None of that makes TeXposit the "provider" or "deployer" Article 50 regulates, and using it doesn't make an author compliant by fiat; that obligation sits with AI system operators, not the person writing the paper. What it gives you is smaller and more useful than compliance: when a rule written for images and video turns out to reach further into a manuscript than expected, there is already a record to point to instead of a promise to make.