From 5 October 2026, the texts that ChatGPT and Codex draft for European users carry a signature no one can see. OpenAI has called it textGrain: it is the company’s response to Article 50 of the AI Act, which requires providers of generative AI to make their content recognisable by a machine.
The mechanism adds nothing to the text. There are no hidden characters or invisible spaces: copying and pasting into another program does not remove the trace. The watermark acts on word choice. Among the possible alternatives, the model favours some according to a secret key, and across an entire passage these preferences form a statistical pattern that only a detector holding the same key can recognise.
Protection with variable geometry
The watermark is enabled by default on ChatGPT and Codex only in the European Union. Via the API it is available worldwide, but must be switched on by the customer. It is also fragile: according to OpenAI’s data, it is detected in around 80% of 200-token texts and 95% of 400-token texts, but if a quarter of the words are replaced, detection falls to 17%. Translation and paraphrasing easily erase it.
The regulatory framework explains the choice. Article 50(2) requires providers to mark outputs in a machine-readable format. The Digital Omnibus did not postpone this: it merely granted a grace period until 2 December 2026 for systems already on the market before 2 August. Article 50(4), by contrast, requires deployers to label visibly deepfakes and texts published to inform the public on matters of public interest. This second obligation has applied since 2 August and does not fall away when a watermark is present: marking speaks to software, not to readers.
How robust the watermark is
Percentage of texts in which the detector recognises the signal, according to OpenAI’s data.
200-token texts80%
400-token texts95%
After replacing a quarter of the words17%
Translation and paraphrasing erase the signal just as easily. Source: OpenAI.
What changes for professional firms and businesses
For professional firms, the watermark makes verifiable, at least in theory, what until now depended solely on the professional’s integrity. Law 132/2025 already requires clients to be informed about the use of AI systems: it is advisable to do so in writing, in the engagement letter, in a way consistent with the firm’s actual practice. Paradoxically, legal documents are among the least traceable texts: recurring formulas, statutory citations and technical language leave little room for the lexical variation the watermark needs. The obligation towards the client, however, does not depend on how detectable the signal is.
For businesses, the decisive question is whether they are mere users or providers. A company that uses ChatGPT internally is a deployer: marking falls to OpenAI, while the company remains responsible for labelling content subject to Article 50(4). A company that integrates the models via API into its own product, offered to the market under its own brand, may instead take on the role of provider. In that case, leaving the watermark switched off is not a neutral choice but a compliance decision that must be justified and documented. The same applies to contracts with technology partners, which should make clear who activates the marking, who retains proof of it and how responsibilities are allocated.
There is also a practical limitation: the detector is not public. OpenAI grants access case by case to research bodies and academics and excludes commercial uses, including corporate verification, as well as the checking of students’ work. Today, a business cannot establish on its own whether a document it has received was written with ChatGPT.
The dates that matter
2 August 2026
The visible labelling obligation takes effect
Article 50(4) requires deployers to label recognisably deepfakes and texts published to inform the public on matters of public interest. This is the obligation that directly concerns professional firms and businesses.
5 October 2026
OpenAI switches on textGrain
The watermark is enabled by default on ChatGPT and Codex, for European Union users only. Via the API it is available everywhere, but must be switched on by the customer.
2 December 2026
The grace period ends
The Digital Omnibus did not postpone Article 50: it only gave systems already on the market before 2 August until this date to comply with machine-readable marking.
The two obligations are not interchangeable: the watermark speaks to software, the label to readers. Having one does not exempt you from the other.
A clue, not proof
The watermark is likely to be invoked in court, in copyright disputes as well as in professional liability claims. Its weight, however, is limited. In civil proceedings, a positive result can at most serve as circumstantial evidence from which to draw a simple presumption under Article 2729 of the Italian Civil Code. The result is probabilistic, with a margin of false positives. The opposing party cannot repeat the test, because the key is secret, which raises evident problems for adversarial technical scrutiny. And OpenAI itself acknowledges that the watermark does not reveal who the author is, nor how much a person contributed.
The reverse is also true: the absence of the signal does not prove that a text is human. On copyright, moreover, the criterion remains human creative contribution, and a watermark cannot on its own deny protection to a work created with the help of AI.
The real change, then, is not technological. It is that transparency about the use of AI ceases to be a statement of principle and becomes, at least in part, a verifiable fact. For professional firms and businesses, now is the time to put things in order: disclosures, internal policies, contracts.
Domande frequenti
What is textGrain?
TextGrain is the text watermarking technology developed by OpenAI. It does not insert hidden characters or invisible spaces: it statistically alters the choice of words and tokens during generation, creating a pattern that a detector holding the correct key can identify.
Does copying or editing a text remove the ChatGPT watermark?
Simply copying and pasting does not remove textGrain, because the signal is embedded in the wording of the text. Deeper edits can, however, weaken it: in tests published by OpenAI, replacing 25% of the words with synonyms significantly reduced the detection rate. Substantial paraphrasing and translation can also make the watermark undetectable.
Is it possible to check whether a text was written with ChatGPT?
Not freely. OpenAI has not made the textGrain detector public and currently grants access only to academic and research organisations with eligible use cases. Moreover, a positive result indicates only that an OpenAI model probably generated or processed the text: it does not identify the author, nor does it measure how much a person contributed.
Fonti e riferimenti
- Commissione europea, Code of Practice on Transparency of AI-generated Content
- Commissione europea, Guidelines on transparency obligations for providers and deployers of certain AI systems
- Commissione europea, Transparency obligations under Article 50 of the AI Act
- EUR-Lex, Regolamento (UE) 2026/1744
- Normattiva, Legge 22 aprile 1941, n. 633, Protezione del diritto d’autore
- Normattiva, Legge 23 settembre 2025, n. 132, Disposizioni e deleghe al Governo in materia di intelligenza artificiale
- OpenAI, Il nostro approccio alle norme UE sulla provenienza dei testi
- OpenAI, Segnali di provenienza nei contenuti generati da OpenAI

