OpenAI to watermark ChatGPT text in the EU as AI Act transparency rules bite
OpenAI will start adding invisible watermarks to ChatGPT's text output in the European Union. The company described the phased rollout, aimed at the AI Act's transparency rules, on 5 October.

OpenAI will start watermarking text that ChatGPT generates for users in the European Union. The company described the approach in its own account of EU text provenance rules, dated 5 October. It is a phased rollout, not a single switch-on.
The mark is meant to let tools identify output as machine-generated. That matters to anyone who reads, edits or publishes ChatGPT text in Europe. It also matters to platforms that must decide what to do with a detection signal they cannot see.
OpenAI has not said publicly how reliable the mark is. Nor has it said how the mark survives editing or translation, or whether third parties will get access to a detector. Those gaps sit at the centre of the debate.
A rule with a deadline behind it
The AI Act is the EU's broad framework for artificial intelligence. Its transparency obligations are the part now reaching products people use daily. OpenAI's move is a compliance step, not a product feature. The rest of the industry is watching how Brussels reads it.
Coverage of the rollout was wide and immediate. TechCrunch reported that OpenAI will start watermarking ChatGPT's text in the EU. PYMNTS framed the same step as meeting AI Act mandates. ActuIA noted the split that matters most for developers: the watermarking applies to the consumer product, but the API stays opt-in. MLex and The Tech Buzz both reported the watermarking plan. Interesting Engineering described the marks as invisible stamps added to ChatGPT output under EU rules. Crypto Briefing reported the optional API feature in the same terms.
That distinction is the practical core of the story. A chat window and an API call produce the same kind of text. Only one of them gets marked by default. Anyone building a product on top of OpenAI's models must decide whether to switch the feature on.
The API opt-in means the watermark's reach depends on developers, not on the regulator.
None of the coverage reviewed here puts a figure on how many EU users are affected. None says when the phased rollout completes.
Why the text side is harder than images
Watermarking pictures is a solved enough problem that platforms ship it quietly. Text is different. A sentence can be paraphrased, translated, shortened or pasted into another document. Each step is a chance for a statistical signal to weaken.
OpenAI calls its system text provenance, a broader idea than a single hidden marker. It covers ways of showing where a piece of text came from. The company has not spelled out in public whether that means a durable watermark, metadata, or a combination.
The stakes are highest for the people who would use a detector. Teachers grading essays, editors checking submissions and platforms moderating content all want a yes-or-no answer. A probabilistic signal gives them something weaker, and a wrong accusation is costly.
Brussels has its own reasons to want this to work. The AI Act's transparency rules are among the few parts of the framework that touch ordinary users directly. Enforcement will be judged on visible cases.
The wider argument about whether rules can be tested
The watermarking rollout lands in the middle of a louder fight about whether the EU's approach can be enforced at all. Cybersecurity Insiders argued on 4 October that the AI Act cannot regulate what regulators have not learned to test. That is a pointed criticism of the gap between legal text and measurement.
Tech Policy Press made a related case the same day. It described the EU's tech regulation as a paradox that is getting harder to ignore. The two pieces come from different directions but land on the same worry: obligations that sound clear in a statute can be hard to verify in a product.
Watermarking is a test of that proposition. If a mark can be detected reliably, the rule has teeth. If it cannot, compliance becomes a declaration.
There is a geopolitical edge too. Bruegel noted on 5 October that the US and China will have a red phone for AI emergencies while Europe is left to watch. The EU is writing rules it does not enforce alone.
What the ads debate says about the money behind all this
Compliance costs money. The companies paying for it are spending at a scale that shapes what they can afford. In a Forbes Agency Council post, Tanuj Joshi, chief executive of the marketing platform Eulerity, wrote that OpenAI generated $13 billion in revenue in 2025 against an estimated $22 billion in infrastructure costs.
Those figures come from a council contributor, not from OpenAI's accounts. They should be read that way. The same piece says OpenAI has reportedly committed to an estimated $600 billion in infrastructure spending through 2030. It also says Anthropic projects roughly $200 million on compute over the next three years.
Joshi also describes the fight over advertising in AI. He cites OpenAI's January 2026 announcement that it would bring ads to free ChatGPT users, an Anthropic Super Bowl ad mocking the decision, and Sam Altman's response on X. The piece argues that roughly 850 million free users are unlikely to pay a $20 monthly subscription.
That context explains why a watermark matters commercially as well as legally. An ad-supported chatbot at scale needs trustworthy signals about what is machine-written. Advertisers want that comfort.
What to watch
The next signals are concrete. Watch whether OpenAI publishes detection rates. Watch whether the API opt-in becomes a default. Watch whether other model providers follow with their own provenance schemes or wait for Brussels to specify one.
Watch, too, whether the Commission treats a phased rollout as compliance or as a work in progress. Neither OpenAI nor the EU has said publicly how the rollout will be assessed.
For European users, the change will be invisible by design. For everyone building on these models, the opt-in decides whether the rule reaches them at all.
Sources
1All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.