OpenAI Tags Every EU ChatGPT Answer — a Thesaurus Beats It

Every ChatGPT answer generated in Europe is about to carry an invisible tag — and OpenAI's own tests show a thesaurus can defeat it. On 5 October, the company announced textGrain, a statistical watermark it will bake into ChatGPT and Codex text across the European Union over the coming weeks, to satisfy the EU AI Act's transparency rules. The twist, printed in OpenAI's own 20-page technical report: replace one word in ten with a synonym and detection collapses from 92% to 66%. Replace a quarter of them, and it sinks to 17%.

So Europe gets a watermark that misses a third of lightly edited text — while the detector that reads it stays locked away from the public. This is the strangest product launch of OpenAI's year: a feature designed to be invisible, admitted to be fragile, and switched on only because Brussels said so.

OpenAI chief executive Sam Altman onstage at TED. His company will begin adding its invisible textGrain watermark to ChatGPT and Codex text in the EU on a phased rollout announced 5 October 2026. | Photo: Steve Jurvetson, CC BY 2.0, via Wikimedia Commons

What OpenAI actually announced

The announcement came in a company blog post titled "Our approach to EU text provenance rules," published on Monday 5 October and reported the same day by TechCrunch, Gizmodo, The Verge and The Hindu. It contains three moving parts.

First, the EU rollout. Over the coming weeks, OpenAI will add an invisible watermark to "eligible ChatGPT and Codex text output in the European Union," across every plan, free tier included. In the EU it is not a setting and there is no off switch — once the rollout reaches you, your words are tagged.

Second, the API. From 5 October, developers anywhere in the world can opt in to watermarked output for select models — but the switch stays off unless they flip it. OpenAI says it is also working with cloud partners to offer the option for its models served through their platforms; The Decoder named Microsoft Azure among them.

Third, the detector. Applications opened on 5 October, but only approved researchers and expert organisations get in, decided case by case. OpenAI says it plans to release textGrain as open source so others can build on it, and published a technical report co-written with researchers from the University of Pennsylvania and Yale.

On quality, the company's numbers are reassuring — for its newest frontier model, Astra, OpenAI reports an Artificial Analysis Intelligence Index of 49.57 unwatermarked against 49.76 watermarked, and 94.44% versus 93.94% on GPQA Diamond: moves in both directions that it calls normal run-to-run noise. Its existing public tools for images and audio — the openai.com/verify web page and the Content Provenance API — stay open to everyone.

How textGrain works — and where it breaks

There is no hidden ink here. As the technical report describes it, textGrain uses a secret key to sort the model's next-word predictions, nudging its choices toward a statistical pattern a reader cannot see. Add hundreds of those nudges across a long passage and a detector holding the key can spot them. No zero-width characters, no invisible spaces, no odd punctuation — and because the signal lives in the words themselves, it survives copying and pasting into another document.

That elegance has hard limits, and OpenAI is unusually blunt about them. With the detector tuned to a 1% false-positive target, it found the watermark in about 95% of 400-token psychology passages but only about 80% at 200 tokens. Mathematics scored "substantially lower," at roughly 60%, because equations leave little room for wordplay. Short passages, translations and edited text all degrade detection — and the EU's own guidance excludes code snippets and responses under 200 tokens, about 150 English words, from watermarking expectations anyway.

Then comes the thesaurus test. We checked the figures against three independent write-ups of the technical report — TechCrunch, Ars Technica and BleepingComputer all land on the identical numbers, so this isn't a game of telephone. In OpenAI's evaluation of 400-token answers drawn from the ELI5 dataset, swapping 10% of words for synonyms cut detection from about 92% to 66%; swapping 25% cut it to 17%. One outlet, ActuIA, citing the report, puts non-English detection at 69% for Spanish and 42.2% for Romanian.

OpenAI's framing is careful to the point of legal: a detected watermark answers exactly one question — "was this likely generated by an OpenAI model?" It does not say who wrote the text, who owns it, how much a human contributed, or whether any of it is true. And, in the company's own words: "The absence of a detected watermark does not prove human authorship."

That 1% false-positive target deserves a second look, too. It sounds tiny. Applied to a thousand genuinely human-written essays, it means roughly ten innocent people flagged as machine-made — which is precisely why the detector is staying out of public hands for now.

The ghost of 2022: OpenAI built this before — then shelved it

None of this is new technology. It is new politics.

Back in 2022, computer science professor Scott Aaronson — then working as a guest researcher at OpenAI — described the idea in a lecture: whenever a system such as ChatGPT generates text, he said, the tool would incorporate an "unnoticeable secret signal," revealing where the content came from. The prototype was designed by OpenAI developer Hendrik Kirchner, and by Aaronson's account it worked fairly well.

OpenAI sat on it. In August 2024, after the Wall Street Journal revealed the company had a working text watermark it had chosen not to release, OpenAI admitted in a blog update that the method was "trivial to circumvention by bad actors" and could "stigmatize use of AI as a useful writing tool for non-native English speakers," as TechCrunch reported at the time. An internal worry, widely reported, was simpler: watermarked ChatGPT might lose users to rivals that didn't tag their text.

What changed is the law. Brussels didn't ask nicely — it legislated, and the watermark OpenAI once feared would cost it customers is now the price of operating in Europe.

Brussels made them do it

The Berlaymont building in Brussels, seat of the European Commission — the regulator whose AI Act transparency rules forced OpenAI's hand on text watermarking. | Photo: tpholland, CC BY 2.0, via Wikimedia Commons

The forcing function is Article 50(2) of the EU AI Act, in force since 2 August 2026, which requires providers of generative AI to mark synthetic text, audio, images and video in a machine-readable, detectable format. The European Commission published detailed guidelines on 20 July, and on 24 July said systems already on the market before 2 August get a limited grace period — full compliance on marking and detection by 2 December 2026.

Around 190 providers had signed the EU's voluntary Code of Practice on Transparency of AI-Generated Content by the end of July, including Anthropic, Google, Meta, Microsoft and OpenAI — The Hindu notes that Anthropic, Google, Meta, Microsoft, Mistral and OpenAI all signed its first section, on marking and detection. As Henna Virkkunen, the Commission's Executive Vice-President for Tech Sovereignty, Security and Democracy, put it when the guidelines landed: "With today's guidelines, the Commission supports the smooth and effective application of the AI Act to make AI systems interacting with people such as chatbots and AI agents and AI content more transparent and trustworthy."

For the record on timing: we reported earlier this month that these labelling obligations began in August — not, as one UK site claimed, on some later Sunday. textGrain is simply OpenAI's concrete implementation arriving two months after the deadline, and it does not contradict that earlier story.

It is also worth noting what Brussels is not demanding: the law doesn't mandate watermarks specifically, only machine-readable marking "as far as technically feasible." OpenAI chose the watermark; the Commission will judge whether 66%-after-a-thesaurus counts as effective.

The watermark stops at the EU border

Here is the split-screen the announcement creates. Ask ChatGPT in Paris for a 300-word essay and OpenAI's own detector should spot its hidden watermark about 95% of the time. Write the same essay in Manchester and nothing is tagged at all — Britain left the EU, the AI Act doesn't apply, and Westminster has written no equivalent rule. An ordinary ChatGPT user in Leeds gets no watermark; a British company building on the API can switch one on, if it bothers.

Across the Atlantic, there is no federal equivalent either. While Washington's most recent AI executive order was about renaming the technology "Super Intelligence," Brussels was building the machinery to make machines confess. The Commission's tech chief has been carrying that message to Washington in person.

OpenAI is not alone in the tagging business. Anthropic began watermarking Claude text worldwide in August, built on Google DeepMind's SynthID approach, for the same EU reason. Google itself has shipped SynthID for years — then, in a move Gizmodo called strange, announced in August that users could strip watermarks from images, video and music in countries where the law doesn't require them. Regulation, it turns out, is the only thing currently holding the industry's tags in place.

As OpenAI's official account wrote on X on 5 October: "We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology." Read that again: even the launch announcement is, in part, a disclaimer.

Our verdict: a shoddy shield, a decent checkbox

Let's be blunt. As an anti-cheating device, textGrain is colossally inadequate: any student with a thesaurus and five spare minutes can drag detection down to a coin flip, and OpenAI's own report proves it. As a plagiarism shield it is shoddy; as a courtroom exhibit it would be laughed out. A watermark that survives copy-paste but not paraphrase is a lock on a screen door.

But that was never really the point, and to OpenAI's credit, it isn't pretending otherwise. The company published the damning numbers itself, kept the detector away from trigger-happy accusers, and told the truth — in writing — that absence of a watermark proves nothing about authorship. That is more honesty than the AI industry usually manages about its own detection tools.

What textGrain actually is: a compliance checkbox, decently engineered, that lets OpenAI tell Brussels "we marked it in a machine-readable format" while telling everyone else "don't rely on it." Whether the Commission accepts 66% as "effective, interoperable, robust and reliable as far as technically feasible" is the next fight — and with the grace period closing on 2 December, it is coming soon.

Until then, Europe's ChatGPT answers will carry a tag most users will never see, readable only by a detector most people will never touch, defeated by an instrument most schoolchildren own. The thesaurus, it turns out, was the most powerful AI-safety tool all along.

Sources

TechCrunch, Gizmodo, The Hindu, The Times of India, News18, Ars Technica, BleepingComputer, The Verge, Unite.AI, TechQuire, TechRepublic, CHOSUNBIZ, ETV Bharat, Pivot News and Dig.Watch — reporting on OpenAI's 5 October 2026 announcement, its technical report, and the EU AI Act transparency framework. Key figures were cross-checked across three independent write-ups of OpenAI's evaluation data.

— TechPulse USA

TechPulse USA is an independent technology news publication covering artificial intelligence and the global tech industry for a US-first audience. Every claim in our reporting is verified against primary sources before publishing.