OpenAI has confirmed that an invisible watermark called textGrain will be attached to text generated by ChatGPT and Codex across the European Union within the coming weeks, on every plan. Announced on October 5, 2026, the measure changes nothing about how output looks: read, copy or paste it and it appears identical. The signal lives in the statistical pattern of word choices, and only a dedicated detector — not currently public — can surface it. The driver is legal: under the EU’s AI rulebook, synthetic text must be identifiable by machine.
That matters for anyone writing with ChatGPT from Europe, though OpenAI’s own figures describe a fragile mark and a detector that proves very little.
- Invisible by design: textGrain adjusts which words the model picks rather than embedding hidden characters, so an ordinary copy-paste will not remove it.
- Fragile under editing: OpenAI’s tests show that replacing a quarter of the words in a passage cuts detection to roughly 17%.
- Compliance, not proof: the company is explicit that the mark cannot establish who wrote a text or how much human work shaped it.
What textGrain Is
It is a statistical watermark. Rather than inserting hidden characters or odd spacing, it slightly biases which words the model selects, leaving a pattern a matching detector can recognise. OpenAI says the approach matched or beat the alternatives it tested, including SynthID for text from Google DeepMind, without appreciably harming response quality, and that it plans to open-source the system — though no date is set.
The rollout has three tiers. In ChatGPT and Codex the mark is automatic for EU users, with no opt-out mentioned. Through the API it is optional worldwide and off by default. Outside the EU it will not be the default behaviour, at least at launch.
How Easily the Mark Is Erased
Mostly by editing. Testing 400-token passages, OpenAI found that swapping 10% of words for synonyms pulled detection from around 92% to 66%, while swapping 25% left it near 17%.
Length counts too: at a 1% false-positive rate the detector caught close to 80% of 200-token responses and around 95% of 400-token ones. Mathematical text fared noticeably worse, since there are fewer ways to phrase the same result. Language is another variable: a breakdown of the technical report cited by the specialist outlet DataNorth put detection between 42.2% in Romanian and 69% in Spanish — second-hand figures not confirmed in OpenAI’s own post.
OpenAI is candid about the limits, which deserves credit: a failed detection proves nothing about human authorship, and a successful one says nothing about which account generated the text or how much human editing shaped the final version.
Why the Rollout Stops at the EU Border
Article 50 of the AI Act requires providers to mark synthetic content in a machine-readable format, an obligation live since August 2, 2026. Systems already on the market before that date received a grace period to December 2, 2026, introduced by the Digital Omnibus package according to a Cloud Security Alliance note; OpenAI’s “coming weeks” fall inside that window. Penalties reach €15 million or 3% of global turnover.
OpenAI is still behind its closest rival. Anthropic watermarks everything Claude produces and applies it globally, not just inside the EU — two opposite readings of one rule. Restricting the mark to the territory that demands it is minimal compliance, and arguably sound commercial logic, since a global watermark could have pushed users toward alternatives.
What Changes for Publishers and Agencies
Provenance now leaves a trace, even if almost nobody can read it yet; the detector is granted case by case to researchers and expert organisations.
Three consequences stand out. Bulk text generated through the API carries no mark unless you enable it, and that remains your decision. Teams working inside and outside the EU will produce a mix of marked and unmarked output. And running copy through a “humaniser” to strip the signal changes nothing about who is accountable for it. The market for watermark-removal tools will grow, and OpenAI’s numbers are its best advertisement: if altering one word in four leaves detection at 17%, any automated rewrite will do.
TechNews DevTop’s view is that a system failing roughly one time in five on short texts, unable to tell generating from correcting, and flagged for false positives by its own maker cannot sustain an accusation against a student or a staff writer. The mark delivers the legal minimum and little else. Earlier this year, industry voices warned that Article 50 labelling would go widely unenforced; today the two biggest labs have watermarked text or committed to a date, so the rule has worked formally rather than materially. A mark that vanishes the moment you edit guards against careless use, never deliberate misuse — and serious disinformation was never careless. The defence worth having is knowing who reviewed each text before publication, the first thing a reader, a client or a court will ask.