What happened
On October 5, 2026, OpenAI announced that it will add an invisible watermark to text generated by ChatGPT and Codex for users in the European Union. The company ties the change to the EU AI Act’s transparency obligations, which became applicable on August 2, 2026 and require AI-generated content to be marked so it can be identified. The watermark rolls out to eligible ChatGPT and Codex users on all plans in the EU over the coming weeks.
What’s new
- Method: called textGrain, described in a technical report written with researchers from the University of Pennsylvania and Yale. A secret key biases the model’s next-word choices slightly; across many words this forms a statistical pattern a detector holding the key can recognize. Readers see no visible mark.
- Persistence: because the signal is in the wording itself, it survives copy and paste.
- Privacy: OpenAI says the watermark does not identify the user, and it measured no meaningful change in model quality with watermarking on.
- Scope: EU-only for ChatGPT and Codex. API developers in any region can enable it for select models from October 5; it is off by default.
- Detection access: initially limited to approved researchers and expert organizations, not the public.
Limits OpenAI disclosed
- Editing weakens the signal: in OpenAI’s tests, replacing 10% of words with synonyms lowered detection from about 92% to 66%.
- Short passages, math answers and translated text are harder to detect.
- The absence of a watermark does not prove a human wrote the text; it may be short, heavily edited or produced by another provider’s model.
Why it matters
This is one of the first large-scale deployments of text watermarking by a major chatbot provider, and it shows how the AI Act’s transparency rules are being implemented in practice. According to TechCrunch, OpenAI built a text watermark in 2024 but held it back, and Anthropic announced worldwide watermarking of Claude text in August 2026.
For businesses and educators, the practical point is restraint: a watermark is a probabilistic signal, detection is not publicly available, and the method is weakened by editing. It should not be treated as proof of authorship in disciplinary or legal decisions. For developers serving EU users, the opt-in API setting is a way to support compliance; check whether your use case falls under the Act’s marking obligations.
What to do
- If you deploy generative text to EU users through the API, review the AI Act transparency guidance and test the watermark option on supported models.
- Do not build workflows that rely on public watermark detection; access is restricted for now.