Secret stamp: OpenAI adds invisible watermarks to ChatGPT text under EU rules

OpenAI will begin adding invisible watermarks to text generated by ChatGPT and Codex in the European Union. The rollout addresses transparency requirements under the EU AI Act.

The system embeds a hidden statistical signal into model-generated writing. Readers will not see the watermark, but OpenAI’s detector can analyze text for evidence of that signal.

The technology arrives as major AI companies face growing pressure to identify synthetic content. Anthropic announced a worldwide watermarking system for Claude text two months ago.

Some Claude users pushed back against that move. They argued that users provide the instructions, context and decisions while AI remains a tool.

OpenAI had also developed text watermarking before, but delayed its release. The company reportedly worried that users could move to competing services without similar restrictions.

OpenAI targets EU users

OpenAI calls its system textGrain. It works by subtly changing statistical patterns in the words selected during generation. The company expects the watermark to reach eligible ChatGPT and Codex users across all EU plans over the coming weeks.

API customers worldwide will get a separate option. They can enable watermarked output for selected models, while the default setting remains off.

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.

Our tools already help verify whether an image or audio file was created with our…

— OpenAI (@OpenAI) October 5, 2026

OpenAI has also opened applications for its text watermark detector. Researchers and expert organizations will receive access initially.

The cautious rollout reflects weaknesses in current detection technology. OpenAI says its detector can generate false positives and miss genuine watermarks.

Editing weakens detection

OpenAI’s tests show that longer passages give the detector a better chance of finding the signal. It detected watermarks in about 80% of 200-token psychology passages.

That figure reached roughly 95% for 400-token passages. Results dropped significantly for mathematics, where writers have fewer choices in wording.

Human editing creates an even bigger challenge. Replacing 10% of words with synonyms reduced detection from about 92% to 66%.

Replacing 25% of words pushed detection down to roughly 17%. Translation and extensive rewriting can also weaken the signal.

That means a watermark cannot establish that AI produced an entire passage. It also cannot measure how much human work went into the final text.

A positive detection does not identify the user, account or conversation behind the writing. It also says nothing about accuracy, ownership or responsibility.

An absent watermark proves little as well. Short text, editing, translation and unsupported models can all produce undetected content.

Provenance takes multiple forms

OpenAI plans to release textGrain as open-source technology and continue testing its performance. The company also wants to study how translation and editing affect watermark durability.

The effort fits into a broader industry push around AI provenance. Anthropic, Google, Meta, Microsoft and OpenAI have committed to following the EU’s code of practice for AI-generated content.

OpenAI already uses Content Credentials and invisible watermarks for supported images and audio. Text presents a harder engineering problem because people can rewrite words without changing the underlying meaning.

The company says it will expand detector access as reliability improves. It will also adjust its approach as technical evidence and regulatory standards develop.