OpenAI Is Watermarking ChatGPT Text in the EU. Here's What textGrain Can and Can't Prove.

OpenAI has begun watermarking AI-generated text. A technology it calls textGrain is rolling out to ChatGPT and Codex users in the EU and is available as an opt-in for developers worldwide, a response to European rules that require generated text to be identifiable by machines. The detector, though, is not public, and the watermark is easier to weaken than the announcement's headline suggests.

What OpenAI Announced

In a post dated Oct. 5, 2026, OpenAI says "the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way," and that textGrain is its answer. The page does not cite a specific article of the Act. The technology adds a statistical signal, invisible to readers, to the model's word choices. OpenAI's detector looks for that signal and reports whether an OpenAI watermark is detected.

• Developers (worldwide): API customers can opt in for select models starting Oct. 5. "Text watermarking will remain off by default in the API," OpenAI says.
• ChatGPT and Codex (EU only): an invisible watermark will be added to eligible text output for users on all plans in the EU, over the coming weeks. It is not a global default.
• Cloud partners: OpenAI says it is working with them to offer watermarking for OpenAI model outputs accessed through their services in the coming weeks.
• Detection: approved researchers and expert organizations can apply for access starting Oct. 5, case by case, under the EU Code of Practice on AI-generated content. The detector is not public at launch.
• Open source: OpenAI plans to release the technology so others can build on it, with no date or license given, and says a technical report will be updated with more detail in the coming weeks.

How Well Does It Work?

OpenAI's own numbers show the limits. At a 1% false positive rate, on psychology content, the detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages. Detection was substantially lower for mathematics, where word choice is more constrained. Editing hurts: for 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI reports it matched or exceeded other methods it tested, including SynthID for text, though these are the company's own evaluations.

The company also says watermarking does not meaningfully change model quality. On the benchmarks it uses for Astra, its latest frontier model, GPQA Diamond scored 94.44% unwatermarked and 93.94% watermarked.

What a Watermark Does Not Tell You

OpenAI is explicit about the limits. A watermark does not measure how much human contribution a text involved, and it does not establish ownership, lawfulness, disclosure requirements, or responsibility. It does not identify the user, account, prompt, or conversation, and it does not verify accuracy. Critically, a missing watermark does not prove human authorship: short, edited, translated, or unsupported-model text may go undetected. The detector is designed not to reveal users or their conversations.

Images and audio are handled separately. OpenAI says they carry Content Credentials (C2PA) and invisible SynthID watermarks, with public verification through openai.com/verify and the Content Provenance API.

What It Means If You Publish With AI Tools

This section is practical analysis, not reporting. For publishers, marketers, and teams that use AI to draft or edit copy, the announcement raises a few concrete questions:

• Which tool and channel produced the text? The watermark is added by OpenAI's systems, only where it is enabled: EU ChatGPT and Codex users, and API customers who opt in. Text from other providers is not covered.
• Editing is not a switch-off, but it weakens the signal. Heavy rewriting lowers the detection rate, per OpenAI's own figures, so a detector result should never be treated as proof either way.
• Disclose on your own terms. Because a watermark is neither a disclosure nor proof of authorship, your own disclosure policy still does the real work. Our beginner's guide to using AI and power-user playbook cover verification habits that apply regardless.
• Watch the EU rules. Our EU AI Act compliance guide explains how provider obligations like this one fit into the wider law.

Why It Matters

• Text provenance is moving from research to product. One of the largest AI providers is now watermarking text by default for EU users.
• Detection is not public. Only approved researchers and organizations can check for the watermark at launch.
• It is a signal, not proof. OpenAI's own data shows detection falling sharply as text is edited.
• Regulation is driving product design. The EU requirement is the stated reason the feature exists, and the global opt-in shows how it may spread.