OpenAI Rolls Out Invisible Text Watermark in the EU: ChatGPT Output Gets Tagged

What Happened

In a blog post on Monday, October 5, OpenAI announced it will add invisible watermarks to text generated by ChatGPT and Codex in the EU. The watermark will roll out over the coming weeks to ChatGPT and Codex users in the EU across all paid tiers; developers using the API can manually enable it globally for designated models — off by default. OpenAI was explicit: this is not becoming a global default.

The direct driver is Article 50 of the EU AI Act. The transparency rule took effect on August 2, 2026, requiring AI companies to mark AI-generated content in machine-readable form. The watermark is OpenAI’s compliance answer.

How the Watermark Works

The technique is called textGrain. OpenAI simultaneously published a technical report co-authored with researchers from the University of Pennsylvania and Yale.

The principle is simple: during generation, the watermark uses a secret key to nudge the ranking of candidate words. Each word’s shift is tiny — nothing reads as off; hundreds of tiny shifts stacked together form a statistical pattern. A detector holding the text and the key can measure that pattern.

Because the watermark lives in word choice, there are no hidden characters, no invisible spaces. Readers see nothing, and there’s nothing to delete. It survives copy-paste.

Detection Rates and Limits

Per OpenAI’s published data: at a 1% false-positive rate, detection hits about 80% on 200-token passages and about 95% on 400-token passages. (A token is roughly a word or word-chunk.) Those numbers were measured on psychology-style content; math answers score much lower because there’s less room for word choice.

Paraphrasing degrades the watermark significantly. In 400-token tests, swapping 10% of words for synonyms dropped detection from ~92% to 66%; swapping 25% dropped it to ~17%. Short texts and translated content are also harder to detect.

OpenAI says the watermark doesn’t affect model quality. In tests on its Astra model, the Artificial Analysis Intelligence Index scored 49.57 without and 49.76 with the watermark; GPQA Diamond 94.44% vs 93.94% — small gaps.

The Detector Isn’t Public

Adding the watermark is one thing; who gets to check is another. Detectors go first to approved researchers and expert organizations by application. OpenAI says it wants to assess detection reliability and responsible use before discussing wider release. Some outlets mention plans to eventually open-source the technology; no timeline was given this time.

OpenAI also stresses: the watermark doesn’t identify users, and it can’t see prompts or chat history. Nor does it solve attribution or copyright — it can only answer “this text may carry an OpenAI-generation signal.” It can’t prove who wrote it or whether the content is true. No watermark detected doesn’t mean a human wrote it either.

Why Only the EU

Only the EU currently exerts this kind of compliance pressure. Article 50 targets systems already on the market; per reports, existing AI systems must comply by December 2, 2026. OpenAI has also signed the EU’s AI code of practice, committing to cooperate on transparency rules.

It’s not the only one moving. Anthropic added a similar watermark to Claude back in August 2026. OpenAI reportedly built a text watermark as early as 2024 and chose not to ship it; two years later, the EU brought it out anyway.

What It Means for Users

If you use ChatGPT in the EU, text you paste into emails, reports, and websites will henceforth carry a machine-detectable trace of AI generation — invisible to human eyes. OpenAI says the watermark won’t identify users and text quality won’t change.

Outside the EU, the watermark stays off by default for now. OpenAI’s verification tools (openai.com/verify and the Content Provenance API) remain publicly available for image and audio provenance.

AI-generated text now has a detectable fingerprint — but it’s a transparency tool, not a verdict. A detection result can only say “this text was likely AI-generated.” Whether it’s accurate or who owns it, the watermark can’t answer a word of that.