OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU -

Quick answer: OpenAI said on 5 October 2026 that it will add an invisible watermark to ChatGPT and Codex text for users in the European Union over the coming weeks. You won’t see it, you can’t switch it off in ChatGPT, and it doesn’t change how the text reads. Anthropic has already started marking Claude output worldwide.

The change comes from the EU’s AI Act, which began requiring AI providers to mark synthetic content in a machine-readable way from 2 August 2026. If you live in the US, UK, Canada or Australia, the OpenAI rollout starts in Europe first, but the wider trend affects everyone who copies AI text into emails, essays or reports.

Here’s what is actually changing, what a watermark can and can’t prove, and what to do if you use AI to help with writing. More on the wider AI and tech picture is in our Tech News section.

What is OpenAI doing with ChatGPT text in the EU?

OpenAI says it will add an “invisible watermark to eligible ChatGPT and Codex text output in the European Union” over the coming weeks, according to BleepingComputer. The technology is called textGrain. It nudges the model’s word choices in a secret statistical pattern that people can’t spot but a detector can pick up.

TechCrunch reports the watermark will apply across all ChatGPT plans in the EU, with no separate cost, and EU users can’t opt out. API developers anywhere in the world can already choose to turn it on for select models, but it’s off by default.

Detection isn’t public. OpenAI says access to its detector is limited for now to approved researchers and organizations, which have to apply.

What has Anthropic done with Claude?

Anthropic moved first. IEEE Spectrum reports that on 11 August 2026 the company announced that its models would carry watermarks identifying text as AI-generated. Implicator reports that this applies to Claude models launched after 2 August 2026 and worldwide, not only in the EU, across the Claude app, the API and cloud partners.

Images are handled differently. Supported image files (SVG, PNG and JPG) reportedly get signed provenance metadata, similar to the labels already used by some camera makers and image tools. Anthropic hasn’t published full technical details or a public detection tool, according to the same report.

Treat the details as still moving. Several of these reports rely on company statements rather than independent tests, and OpenAI and Anthropic haven’t opened their detectors to the public.

Why is this happening now?

Article 50 of the EU AI Act says providers of generative AI must mark synthetic content “in a recognisable and machine-readable manner,” as summarized by law firm Houthoff. The transparency obligations came into force on 2 August 2026. There’s a transition period until 2 December 2026 for systems that were already on the market before 2 August.

The penalties are meaningful. IBTimes UK reports fines for non-compliance can reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. That explains why companies are acting even though watermarking text is technically imperfect.

Can a watermark prove something was written by AI?

No, and the companies say so themselves. OpenAI states that the absence of a detected watermark does not prove human authorship. It also can’t tell who wrote the text, what the prompt was, or how much a person edited it.

The weaknesses are practical:

Light editing hurts detection. OpenAI’s own testing found that swapping 10 percent of words for synonyms cut detection from about 92 percent to 66 percent. At 25 percent, it fell to 17 percent.
Short text is hard. A one-line reply gives a detector very little to work with. Math and translated text are also harder.
Code is different. Reports on Anthropic’s approach say exact code output is excluded, though code comments can be marked.
Images can be stripped. Provenance metadata can disappear when you screenshot, convert or re-save a file.

So think of a watermark as one clue that a detector may use, not a verdict. It’s a poor basis for accusing a student or an employee of cheating.

What does it mean for you?

If you use ChatGPT or Claude, your day-to-day experience shouldn’t change. The text reads the same, and copy and paste still works. Nothing about your account or identity is embedded in the watermark, according to OpenAI.

A few practical points:

Students: don’t assume a school or university can detect AI text from a watermark, and don’t assume it can’t. Follow your institution’s AI policy and disclose use when it asks.
Workers and small-business owners: if you publish AI-assisted copy, edit it properly and check the facts. That’s good practice whether or not a mark is present.
Privacy-minded users: the mark doesn’t reveal your prompts or account. If you want control over what AI companies do with your chats, that’s a separate setting. Our separate guide on stopping AI chatbots from training on your chats covers that.
Outside the EU: ChatGPT users in the US, UK, Canada and Australia aren’t named in OpenAI’s rollout yet. Watch for updates, because Anthropic has already gone global.

Data leaks are a different kind of AI-adjacent worry. If you’re unsure whether your own details have been exposed anywhere, see our guide on how to check whether you were in a data breach.

What should schools and employers take from this?

Watermarks may help platforms and researchers study how much AI text is in circulation. They are a weak tool for judging one person’s work. Short answers, edited drafts and translated text can all slip past a detector, and OpenAI’s own numbers show how quickly detection falls.

A fairer approach is to ask for drafts, notes or a short conversation about the work. If your school, university or employer has an AI policy, that policy matters more than any hidden mark.

Frequently asked questions

Will the watermark change how ChatGPT text looks?

No. It shifts word choices very slightly in a hidden pattern, so the text reads normally. You can’t see it in a document, an email or a web page.

Can I remove an AI watermark?

Heavy rewording, paraphrasing tools or translation can weaken or break a text watermark, and the companies admit it. We’d still encourage you to be open about AI use where rules require it, rather than trying to hide it.

Does the EU rule apply if I live in the UK, US, Canada or Australia?

The law applies to providers serving the EU. OpenAI’s ChatGPT rollout is EU-only for now, while Anthropic reportedly applies its marking worldwide. Other providers may follow whichever path is simplest for them.

Does this mean AI detectors are now reliable?

No. Watermark detection only works for text from a system that embedded a mark, and only if it hasn’t been heavily edited. A clean result proves nothing either way.

OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU