ChatGPT hides a mark in text written in the EU but not in Indonesia

The short version

• OpenAI now adds an invisible watermark to ChatGPT text for users in the European Union only. It chose not to turn the mark on worldwide.
• Companies anywhere that use OpenAI's paid developer access can switch the mark on. It stays off unless they do.
• Pictures work differently. OpenAI says images made with ChatGPT carry 2 kinds of mark, and anyone can upload a file to check it.

What OpenAI changed

OpenAI announced its plan on October 5. The EU AI Act requires AI providers to make generated text identifiable by machine. OpenAI's help page now says ChatGPT text in the EU carries the mark.

The technology is called textGrain. It nudges how the model chooses between possible words as it writes. Across a passage, those choices form a pattern a detector can look for. No hidden characters are added, so a reader sees nothing unusual.

Developers get a choice. Those using OpenAI's paid access can turn the mark on in their project or organization settings, wherever they are. OpenAI says this lets each customer decide how the mark fits its own legal duties.

Anthropic, which makes Claude, went the other way. The Decoder, a trade publication, reports that Claude's text watermark applies worldwide, whichever way people use the model.

Why the mark is easy to lose

OpenAI published its own test results, and they show real limits. OpenAI set the detector to raise a false alarm about 1% of the time. At that setting, it found the mark in about 95% of psychology passages of roughly 300 words. At half that length, it found about 80%. Results were much weaker for math, where the model has fewer word choices.

Light editing wears the mark away. In OpenAI's editing test, replacing 10% of the words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%.

OpenAI says the mark made no meaningful difference to the model's performance across 8 standard tests. The Decoder points out that those tests don't measure writing quality.

What the mark can't tell you

OpenAI is plain about this. A detected mark doesn't show how much human judgment went into a passage. Nor does it settle who owns the text or who is responsible for it. It also can't identify the user or show the text is accurate.

A missing mark proves nothing either. The text may be too short, or it may have been edited or translated. It may also come from another company's tool.

You also can't run the text check yourself. OpenAI is giving detector access to approved researchers and expert groups, case by case. Its help page says general business uses are outside the program for now, and so is checking student essays. The Decoder notes that OpenAI hasn't said when access will widen.

What this means for a business in Indonesia

By OpenAI's description, text your team writes in ChatGPT from Jakarta isn't marked. That changes once work crosses a border. A colleague or agency partner in the EU who drafts in ChatGPT is producing marked text, and that text can end up in your documents.

The rules also differ by tool. If your teams use both ChatGPT and Claude, one company policy has to cover 2 different setups. So don't build a policy that depends on whether a given draft carries a mark.

Pictures need more care than text. OpenAI's help page says images made with ChatGPT or its developer access carry both a signed label and a hidden watermark. It lists no regional limit. Anyone can upload a file at openai.com/verify. So a journalist or a rival could test a campaign image.

Those marks can still fall away. OpenAI says a screenshot or a file conversion can remove the label, and heavy cropping or compression can defeat the watermark. Which marks a file carries also varies by product and file type. As with text, a check that finds no mark on a picture proves little.

What to do next

• Write your own rule on when your company discloses AI use. Hidden marks are too patchy to do that job.
• Ask EU offices and suppliers whether they use ChatGPT on your work, and whether your contracts allow it.
• Keep a record of which campaign images were made with AI, and never claim otherwise in public.
• If you build products on OpenAI's models and have EU users, ask your lawyers whether to switch the mark on.
• Don't treat a detector result as proof, in either direction.