ChatGPT Now Adds Invisible Watermarks in Europe — What It Means for You

OpenAI is rolling out invisible watermarks for ChatGPT and Codex in the European Union. This is to comply with the EU AI Act, which requires transparency for AI-generated content. The watermark, called textGrain, subtly changes how the AI picks words, creating a pattern that detection software can spot. It doesn't add hidden characters, so copying and pasting won't remove it. However, editing, translating, or paraphrasing can weaken or erase the signal.

The watermark only works on longer text—at least 200 tokens, about 150 English words. Code snippets are excluded because there are fewer word choices, making watermarking unreliable. Detection is also limited: only approved researchers and expert organizations will have access to the detector, not the general public. A positive match doesn't identify the user or prove authorship, and a negative result doesn't guarantee the text is human-written.

For everyday users, this means your ChatGPT outputs in the EU may carry a hidden marker. If you use AI for work, school, or publishing, you should still keep your own records of how you used AI, as the watermark can't explain your process or compliance. The watermark is just one tool for identifying AI text, but it's not foolproof. Ordinary documentation remains more reliable for disclosing AI use.

OpenAI's move is part of a broader trend of tech companies experimenting with ways to label AI content. While the watermark may help researchers and organizations detect AI-generated text, it's not a silver bullet. Users should continue following their own disclosure policies and keep a trail of their AI usage.