AI-rewritten: This is a summary of an article from Ars Technica, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

OpenAI has announced that text generated by its ChatGPT models will be automatically watermarked within the European Union starting in the coming weeks. This decision is primarily intended to comply with the EU AI Act, which mandates marking content produced by artificial intelligence so it can be detected by other tools. While OpenAI plans to offer this feature globally, it will remain disabled by default outside of Europe unless users manually enable it.

The company uses a proprietary method called textGrain to embed patterns into word choices that are invisible to human readers but detectable by specialized software with the correct key. Access to these detectors is currently limited to a select group of researchers and organizations, with a request-for-approval process available for others. Although OpenAI reports a 92 percent detection success rate in its tests, the method is not perfectly reliable. The company notes that altering just 10 percent of the text can drop the detection rate by nearly 30 percent, while changing 20 percent can reduce it by almost 75 percent.

Competitor Anthropic introduced similar watermarking for its models in August but enabled it globally rather than restricting it to specific regions. OpenAI’s implementation will apply to both the ChatGPT and Codex apps within the EU. The company acknowledges that while textGrain is effective, its success rate tends to be lower for shorter texts or translated content compared to longer original outputs.