Claude and ChatGPT now hide a secret mark in the text they write

Claude, made by Anthropic, now quietly marks the text it writes. ChatGPT does the same, but only for people in the European Union. You cannot see the mark, and it stays intact when you copy and paste. Anthropic has marked text from new Claude models worldwide since August. OpenAI, the maker of ChatGPT, has no plans for a global rollout.

What changed

Both moves answer the EU AI Act, which requires machine-readable labeling of AI-generated content as of August 2, 2026. Anthropic's mark covers the Claude app, Claude Code, the API and Claude offered through AWS, Google Cloud and Microsoft. Older Claude models are scheduled to be updated. Google already uses a similar method called SynthID, which also covers images, audio and video.

OpenAI calls its method textGrain and announced it in early October. It automatically marks text from ChatGPT and Codex only for EU users. In the API, the feature works worldwide but is switched off by default.

How the mark works

A language model picks each word from several good options. A secret key nudges that choice slightly toward certain options. In one sentence, nothing stands out. Over a longer text, a statistical pattern builds up, and a detector with the right key can spot it. Anthropic says content, quality and readability do not change, and no hidden characters are added. Images and other files get signed C2PA provenance data, which a free checker can verify.

Why it is not proof

Both companies say the mark is not proof of anything. A missing mark does not mean a human wrote the text. It may be too short, heavily edited or written by another AI. A found mark only shows the model was involved, perhaps just to correct, translate or summarize.

• OpenAI's own tests show detection falling from about 92 to 66 percent if one in ten words is swapped.
• Waldemar Berchtold of Fraunhofer SIT says reliable detection needs about 200 to 300 words.
• Detectors give only a probability, which is hardly enough as forensic evidence.

Who it affects

Critics worry that students and job applicants could come under suspicion even when AI only fixed their grammar. They also note that tech-savvy users can rephrase text to shake the mark more easily than anyone else.

For most people, day-to-day life changes little. The one thing to know is that longer, unedited AI text may be traceable once verification tools become more widely available. The material does not say when that will happen.