OpenAI Plans Invisible Watermarks for European Text to Follow Local Regulations

New European Union transparency requirements, set under Article 50 of the AI Act, demand that automated writing services offer machine-readable markers. Established companies face a December 2 deadline to follow these guidelines, while newer entrants must comply immediately.

OpenAI will soon begin placing unseen markers into writing and programming lines produced by ChatGPT and Codex inside the European Union.

Earlier this year, Anthropic introduced comparable steps. Other established firms such as Microsoft, Google, and Meta fall under the same regional obligations.

The tracking system, named textGrain, adjusts which words the software picks. These slight variations form a subtle statistical structure that special tools can uncover later.

A human eye cannot detect the difference while reading or duplicating the sentences, and OpenAI asserts that the output standards of its GPT-6 Astra model remain largely steady during testing.

Across the rest of the globe, the system will not activate on its own. Outside the European market, engineers building on OpenAI's interface can decide for themselves whether to turn the feature on for supported options.

OpenAI plans to publish the underlying technology under an open-source license so other programmers can adapt it.

The software designed to identify these hidden signatures will not be widely accessible right away. Instead, OpenAI is reviewing applications and handing early keys to selected research bodies and specialist groups.

The scanner cannot expose private account data, individual user identities, underlying prompts, or complete chat transcripts. It also cannot determine which sections a person rewrote.

The developer acknowledges that the process struggles outside controlled settings. Detection works best on longer answers, while brief excerpts frequently fail to show the stamp.

In tests covering 200-token psychology replies, the scanner discovered the marker about 80 percent of the time at a 1 percent false-positive setting, rising to roughly 95 percent on 400-token passages.

Mathematics problems show weaker results because the strict logic limits alternative phrasing, giving the software fewer opportunities to embed its pattern.

Regular editing also weakens the signature rapidly. In 400-token samples, swapping out one word in ten drops the discovery rate from about 92 percent down to 66 percent, and changing a quarter of the words sinks it to 17 percent.

Because translations, small trims, or brief passages can wipe the hidden data, OpenAI notes that an empty scan does not mean an article came from a human writer.