The European Union’s AI Act, which became enforceable in August, obliges developers to label content generated by artificial-intelligence systems in a way that a separate tool can verify. To satisfy that mandate, OpenAI announced it will embed a watermark in ChatGPT and Codex outputs for users located in the EU, making the feature the default setting there.
OpenAI’s proprietary scheme, named textGrain, inserts subtle statistical patterns into the choice of words produced by the model. These patterns are invisible to human readers and do not noticeably alter the fluency or relevance of the text, but they can be identified by a specialized detector that possesses a secret key. The company released a technical paper outlining the algorithmic details while keeping the exact implementation confidential.
In internal testing OpenAI reported a 92 percent success rate for detecting its watermark under ideal conditions. However, the same tests showed that modifying just ten percent of the generated text cuts detection efficacy by roughly thirty percent, and altering twenty percent of the content drops the success figure by about seventy-five percent. The company also noted lower detection rates for shorter passages and for text that has been translated, mirroring trends observed in other watermarking approaches.
OpenAI plans to distribute the detection tool to a small pool of researchers and organizations initially, according to the company’s statements. Additional parties may gain access through a request-for-approval process that OpenAI intends to expand over time. This limited rollout reflects the current difficulty of providing a universally reliable verification method while still meeting regulatory expectations.
Existing labeling initiatives such as SynthID and the C2PA project also aim to embed traceable markers in AI-generated media, yet analysts note they can be bypassed by users with modest technical skill. Competitor Anthropic introduced a similar watermark across all its models worldwide, whereas OpenAI restricts its default activation to jurisdictions where the law mandates it. The watermark remains optional in OpenAI’s API, allowing developers to disable it if desired.
OpenAI indicated that the EU-wide watermark will be enabled for ChatGPT users within the next few weeks. The company has not specified an exact date, but the rollout will coincide with its broader compliance strategy for the AI Act. Users outside the European Economic Area will continue to see the feature turned off by default, though they may request its activation via the API.