OpenAI starts opt-in text watermarking before EU expansion

OpenAI began letting API customers worldwide opt in to text watermarking for select models on October 5, with the feature off by default. The company said eligible ChatGPT and Codex text output in the European Union will gain an invisible watermark across all plans over the coming weeks.

The first phase gives developers an optional way to add a machine-detectable provenance signal to model output. OpenAI said it will not make watermarking the global default at launch. Support for output generated through unnamed cloud partners is also planned for the coming weeks, according to the company.

The rollout follows the EU’s Article 50 transparency requirements. The European Commission says those obligations have applied since August 2, 2026 and require providers to mark AI-generated or manipulated output, including text, in a machine-readable way when technically feasible and appropriate for the content. The Commission describes participation in its transparency code of practice as voluntary, but the underlying Article 50 requirements as legal obligations.

OpenAI calls its system textGrain. It embeds a statistical signal through the model’s token choices, so the visible text does not carry a label or obvious marker. In the company’s technical description of textGrain, generation ties token selection to keyed pseudorandomness and detection uses the secret key to reconstruct statistical scores from a passage.

Applications for OpenAI’s detector also opened October 5, but initial access is limited case by case to approved researchers and expert organizations. OpenAI said it is withholding public access at launch because the detector can produce false positives and miss watermarks. According to the company, the detector reports whether it finds an OpenAI watermark without identifying a user or revealing prompts or conversations.

OpenAI also set out boundaries for interpreting a result. A detected watermark can indicate that an OpenAI system generated or processed part of a passage, but it cannot measure the amount of human judgment, editing or creativity involved. It does not establish ownership, responsibility, lawful use or whether disclosure was required. It also cannot determine whether text is accurate, misleading, harmful or presented in the proper context, and it does not link the text to a person, organization, account, prompt or conversation.

The absence of a detected watermark is not proof that a person wrote the text. OpenAI said detection may fail when text is short, edited or translated, or when it came from an unsupported model, predates the rollout or was produced with another company’s tools.

Performance also varies by length and subject. At a target 1% false-positive rate, OpenAI said its detector found watermarks in about 80% of 200-token psychology passages and about 95% of 400-token psychology passages, with substantially lower detection for mathematics content. In another company evaluation using 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. The opened sources did not provide an independent replication of the deployment configuration.

OpenAI said it plans to release textGrain as open source, but gave no date. It also did not identify the API models supported at launch, define which ChatGPT and Codex output is eligible, name participating cloud partners or give exact deployment dates within the coming-weeks window.