Images generated through ChatGPT, Codex, or the OpenAI API carry Content Credentials and SynthID watermarks; ChatGPT and API audio carries SynthID. Sora is discontinued, while its video outputs are described as using Content Credentials.
For text, OpenAI says ChatGPT outputs in the EU include its invisible textGrain watermark. The system changes randomness in word or word-piece selection rather than inserting hidden characters or unusual punctuation. API customers worldwide can enable text watermarking from organization or project settings for selected models. OpenAI reports detection rates from 69.0% for Spanish to 42.2% for Romanian at a 1% false-positive rate across the 24 official EU languages.
The company says text detector access is limited to qualifying organizations, including research uses; academic-integrity checks of student work are outside the program scope. Image and audio checks are available through openai.com/verify and the Content Provenance API, with free but limited checks. The source does not establish that a provenance signal proves accuracy, legal ownership, unedited status, or correct context. It also says substantial paraphrasing or translation can make a text watermark undetectable.