OpenAI set out its approach to text provenance under the EU AI Act on 5 October 2026, beginning with an option for API customers worldwide to apply text watermarking to select models.
Over the coming weeks, the company plans to place an invisible watermark on eligible text produced by ChatGPT and Codex in the EU. The system, called textGrain, builds a statistical signal into the model’s choices of words.
OpenAI is introducing the technology to meet requirements under the EU AI Act. Businesses using its models could also use the watermark to meet marking duties of their own under the law.
The feature will stay switched off by default for the API, and OpenAI does not intend to make it the global default at launch. Anthropic introduced text watermarking before OpenAI.
There are clear boundaries to what the mark can establish. A detector may find evidence that an OpenAI system generated or processed part of a text, yet it cannot identify the person who used the system, measure a person’s contribution, establish ownership or assess whether the material is accurate.
OpenAI described the technology as having “significant limitations”. Detection is particularly unreliable for short passages and text that has later been edited, while a missing signal does not prove that a person wrote the material.
That weakness reflects a broader problem with provenance tools: no individual method is perfect, and signals can be removed or lost through editing or changes between platforms.
For now, access to the detector will be confined to approved researchers and expert organisations. Organisations with a qualifying use case can apply in line with OpenAI’s commitments under the Code of Practice on the transparency of AI generated content.
OpenAI also publishes summaries of material used to train general-purpose AI models, in accordance with Article 53(1)(d) of the EU AI Act. Article 5 of the act prohibits certain AI uses that carry an unacceptable level of risk.
Comments (0)