“In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act,” OpenAI said.
The rollout will cover eligible ChatGPT and Codex users across all plans in the EU. OpenAI is not enabling the technology globally by default, saying its initial regional deployment will allow the company to study how the system performs in real-world use.
API customers outside the EU will have a different option. OpenAI is allowing developers worldwide to turn on watermarking for select models, but the setting will remain disabled unless they choose to enable it.
OpenAI says textGrain is designed to make AI-generated text identifiable without attaching information about the person who created it. A detected watermark does not reveal a user, account, prompt or conversation.
The company is also cautioning against treating the watermark as definitive proof about authorship. A positive result only indicates that an OpenAI system may have generated or processed some portion of the passage.
“ can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company said.
Detection also becomes less reliable under certain conditions. Short passages provide fewer opportunities for the model to create a detectable statistical pattern, while highly constrained material such as mathematics offers less flexibility in word choice.
Editing can further weaken the signal. In one OpenAI evaluation involving 400-token passages, replacing 10% of the words with synonyms reduced the detection rate from about 92% to 66%. Replacing 25% cut it to 17%.
That means the absence of a watermark cannot establish that text was written by a person. Content may have been edited, translated, generated by an unsupported model or produced before watermarking was enabled.
OpenAI is also limiting access to its detector at launch. Approved researchers and expert organizations can apply to use the tool, but the company is not releasing it publicly because of the risk of both missed detections and false positives.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” OpenAI said.
The company says watermarking does not meaningfully reduce model quality in its internal testing. Across benchmarks used to evaluate Astra, its latest frontier model, OpenAI reported similar results with and without the watermark enabled.
OpenAI plans to publish more technical information about textGrain and says it eventually intends to make the technology available as open source.
The company’s approach differs from Anthropic’s decision to apply text watermarking globally. OpenAI is instead keeping the default rollout limited to the EU while it gathers feedback and studies how the system behaves outside controlled evaluations.
“This regional approach gives us room to learn from real-world use and feedback,” OpenAI said.
Text watermarking is one part of OpenAI’s broader content-provenance work. The company already uses Content Credentials and invisible watermarks for supported image and audio outputs, while offering verification tools for those media formats.
Text presents a different challenge because it can be easily rewritten, shortened or translated. OpenAI says it expects to continue adjusting its approach as detection technology, regulatory requirements and standards develop.
This analysis is based on reporting from PC Mag & OpenAI.
Image courtesy of The Verge.
This article was generated with AI assistance and reviewed for accuracy and quality.
