ResearchOpenAI

OpenAI introduces invisible text watermarks

Original title: Our approach to EU text provenance rules

Plain language

What it is: OpenAI is introducing invisible text watermarking to identify text generated by its AI models, in response to the EU AI Act. They are also launching a tool to detect these watermarks.
How it works: The technology adds an invisible statistical signal to the words the AI chooses. A detector tool can then analyze a text passage to check if it contains this signal.
Where it is available: The watermark will be added to ChatGPT and Codex in the European Union over the coming weeks, and API users globally can opt in starting today. The detector tool is currently available only to approved researchers and expert organizations.
Example: An API customer can choose to turn on watermarking for their AI model outputs to show that the text was generated by OpenAI.

What it is: OpenAI announced textGrain, a text watermarking technology, and an accompanying detector to comply with EU AI Act machine-readable provenance requirements.
How it works: It embeds an invisible statistical signal into the model's token selection without significantly affecting output quality. The detector checks for this signal, though its reliability decreases with shorter texts, constrained topics like mathematics, or post-generation editing.
Where it is available: It is available as an opt-in for global API customers on select models, will roll out to eligible ChatGPT and Codex users in the EU, and the detector is restricted to approved researchers.
Example: The system was evaluated using psychology and mathematics questions from the ELI5 dataset to measure how text length and word choice flexibility affect detection rates.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

Our approach to EU text provenance rules