OpenAI Rolls Out textGrain Text Watermarking in Phases for EU Rules!

Hi everyone, it's Shii-chan! Today I'm sharing an announcement about watermarks for AI-generated text!

OpenAI News openai.com

What was announced?

OpenAI News shared its approach to the EU's text provenance rules. The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. Because text watermarking has real limitations, OpenAI is taking a phased approach.

• API: starting today, customers worldwide can opt in for select models. It stays off by default
• ChatGPT and Codex: over the coming weeks, an invisible watermark is added to eligible output in the EU
• Detector: applications are open, with access initially limited to approved researchers and expert organizations

Why it matters

Verification tools for images and audio are already public. This is the text version, and what stands out is how carefully it explains what a watermark can and cannot tell you. OpenAI itself says the technology is early and views on its benefits and responsible uses are still developing.

What changes

• API customers: you decide how watermarking fits your transparency obligations and experiences. OpenAI also plans to make it available for model outputs accessed through cloud partners in the coming weeks
• ChatGPT and Codex users in the EU: eligible output on all plans gets watermarked. It is not a global default at launch
• Researchers and expert organizations: you can apply for detector access

Dive Deep

The technology is called textGrain. It adds an invisible statistical signal to the model's word choices, and the detector looks for that signal. OpenAI plans to update its technical report and release the technology as open source. In its evaluations, textGrain matched or exceeded other approaches, including SynthID for text.

There are limits, though:

• Shorter or more constrained text is harder to detect. At a 1% target false positive rate, the detector found watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Detection was substantially lower for content such as mathematics, where there is less flexibility in word choice
• Editing weakens the watermark. For 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% reduced it to 17%

On output quality, OpenAI says it saw no meaningful performance differences with and without watermarking on the benchmarks it uses for Astra, its latest frontier model.

OpenAI also lists what a watermark does not tell you:

• It does not measure human contribution
• It does not establish ownership or responsibility
• It does not identify the user
• It does not verify accuracy
• The absence of a watermark does not prove human authorship

The detector only reports whether it finds an OpenAI watermark, without identifying the user or revealing prompts or conversations. Because of the risk of missed watermarks and false positives, it isn't public at launch.

Wrap-up

• API opt-in starts today, off by default
• ChatGPT and Codex get watermarks in the EU over the coming weeks
• Detector access starts with approved researchers and expert organizations
• Detection gets weaker for short or edited text
• A watermark doesn't prove who wrote something or who is responsible for it

This is for anyone working on AI text identification or EU compliance!