Artificial Intelligence
OpenAI plans EU ChatGPT watermarks and optional API access
OpenAI said it would switch on invisible text watermarking for ChatGPT and Codex in the EU, while keeping API watermarking optional worldwide and limiting detector access at first.
- OpenAI said it would enable textGrain watermarking for ChatGPT and Codex users in the European Union over the coming weeks.
- The company said API watermarking would be optional worldwide and would also be offered through cloud partners such as Microsoft Azure.
- OpenAI said textGrain would be released as open source and that detector access would begin with selected researchers and specialist organisations.
- OpenAI said detection rates fell when passages were shorter or edited, with about 95 percent detection for one 400-token test and 17 percent after a quarter of the words were replaced.
OpenAI said it would turn on invisible watermarking for ChatGPT and Codex in the European Union over the coming weeks to meet EU requirements. It said API users around the world would be able to choose whether to use the same system, and that the API route would also run through partners such as Microsoft Azure.
The company described textGrain as a statistical signal built into word choice and compared it with Google’s SynthID system. OpenAI said its own testing put textGrain on a par with or ahead of other approaches, and it planned to release the technology as open source.
OpenAI said the detector’s performance varied with passage length and subject matter. With the false-positive target set at 1 percent, it said the detector found watermarks in about 95 percent of 400-token psychology passages and about 80 percent of 200-token passages; it said results were lower for maths content.
The company also said edits made detection much harder. Replacing 10 percent of the words with synonyms reduced detection to 66 percent in one 400-token test, while replacing a quarter of the words cut it to 17 percent, and OpenAI said that was one reason it would restrict detector access at first.
Named in this story
Companies
- OpenAI
- said it would add watermarking to its text products and API
- Anthropic
- was used as the comparison for mandatory watermarking
- was the company behind the related SynthID technology
Governments and agencies
- European Union
- was the jurisdiction where OpenAI said watermarking would be enabled
Products and systems
- ChatGPT
- would receive watermarking for users in the European Union
- Codex
- would receive watermarking for users in the European Union
- Claude
- was Anthropic's model with mandatory watermarking
- textGrain
- was OpenAI's invisible watermarking system
- SynthID
- was Google’s text watermarking system used in comparisons
- Microsoft Azure
- was named as a cloud partner for OpenAI API watermarking
- Astra
- was the frontier model OpenAI cited in output-quality tests
How the source tells it
The source read as a plain technical report with no distinct emotive framing.
No emotive framing was found in the original.