Technology and Science05 October, 20262 min read
OpenAI Automatically Watermarks ChatGPT Text In The European Union With textGrain
32outlets comparedWhere the coverage agrees
16 Other5 Western Alternative4 West Asian7 Other source typesThis article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
OpenAI will add invisible watermarks to ChatGPT and Codex text outputs in the EU. Rollout will be gradual over coming weeks; API users worldwide can opt-in watermarking.
Save coverage image
32 outlets in this story. Read their headlines below.
Outlets did split on US Loses Measles-Elimination Status After Pan American Health Organization Finds 12-Month Spread →
textGrain in the EU
OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, and OpenAI will also offer the watermarking feature in other regions but keep it off by default outside the EU. OpenAI’s method, called textGrain, works by putting patterns in word choices that are not clear to a human reader, and a detector with a key can use it to find the watermark.
OpenAI will provide access to the detector to a limited number of researchers and organizations, and OpenAI will use a request-for-approval process for others to be added over time. OpenAI will provide access to the detector to a limited number of researchers and organizations, and it will use a request-for-approval process for others to be added over time.
Errors and limits
OpenAI’s text watermark and its detection remain incipient technologies with notable limitations, and OpenAI warned that high performance in ideal conditions does not guarantee reliable detection in everyday use. Detectors can make two types of errors, they can indicate a watermark where there is none, a false positive, or miss a watermark that is present, a false negative.
Substituting 25% of the words with synonyms reduces detection from 92% to 17% in 400-token fragments, and edited texts can reduce detectability because editing drastically reduces detectability. These limitations contribute to our decision to provide initial access to the detector only to researchers and authorized expert organizations.
API rollout and what it means
OpenAI will deploy the textGrain system in the European Union over the coming weeks so that invisible watermarks can identify artificially generated content in both ChatGPT and Codex. The rollout will be phased, starting this Tuesday with global API users, and the feature will remain disabled by default in the API during this initial implementation phase.
The watermark “does not indicate who originally wrote the text or who owns it,” and the watermark “does not identify a person, organization, account, conversation, or message.” OpenAI will allow API customers worldwide to opt to enable the watermark in certain models, while OpenAI will add the watermark automatically for EU users on both platforms in the coming weeks.