OpenAI will begin automatically watermarking text generated by ChatGPT in the European Union to comply with the EU AI Act, which mandates detectable markings on AI-generated content. Outside the EU, the feature will remain disabled by default but available as an option, including in its API. The technology, dubbed textGrain, embeds hidden statistical patterns into output text that can be detected using specialized keys.
According to OpenAI, textGrain achieves a 92% detection rate in internal testing under controlled conditions. However, performance degrades rapidly when text is modified: altering just 10% of an output reduces detection success by nearly 30%, while a 20% change reduces it by almost 75%. Shorter and translated texts also yield lower accuracy.
The rollout comes as competitors like Anthropic adopt global watermarking strategies. OpenAI plans to grant limited access to its detection tool to select researchers and organizations via an approval process as the system rolls out in Europe in the coming weeks.
Why it matters
EU AI Act compliance mandates watermarking features, creating regional product fragmentation for AI developers.
Current LLM watermarking solutions remain vulnerable to minor text editing, limiting their effectiveness against malicious misuse.
Enterprise API users retain optionality, preserving raw output flexibility for downstream application integration.
Source: arstechnica.com



