OpenAI to Enable Automatic Watermarking of ChatGPT Outputs by Default in EU
Image: Ars Technica · Source
OpenAI will begin automatically watermarking ChatGPT-generated text in the European Union to comply with the EU AI Act, using a proprietary watermarking method called textGrain. The feature will be optional elsewhere.
OpenAI announced plans to automatically watermark text generated by ChatGPT within the European Union (EU), a move designed to comply with the recently enacted EU AI Act that mandates marking AI-produced content for detection by specialized tools. This watermarking system, named textGrain, subtly alters word choices in generated content to embed identifiable patterns—imperceptible to human readers but detectable by tools possessing a specific key.
The company’s approach aligns with ongoing efforts in the AI industry to provide traceability of machine-generated text. However, OpenAI acknowledges the limitations of textGrain, noting a best-case detection rate of approximately 92 percent under controlled tests. Detection reliability diminishes notably when slight text modifications occur; changing just 10 percent of the text reduces successful detection by nearly 30 percent, while a 20 percent change drops accuracy by up to 75 percent. Shorter or translated texts also exhibit lower detection rates.
Unlike OpenAI’s regional mandate, competitor Anthropic has implemented watermarking globally for its AI models. Outside the EU, OpenAI will keep watermarking off by default, though it will offer the feature as an option in its API. The watermark detector will initially be accessible only to select researchers and organizations, with plans to expand access through a pending approval process.
OpenAI plans to roll out this watermarking feature to EU users “in the coming weeks,” introducing the capability primarily through ChatGPT and Codex applications. The move underscores increasing regulatory influence on AI transparency and the challenges faced in creating robust identification mechanisms for AI-generated content.

Sources and original reporting
Read the original source ↗

Comments (0)
No comments yet. Start the discussion.
Write a comment
Comments are published after moderation. Your name and comment will be visible publicly. Account