OpenAI has announced it will begin automatically watermarking text generated with ChatGPT for users in the European Union. While the feature will be available in other regions, it will remain disabled by default outside the EU.
Compliance with the EU AI Act
The move is a direct response to the EU AI Act, which took effect in August. The regulation mandates that AI-generated content must be marked in a way that specialized tools can detect. However, the industry currently lacks a completely effective or foolproof standard for such identification.
The 'textGrain' Method and Its Limitations
OpenAI’s proprietary watermarking method, dubbed textGrain, functions by embedding specific patterns into word choices. These patterns are invisible to human readers and do not significantly alter the quality of the output, but they can be identified by those with access to a specialized detector key.
Technical data suggests the system is far from invincible:
- OpenAI reported a 92% success rate in controlled tests, but this efficacy drops sharply with minor edits.
- Changing just 10% of the generated text reduces detection success by nearly 30%.
- Altering 20% of the text can lower the success rate by almost 75%.
- Detection is notably less reliable for shorter snippets or translated text.
Access to the detector tool will initially be restricted to a limited group of researchers and organizations, with others required to undergo an approval process.
Global Fragmentation and Rollout
The strategy highlights a divergence in the industry. While competitor Anthropic enabled watermarking globally in August, OpenAI is restricting the default setting to regions with active regulatory mandates. The feature will be the default in ChatGPT and Codex apps within the EU, while remaining optional for API users. The rollout is expected to reach EU users "in the coming weeks."
