OpenAI will automatically watermark text generated by ChatGPT in the European Union starting today, while the feature remains disabled by default elsewhere. This change satisfies the requirements of the EU AI Act, which mandates that artificial intelligence systems mark their content so detection tools can identify it. The regulation came into force in August, forcing providers to adopt a standard that currently lacks universal reliability.
The company describes its method as textGrain, a proprietary technique that embeds subtle patterns into word choices without altering the overall quality for human readers. A specialized detector, accessible only to approved researchers and organisations via a request process, is required to find these markers. The approach mirrors previous watermarking attempts, relying on statistical anomalies rather than visible tags. Critics note that determined users with basic technical knowledge can still bypass such systems, meaning the watermark serves primarily as a compliance checkbox rather than a foolproof shield against misuse or deepfakes.
- Watermarking is mandatory only within the European Union
- The detection key is restricted to approved researchers
- Existing standards like SynthID remain vulnerable to circumvention
