OpenAI Introduces Text Watermarking for EU Compliance and Global API Opt-In

OpenAI has introduced a phased approach to text watermarking to comply with the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. The company emphasizes that text watermarking remains an early technology with significant limitations, prioritizing transparency about what the signals can and cannot confirm.

Phased Rollout Strategy

The implementation follows a three-part plan designed to balance regulatory requirements with technological constraints:

• API Customers: Starting today, API customers globally can opt in to text watermarking for select models. This feature remains off by default.
• EU Users: Over the coming weeks, an invisible watermark will be added to eligible ChatGPT and Codex text output for users in the European Union.
• Detector Access: Applications are now open for approved researchers and expert organizations to access the text watermark detector. Public access is restricted to prevent misinterpretation of results.

Technology and Performance Limitations

OpenAI’s watermarking technology, textGrain, adds an invisible statistical signal to model word choices. While evaluations show textGrain matched or exceeded other approaches like SynthID for text, real-world reliability varies significantly based on text length and editing.

• Text Length Impact: At a 1% false positive rate, detection identified watermarks in about 80% of 200-token passages versus 95% of 400-token passages for general content. Detection rates were substantially lower for mathematics, where word choice flexibility is limited.
• Editing Vulnerability: Replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words dropped detection to 17%.

Impact on Model Quality

Benchmarks for Astra, OpenAI’s latest frontier model, show no meaningful performance differences between watermarked and unwatermarked outputs. Key metrics remained stable across tests such as the Artificial Analysis Intelligence Index (49.57 vs 49.76 points) and GPQA Diamond (94.44% vs 93.94%).

Scope and Interpretation

OpenAI clarifies that a text watermark provides a limited signal and does not establish ownership, verify accuracy, or identify the user. The absence of a detected watermark does not prove human authorship, as text may be too short, edited, or generated by unsupported models. The company plans to open-source the technology and revisit its approach as standards and evidence evolve.