The news
OpenAI announced a phased plan to add machine-readable text watermarking. The move responds directly to the EU AI Act requirement that generative AI providers make generated text identifiable. Watermarking starts as an opt-in feature for API customers worldwide on select models and will remain off by default. In the European Union, an invisible watermark will appear on eligible ChatGPT and Codex text output within weeks. The company is also opening applications for access to its detection tools.
Context
The EU AI Act sets a legal obligation for providers of generative AI systems to ensure output can be detected as machine-generated. OpenAI states that text watermarking and detection technologies are still early-stage and carry significant limitations. Views on the benefits and responsible uses of these tools continue to evolve. The company’s approach therefore combines compliance steps with explicit transparency about what any watermark can and cannot prove. This marks the first concrete implementation timeline OpenAI has shared for text provenance features.
Details
API customers globally gained the ability to opt in to text watermarking starting today. The feature applies only to select models and stays disabled unless a customer chooses to enable it. Within the European Union, OpenAI will begin embedding invisible watermarks into eligible ChatGPT and Codex text output over the coming weeks. The company simultaneously opened an application process for parties seeking access to its watermark detection capabilities. OpenAI’s announcement notes that the technology remains limited in scope and that detection accuracy is not guaranteed across all conditions or model outputs.
The phased schedule reflects both the legal deadline and the current technical constraints. Watermarking will not apply by default anywhere, including inside the EU. OpenAI emphasizes that the presence or absence of a watermark does not constitute definitive proof of origin. Instead, the marks serve as one signal among others when assessing whether text was produced by an AI system. No additional technical specifications on how the watermark is encoded or how robust it is against editing were released in the announcement.
Why it matters
Developers and organizations that rely on OpenAI’s API now face a practical choice about whether to embed provenance signals in their generated text. Those operating in the EU will see the change appear automatically on covered outputs, which may affect downstream workflows that assume unmarked text. The decision to keep watermarking off by default outside the EU limits immediate impact for most global users while still satisfying the regional legal requirement. Because the company acknowledges the technology’s early limitations, any detection results will require careful interpretation rather than automatic trust. Over time, the availability of an opt-in mechanism and a formal application process for detectors could shape how enterprises handle content provenance in regulated environments. The narrow scope of the initial rollout also signals that broader adoption will depend on further technical progress and clearer standards for what constitutes reliable detection.
---
Sources:
No comments yet