OpenAI to Apply Default Watermarks to ChatGPT Outputs—EU Only

Meta description: OpenAI will automatically watermark ChatGPT text in the EU under the AI Act, and will offer textGrain detector access to approved researchers and organizations.

Tags: OpenAI, ChatGPT, EU AI Act, textGrain, AI watermarking

Featured image: Abstract technology style image of flowing text fragments encoded as faint hidden patterns, with soft blue and green network lines, no real people, no company logos, no text.

OpenAI ChatGPT watermarking is becoming a compliance step rather than a standalone safety feature. The company announced that text generated with ChatGPT will begin to be automatically watermarked in the European Union, while the same capability will be offered in other regions but disabled by default outside the EU. The change matters because it ties a product behavior to a legal signal: content produced by AI models should be marked in a way that another tool can detect.

Regulatory background: August deadline for detectable AI text

The move in Europe is driven by the need to comply with the EU AI Act, which took effect in August. The relevant requirement is not simply that AI-generated text feels machine-like or can be labeled by users, but that it can be marked for detection by tools. That creates a practical question for anyone who needs to assess the provenance of AI-generated text.

The technical difficulty is significant. There is still no completely effective and reliable way to do this. A few standards already exist, such as SynthID and the C2PA project, but they are relatively easy to circumvent for anyone with basic know-how.

Development path: textGrain adds hidden word-choice patterns

OpenAI's method is proprietary and is called textGrain. The company has published a technical paper explaining how it works. In general, it follows the pattern of other large language model, or LLM, watermarking tools: it places patterns in word choices that are not clear to a human reader and do not meaningfully change the general quality of the output.

The watermark is designed to be machine-readable rather than human-visible. Someone with a key can use a specialized detector to find it.

Latest rollout: automatic in EU, optional elsewhere

The rollout is split by region. Inside the EU, ChatGPT will begin automatically watermarking generated text, aligning the service with the detectability requirement under the EU AI Act. Outside the EU, OpenAI will offer the watermarking feature, but it will be off by default.

The detection side is also gated. OpenAI says it will give access to the detector to a limited number of researchers and organizations, and will provide a request-for-approval process for others to be added over time.

Current stage: limited verification and open questions

At this stage, the development is best understood as an early compliance implementation. OpenAI has introduced a hidden marker and a controlled detection path, but the wider uncertainty remains. The same limitation is likely to apply to OpenAI's watermark: if no completely effective and reliable method exists today, and if existing standards can be bypassed by users with basic know-how, the practical strength of textGrain will depend on access, enforcement, and resistance to removal.

For readers, the visible effect is that ChatGPT text in the EU may carry a hidden signal, while users elsewhere must enable the feature manually. The next phase to watch is whether approved researchers and organizations can turn limited detector access into a usable way to identify AI-generated text at scale.