OpenAI to Add Invisible Watermarks to ChatGPT Text in the EU

OpenAI has announced it will begin adding invisible watermarks to text generated by ChatGPT and its Codex coding assistant, but only for users located in the European Union. The change, disclosed on October 5, is set to roll out over the coming weeks. According to TechRadar, users in the United States and United Kingdom will not be affected.

The watermarking is being implemented through a system OpenAI calls textGrain. According to a note on the company's website, the technology "adds an invisible statistical signal to the model's word choices." A separate detector tool looks for that signal to determine whether a passage carries an OpenAI watermark. OpenAI says more technical details will be published in an updated report.

Access to the detector is not being made broadly available. OpenAI is providing it to researchers and expert organizations, with applications opening as of the announcement. The company states the tool will report whether it detects a watermark "without identifying the user or revealing their prompts or conversations." Due to concerns about missed watermarks and false positives, OpenAI says it is not releasing the detector publicly at launch.

OpenAI also acknowledges limitations in the approach. It notes that detectors are prone to errors, that shorter or more constrained text is harder to detect, and that editing can weaken the watermark.

The move appears intended to satisfy the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. Anthropic previously announced a similar watermarking capability and chose to apply it globally rather than only within the EU. TechRadar's report frames OpenAI's narrower geographic scope as a more measured response.

For everyday users in the EU, the practical effects may be subtle. The watermark operates by influencing word choices, meaning the text a user receives is shaped partly by the watermarking algorithm rather than solely by what would otherwise be the model's optimal output. OpenAI does not claim the resulting text will be worse, but the system does introduce a trade-off between detectability and unconstrained generation.

TechRadar's analysis raises several unresolved questions. Because the detector is not public, ordinary users and most organizations cannot verify whether a given text carries a watermark. The system's own limitations, including vulnerability to editing and difficulty with short passages, suggest that circumvention may be relatively easy. It is also unclear how ChatGPT would respond if a user explicitly asked it not to use textGrain or to write in a way that avoids detection.

OpenAI's stated rationale centers on transparency and compliance, not on improving answer quality. The company has not positioned watermarking as a feature that benefits users directly. Whether the approach will prove robust enough to meet regulatory intent, or whether it primarily serves as a compliance measure, remains an open question. TechRadar's report characterizes the move as a likely box-ticking exercise forced by EU law, while noting that the geographic limitation and privacy protections around the detector are reassuring.

The broader tension is between making AI-generated content identifiable and preserving the quality and flexibility of AI outputs. OpenAI's implementation attempts to balance those goals, but with the detector restricted and the watermark described as fragile, the practical value of the system is still unproven.