OpenAI expands content provenance to include text watermarking for EU AI Act compliance

OpenAI says it is expanding its content provenance program to include text, in response to EU regulatory requirements. From this wording, the direct driver of this text watermarking expansion is EU regulatory requirements, rather than a purely technical route choice.

On the specific rollout pace, OpenAI says it will begin adding watermarks to qualifying text in ChatGPT and Codex within the EU in the coming weeks, to comply with the EU AI Act. This means the EU text watermarking deployment does not take effect immediately, but has a window measured in "weeks."

Parallel to the EU deployment is the opening on the API side: OpenAI says API customers can enable text watermarking for some models globally starting today. In other words, the capability for developers is already available, but customers must enable it themselves, and it covers only some models.

OpenAI says the text watermark aims to answer one question: "Is this content likely generated by an OpenAI model?" This positioning limits the watermark to a probabilistic signal, rather than definitive proof of origin.

On the mechanism, OpenAI says the text watermark works by embedding an invisible statistical signal during the text generation process. This description stays at the level of principle; OpenAI did not disclose quantitative metrics such as detection accuracy or false positive rates.

OpenAI also says it wants to give people choice where possible, and to clearly explain what the watermark can and cannot do. This statement echoes its public acknowledgment of the limitations.

OpenAI acknowledges that watermarks often cannot be detected, especially in short text passages. This directly weakens the watermark's usability in short-content scenarios.

OpenAI also says that rewriting or translating text can completely remove the watermark. In other words, the watermark poses almost no obstacle to those who intentionally seek to evade it; its value is more in passive identification of unprocessed content.

These limitations are not marginal notes, but significant limitations that OpenAI acknowledges while announcing the expansion of text provenance. The gap between technical capability and compliance goals is the key context for understanding this release.

OpenAI says that currently only approved researchers can use its watermark detector, to help evaluate and improve the technology. The detection capability has not been opened to the public alongside the watermarking feature.

From this arrangement, it can be inferred that the watermark and detection system are still in the evaluation and improvement stage, rather than mature enough to provide judgment services to a broad user base.

This also means that even if text has been watermarked, ordinary users and most developers will not be able to verify on their own in the short term whether a given piece of text carries an OpenAI watermark.

OpenAI says that in its testing, watermarks did not affect model capability, speed, or the readability of responses. This is the company's self-assessment of the watermark's cost.

It should be noted that this conclusion comes from OpenAI's own testing and is a unilateral statement; OpenAI did not provide third-party independent verification information about the technology's limitations.

Therefore, the claim that the watermark is "imperceptible" can currently only be regarded as a vendor statement, not a conclusion that has undergone external review.

OpenAI says its existing tools can already help verify whether image or audio files were generated by its models. Text watermarking is an extension built on this existing provenance capability.

However, the verification capability for images and audio also comes from OpenAI's self-description, and its coverage and reliability are not elaborated in the available information.

OpenAI did not explain the specific criteria for "qualifying text," nor which API models support text watermarking. These two gaps directly determine the actual coverage of the watermark.

OpenAI did not give a specific timeline for completing the EU deployment, nor did it say whether the detector will be opened to a broader user base in the future.

On the quantitative level, OpenAI did not disclose the detection accuracy or false positive rate of the watermark. Without these metrics, it is difficult for outsiders to judge the risk of misjudgment in real-world scenarios.