OpenAI to Watermark ChatGPT Text in EU, API Users Can Opt Out
OpenAI will embed invisible watermarks in ChatGPT text within the EU, while API customers worldwide may opt out. Detection rates reach 95 percent but fall to 17 percent when a quarter of words are replaced.
OpenAI is introducing invisible watermarks to text generated by ChatGPT for users in the European Union, a move that aligns with growing regulatory pressure to label AI-generated content. The company will use a technique called textGrain to embed subtle, machine-readable signals in the output, according to tests that show detection rates as high as 95 percent under ideal conditions.
Unlike Anthropic, which has taken a stricter approach to watermarking, OpenAI will allow API customers worldwide to opt out of the watermarking feature. This decision could weaken the effectiveness of the system, as third-party developers building on OpenAI's models may choose not to participate. The opt-out option is not available to EU ChatGPT users, who will receive watermarked text by default.
The watermarking method is designed to be imperceptible to human readers while remaining detectable by specialized software. However, its reliability drops sharply when the text is altered. Tests indicate that if just a quarter of the words are replaced, the detection rate falls to 17 percent. This vulnerability raises questions about how robust the system will be against paraphrasing, translation, or other common modifications that can easily strip watermarks.
The European Union has been at the forefront of regulating artificial intelligence, with the AI Act requiring transparency for AI-generated content. Watermarking is seen as one way to help platforms and regulators identify synthetic text, though experts caution that no watermarking system is foolproof. OpenAI's decision to make the feature optional for API users globally suggests a balancing act between compliance and commercial flexibility.
Anthropic, a rival AI company, has reportedly implemented watermarking without an opt-out for its API customers, positioning itself as more proactive on safety and transparency. OpenAI's approach may appeal to developers who worry about watermarking interfering with their applications, but it could also draw criticism from regulators who want consistent labeling across all AI outputs.
The textGrain watermark works by subtly adjusting the probability distribution of words during text generation, leaving a statistical signature that can be detected later. While the method does not alter the meaning or readability of the text, it is not immune to editing. Even minor changes can disrupt the pattern, making detection unreliable in real-world scenarios where AI text is often copied, pasted, and modified.
For now, the watermarking will apply only to ChatGPT in the EU, but OpenAI has not ruled out expanding it to other regions. The company has not specified when the feature will go live or how users will be informed. The move comes as AI-generated content proliferates online, raising concerns about misinformation, academic integrity, and copyright. Watermarking is one of several tools being explored to address these issues, alongside metadata tagging and content credentials.
Critics argue that optional watermarking for API users undermines the goal of transparency, as many AI-powered applications rely on API access. If a significant number of developers opt out, the watermark's reach will be limited. OpenAI has not disclosed how many API customers are expected to participate or whether it will offer incentives for keeping the watermark enabled.
The effectiveness of textGrain also depends on the length of the text. Short outputs may contain too few signals for reliable detection, while longer texts are more robust. OpenAI's tests likely focused on typical ChatGPT responses, but real-world use cases vary widely. The company has not released detailed technical papers on the watermark's resilience, leaving independent researchers to verify the claims.
As AI regulation evolves, watermarking could become a standard requirement for generative models. The EU's AI Act is expected to set a global precedent, and other countries may follow suit. OpenAI's partial rollout and opt-out policy may be a test of how such measures are received by both users and developers. The outcome could influence how other AI companies design their own transparency features.
5
