OpenAI Implements Invisible Watermarking for ChatGPT Text in EU Compliance Push

Regulatory Compliance Drive

OpenAI announced it will deploy invisible watermarks to text generated by ChatGPT and Codex, starting in the European Union to comply with the EU AI Act's transparency requirements that took effect in August. The company said the watermark will gradually roll out over the coming weeks to eligible users across all subscription tiers within the EU region. Developers globally can activate the feature on select models through OpenAI's API starting immediately, though it remains disabled by default and is not being implemented worldwide at launch.

How the Technology Works

The watermarking system, called textGrain, operates by subtly altering the model's word selection patterns to create an undetectable signature that remains intact even when text is copied and pasted. Rather than adding a visible symbol, the method uses a secret key to influence how the AI ranks its next-word predictions, effectively nudging hundreds of word choices in ways human readers cannot perceive but specialized detectors can identify. OpenAI collaborated with researchers from the University of Pennsylvania and Yale to develop the approach and published a technical report detailing the methodology alongside the announcement. Testing revealed that the watermark produces no meaningful degradation in the model's performance.

Significant Limitations

OpenAI's own testing indicates the watermark system has notable vulnerabilities. When researchers replaced just 10 percent of words with synonyms, detection accuracy dropped from approximately 92 percent to 66 percent. The company also acknowledged that short passages, mathematical answers, and translated text prove substantially harder to detect. Due to these reliability concerns, OpenAI said it will initially restrict detector access to approved researchers and expert organizations capable of evaluating responsible implementation. The company emphasized that an absent watermark cannot serve as proof of human authorship, since text might be too brief, heavily edited, sourced from competing AI systems, or involve substantial human revision and creativity that obscures the original signal.

Broader Industry Context

OpenAI's move follows Anthropic's announcement two months earlier that it would watermark Claude-generated text on a worldwide basis. That decision sparked pushback from some Claude users who objected to being marked, arguing they had supplied the intellectual direction while the AI merely functioned as a tool. OpenAI had previously developed watermarking technology but chose not to release it, reportedly concerned that users would migrate to competitors offering unmarked output. The company joins Anthropic, Google, Meta, Microsoft, and others in committing to the EU's code of practice governing AI-generated content disclosure.