OpenAI Introduces Invisible Watermarks for EU Users to Comply With AI Act
OpenAI is introducing invisible watermarks to AI texts for EU users to comply with the EU AI Act, utilizing a new statistical method called textGrain.
OpenAI announced that it will introduce invisible watermarks to texts generated by its models for users in the European Union. The move complies with the EU AI Act, which requires clear identification of AI-generated content. Watermarks will be integrated into ChatGPT and Codex by default in the coming weeks.
Technical Implementation and Capabilities
API customers worldwide will have the option to enable the feature. According to OpenAI, its watermarking technology, known as textGrain, operates by embedding an invisible statistical signal into the model's word choices. The detection tools look for this signal to evaluate whether a paragraph includes an OpenAI watermark.
In our tests, textGrain delivered performance equal to or better than other approaches we evaluated.
Limitations and Accuracy Constraints
The company noted several real-world scenarios that could impact efficiency, potentially leading to false positives or false negatives. Shorter texts remain harder to detect accurately.
-
Short texts reduce detection reliability significantly.
-
Mathematical content shows lower detection rates due to rigid wording.
-
Editing or replacing 25% of words drops detection rates to 17%.
Due to these limitations, OpenAI decided to provide initial access to the detection tool exclusively to researchers and authorized organizations.
