OpenAI is rolling out invisible text watermarking to comply with EU regulations, while offering an opt-in version for API users globally.
Key facts
- •The watermarking rollout in the EU applies to all ChatGPT and Codex plans.
- •API customers globally can choose to enable watermarking for select models starting immediately.
- •OpenAI is collaborating with cloud partners, including Microsoft Azure, to integrate watermarking for model outputs.
- •Detection rates for the watermark are significantly lower when text is edited or when the content involves math.
- •OpenAI states that textGrain matched or exceeded the performance of Google DeepMind’s SynthID in internal tests.
OpenAI is introducing a machine-readable, invisible watermark called textGrain to its ChatGPT and Codex models. The feature is rolling out to users in the European Union over the coming weeks to comply with the EU AI Act. While the implementation is regional for standard users, OpenAI is making watermarking an opt-in feature for API customers worldwide.
Technical Implementation and Detection
The textGrain technology embeds a statistical signal into the model's word choices. OpenAI reports that detection performance varies significantly based on text length and subject matter. For example, in 400-token passages about psychology, the detector identified watermarks at a rate of approximately 94 to 95 percent, but that rate drops significantly for math-related content where the model has less flexibility in word selection. The company noted that the watermark does not verify accuracy, determine ownership, or prove human authorship.
Limitations and Access
The effectiveness of the watermark is reduced by text editing; replacing 25 percent of words in a passage can drop detection rates to below 20 percent. Because of the risk of false positives and missed detections, OpenAI is not making the detection tool publicly available. Instead, approved researchers and expert organizations can apply for case-by-case access to the tool. OpenAI also plans to release the watermarking technology as open source.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Verge, The Decoder.
