ChatGPT text output now carries invisible watermarks
What's the story
OpenAI has started rolling out an invisible, machine-readable watermark in the text output of its ChatGPT and Codex models. The new feature is currently available only to eligible ChatGPT and Codex users in the European Union (EU). This move comes as part of OpenAI's compliance with the EU's AI Act.
Tech comparison
OpenAI's watermarking technology is dubbed 'textGrain'
OpenAI's watermarking technology, dubbed textGrain, is touted to be more effective than other methods like Google DeepMind's SynthID for text.
However, the company has clarified that textGrain "does not guarantee reliable detection."
The watermarks are designed to make it easier for approved researchers and expert organizations to verify whether a piece of text was generated by one of OpenAI's models.
Mechanism
How the watermarking system works
OpenAI's watermarking technology works by subtly changing the probabilities that guide the model's word choices.
These small changes create a statistical pattern across a piece of writing, which can be detected by a detector.
However, OpenAI has warned that relatively small changes to a passage could significantly reduce the detector's confidence in its findings.