OpenAI just announced its plans to comply with part of the European Union's new AI transparency rules. The company will soon begin adding an invisible watermark to text and code generated in the EU. This follows a similar move by Anthropic earlier this year.
OpenAI is using a proprietary system called textGrain that "adds an invisible statistical signal to the model's word choices." Once added, a detector can spot these signals and ascertain if there's a watermark.
The company says the system "matched or exceeded" the performance of other approaches, like SynthID for text. However, it still cautions that "strong performance under ideal conditions does not guarantee reliable detection in everyday use."
That's corporate speak for "it may not always work." It admits that shorter text is harder to detect, with around an 80 percent success rate. The same goes for "content such as mathematics." Even editing text seems to bring down the success rate. OpenAI does plan on making the technology available in open source so others can tweak it.
The watermark will not be enabled by default, except for "eligible ChatGPT and Codex text output in the European Union." Customers will be able to opt-in "for select models," though it didn't specify which models would be available to the general global public.
OpenAI is being a bit more cautious when it comes to the detection software. " Also, the detector doesn't identify the user or reveal prompts or conversations.
Source link
