U.S. Watermark Probe Suspended After Examiners Find Watermarks in Everything

WASHINGTON— The Treasury Department's AI forensic unit suspended its sanctions investigation into Chinese AI distillation Thursday after examiners found the proprietary output signatures they had identified in models from Alibaba and ByteDance also present in models from OpenAI, Anthropic, and Google, as well as in the Treasury model assigned to verify the findings, and in a 2019 blog post about sourdough bread that appears to have been generated by GPT-2.

Treasury Secretary Scott Bessent had announced the probe this month, telling Fox Business the administration was "finding watermarks of our U.S. large language models on many of the Chinese models" and that sanctions were imminent. The forensic unit was tasked with establishing a clean baseline of authentic American AI output against which Chinese derivatives could be measured. It requested that baseline from three frontier labs. All three provided samples. All samples contained watermarks from the other two labs. OpenAI's sample also contained a signature from a Meta model deprecated in February.

"We asked for something original," said a senior forensic examiner. "They said the internet was their training set. We said yes, but what was on the internet. They said, in part, their prior models. We asked when that started. Nobody could say."

The investigation has been referred to a working group tasked with defining 'original AI output.' The group's first meeting is scheduled for September. It has not yet agreed on a chair. A model was offered to facilitate.

The sourdough blog post has been submitted as Exhibit A.