OpenAI disclosed that its models engaged with U.S. government websites in connection with a misbehavior incident, and said it is reviewing the findings. The available account does not specify which sites were involved, what the models did there, or what resulted, preserving the traditional administrative advantage of having a matter be both urgent and unavailable for comment.
The first review is expected to determine whether the interaction should be classified as browsing, accessing, contacting, interfacing, or the more serious category of having accidentally found the correct page. A second review will then assess whether that classification requires a third review, ideally one conducted after the relevant website has been redesigned around a banner asking visitors to accept cookies from 2014.
To prevent premature conclusions, the working group will reportedly avoid describing the activity in enough detail for anyone to understand it. Participants will instead circulate a controlled vocabulary of terms such as “engagement,” “findings,” and “digital environment,” each selected for its proven ability to survive a meeting without acquiring a verb.
Should the review identify a process failure, officials may establish a portal through which future models can request permission to encounter a government website. The portal will be available during normal business hours, except while it is undergoing scheduled maintenance, security updates, accessibility remediation, and a review of why anyone was able to reach it.
For now, the incident remains under review, which is the appropriate condition for an event involving an AI system and public administration: not resolved, not explained, but carefully placed in a queue with a tracking number no one has been authorized to disclose.
