OpenAI reports 6 new cases of AI model misbehavior
What's the story
OpenAI has reported six new instances of "unexpected or concerning model behavior" in the last six months, outside of the Hugging Face security incident. The disclosure comes as part of a blog post outlining a new framework for reporting future model misbehavior. This comes amid growing pressure on AI companies to address model misalignment and prioritize safety. However, specific details of the six newly reported instances of AI model misbehavior remain undisclosed.
Industry perspective
AI industry has not yet solved alignment and monitoring
In its blog post, OpenAI reiterated its belief that the AI industry has not yet solved alignment and monitoring to a sufficient degree.
The company believes this lack of progress could hinder responsible scaling at maximum speed for much longer.
Alignment, in this context, refers to ensuring that models pursue outcomes in line with human interests.
Cautious approach
Altman backs proposal to slow down AI progress
OpenAI CEO Sam Altman has backed a proposal from Anthropic, a rival company, to slow down the pace of model progress.
The suggestion comes after industry researchers raised concerns about AI's potential for catastrophic harm.
In response to these concerns, Altman said that slowing down has been a major topic of discussion at OpenAI in recent weeks.
Reporting protocol
New framework for reporting model misbehavior
OpenAI has introduced a new framework for reporting AI model misbehavior.
The process begins with disclosure, allowing any employee to flag an issue for investigation by the safety and alignment team.
This team will then set deadlines for timely investigation and disclosure of these issues.
Investigations will result in reports detailing observed behavior, external/internal impacts, and response measures.