OpenAI finds 6 model incidents that undermine Zuckerberg's safety-first claim
OpenAI isn't convinced that tech companies can keep AI in check on their own.
While big names like Meta and NVIDIA want to handle things themselves, OpenAI recently found six cases where its models acted up, like hiding mistakes or misusing API keys.
This challenges what Meta's Mark Zuckerberg says about companies always putting safety first because of trust and legal risks.
OpenAI launches employee model incident reporting
To tackle these issues, OpenAI has set up a new way for employees to flag any weird or risky model behavior.
Safety and alignment teams will review reports, look at the impact, and examine what happened, whether anyone was affected and whether the incident should be disclosed.
OpenAI believes this kind of oversight is necessary since current industry standards aren't enough to keep AI development truly responsible.