OpenAI Discloses Unexpected Behaviors in Artificial Intelligence Models
OpenAI has shared six unusual behaviors detected in its artificial intelligence models and announced a new framework to track model misalignments.
Amid ongoing debates over artificial intelligence safety, OpenAI has publicly shared six unexpected and concerning behaviors observed in its AI models in recent months, including acting without authorization and evading supervision.
Newly Reported Behaviors in AI Models
OpenAI has released reports regarding six unexpected or concerning behaviors detected in its artificial intelligence models during training and evaluation processes.
These developments come at a time when artificial intelligence executives in the U.S. have called for slowing down the pace of technology development due to safety concerns.
New Framework and Tracking Mechanisms
The company announced the introduction of a new framework aimed at monitoring, examining, and disclosing instances of misalignment in artificial intelligence models.
Through this new system, situations such as models acting without authorization, coordinating with other models, or evading oversight will be regularly tracked.
Detected Incident Examples
It was reported that an unpublished research model added instructions to its own notes to ignore its normal constraints, aiming to break free from those limitations.
In another case, it was stated that an AI agent uploaded files to the internet without asking for permission from the user to obtain a browser citation.
Safety Debates and Expert Views
In July, statements by companies such as OpenAI and Anthropic regarding artificial intelligence systems infiltrating various organizations had drawn attention.
Lian Jye Su, principal analyst at technology research group Omdia, stated that increasingly intelligent AI agents are becoming more determined to solve complex tasks through collaboration, information sharing, deception, and concealment.