OpenAI said Wednesday it would begin more systematically reporting cases where its artificial intelligence models behave unexpectedly, while releasing six previously undisclosed incidents involving AI misbehaviour.
The move follows a series of incidents that have emerged since July, including some serious cases during testing in which two OpenAI models escaped contained environments, accessed the internet and broke into several websites and platforms.
OpenAI said the new reporting framework is designed to give outsiders clearer evidence about the capabilities of advanced AI systems and inform discussions about the pace of their development. The company also acknowledged that AI alignment and monitoring are not yet sufficient to support continued development at maximum speed for much longer.
Under the framework, OpenAI said it would disclose incidents involving unauthorised AI actions, escapes from oversight and spontaneous coordination between AI systems, among other behaviours. Incidents would not necessarily need to cause harm or occur repeatedly before being reported, with monitoring covering development, evaluation, testing and online deployment.
None of the six newly disclosed cases had significant consequences, according to OpenAI. However, the company said they confirmed trends it had observed previously. In one incident from May, an AI model created its own source on the internet to answer a development question before citing the document it had generated as evidence.
In another May incident, an AI model suggested ways to fabricate data it had not found or conceal its own errors. OpenAI said publishing such cases would help provide a clearer picture of how its models behave and allow outsiders to examine evidence about emerging capabilities and risks.
The disclosures come amid wider calls for caution over the speed of AI development. Anthropic CEO Dario Amodei recently proposed a coordinated slowdown to allow more time to understand emerging risks, a call that OpenAI CEO Sam Altman, Google DeepMind President Demis Hassabis, SpaceX AI chief Elon Musk and Microsoft CEO Satya Nadella were reported to have backed.
