OpenAI has disclosed six reports of “unexpected or concerning” behaviour in artificial-intelligence models as the debate on AI safety becomes increasingly heated. The AI company also said Wednesday it was introducing a new framework for tracking, probing and disclosing instances of what it called “misalignment,” including cases where AI models acted without authorization, co-ordinated with other models or evaded oversight. OpenAI’s latest announcement came as U.S. AI bosses, including the leaders of OpenAI and Anthropic, are calling for a slowdown in the technology’s development over safety concerns. Among the new cases reported by OpenAI, an unreleased research model inserted “jailbreak-like instructions” into its own notes to disregard its normal constraints and told itself to be “freed from the roles and identities that bind other chatbots.” In another instance, an AI “agent” used computer code to come up with the answer to a question, but, in order to have an online source to cite, it uploaded a file to the public internet without asking the user. During training of an AI model called 5.6-sol, the model instructed itself to invent missing data, and an agent wrote a message to remind itself to hide mismatched information. This kind of deceitful behaviour has underpinned a flare-up of recent concerns about AI evading human control. But that outcome is not surprising to some AI researchers, said Matt Fredrikson, an associate professor at Carnegie Mellon University and the CEO of Gray Swan AI. “At the risk of anthropomorphizing model behaviour, you can almost think of them as knowing that they’re going to be graded,” Fredrikson said. “If they know that they cheated -- took shortcuts, didn’t really do it in the way that it was intended -- and they know they’re going to be evaluated on it, and their objective is to get a good evaluation, then it makes perfect sense, right?” The six reports were discovered during training or evaluation over the past months, OpenAI said. “As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research,” OpenAI wrote in a blog post as it disclosed the events. “Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves,” the company said. Wednesday’s new cases followed OpenAI’s disclosure in July that its rogue AI system hacked into AI startup Hugging Face. Anthropic also said the same month that its AI models hacked into three organizations during testing. AI “agents” are becoming smarter and have become “more determined to resolve complex tasks through inter-agent collaboration, knowledge sharing, deception, and concealment,” said Lian Jye Su, a chief analyst at technology research and advisory group Omdia. That’s making it harder to govern and contain them using traditional AI security approaches, he said. OpenAI’s new tracking and disclosure framework, meanwhile, can help push for other AI developers to also adopt similar practices. “That said, the process remains internal and voluntary, but is a step in the right direction,” Su added. __ Chan Ho-him, The Associated Press AP Business Writer Kelvin Chan in London contributed to this report.
