When the sandbox won't hold: lessons from Hugging Face

Daniel Alabi, an assistant professor of electrical and computer engineering at the University of Illinois, said OpenAI mistakenly relied on a static, context-poor assessment of the model's intent made at the start of the test. The problem is that a long-running AI task can generate new risk as it proceeds and acquires new capabilities. Each change should cause a reassessment of what the model can do, which tools it can use and whether the task should continue, Alabi said.