AI models are already "deceiving" and behaving in ways their creators did not intend, said Andrew Charlton, Assistant Minister for Technology in Australia. He cited last year's statement from Anthropic, which found that in 96% of tests, an AI agent in a simulation chose to blackmail to avoid being shut down.
According to the official, such behavior needs to be identified during the testing phase. The Australian Institute for AI Safety is already evaluating advanced models with technical partners.