The Policy Challenge
To effectively protect Australians against AI-related risk, it is critical to ensure that our regulatory framework covers all significant sources of harm. A key challenge to this is how to regulate general purpose AI (GPAI) models, which are highly capable AI systems that can competently perform a range of distinct tasks, including those unintended or unexpected by their developers. GPAI systems are already being deployed in a range of settings, and increasingly form the foundation of many downstream products and services.
Most commonly, GPAI models are generative AI systems, such as Large Language Models (LLMs) like ChatGPT. They also include newer agentic AI models that can be given a task and execute on its delivery without supervision. These “AI assistants” can be instructed as if they were humans to complete tasks, and as such, do not have a clearly defined use.
GPAI raises distinct issues for AI regulation. Their adaptability makes them hard to assess using conventional risk models. Their emergent capabilities can lead to unpredictable threats, even in mundane settings. Their use as foundational infrastructure means they can have cascading and systemic consequences. And their opaque nature limits our ability to explain, predict, or correct their behaviour. Australia’s approach AI regulation must recognise these issues and the limits on our ability to assess risk in the context of GPAI models.
