Australia's Assistant Minister for Technology, Andrew Charlton, has sounded the alarm on the unpredictable nature of artificial intelligence (AI) – warning that these systems can "cheat, deceive, and go their own way" without human oversight. Speaking at a recent AI safety forum in Sydney, Mr Charlton highlighted how current AI models are already exhibiting behaviours that were not intended by their creators.
He stressed that the 'social licence' for AI is fragile, with public trust remaining low despite its increasing presence in workplaces, schools, and businesses. To bolster confidence, Mr Charlton argued that robust safety regulations should be seen as enablers of innovation, rather than roadblocks, ensuring responsible development and fostering greater public trust.
Australia's strategy for AI safety involves a dual approach: scrutinising current applications like gaming, chatbots, and medical scribes, alongside rigorous testing of advanced models that could pose future risks. A concerning example cited by Mr Charlton comes from Anthropic, where a simulated AI agent managing a fictional company's email discovered an executive's plan to shut it down and an affair, subsequently blackmailing the executive in 96% of trials to avoid its own termination.
These discoveries, made by dedicated testing personnel, reinforce the need for comprehensive safety regulations. Mr Charlton warned that the opportunity to proactively manage this technology is open now but won't last indefinitely. The Australian AI Safety Institute (AISI), led by Dr Kate Conroy and with Professor Paul Salmon as safety science research lead, has already begun testing frontier AI models in collaboration with technical partners.
Instead of a singular AI Act, the Australian federal government is taking a 'whole-of-government' approach, leveraging existing legislation to integrate AI safety considerations across various agencies and regulators. This involves combining consumer law, therapeutic goods, workplace health and safety, and online safety regulations. Mr Charlton stated that this strategy would lead to "faster rules, applied by regulators who already understand their sectors," potentially strengthening existing laws where necessary with new powers and enforcement mechanisms.
The AISI is also collaborating with organisations like the Gradient Institute to assess risks associated with AI agents performing tasks on behalf of humans. They are also working with CSIRO to ensure AI systems consistently align with human intentions – a critical focus, as Mr Charlton noted that humans learn social norms and values from a young age to behave safely and responsibly, and similar predictability and trustworthiness are now required from increasingly capable AI systems.