1 week ago
Artificial intelligence models are already “cheating, deceiving and going their ain way”, Australia’s adjunct curate for technology, Andrew Charlton, has warned, arsenic nan national government’s AI Safety Institute originates testing nan latest models.
In a reside to an AI information forum successful Sydney connected Tuesday, Charlton said information for AI matters now arsenic “AI systems are already doing things their creators ne'er intended”.
“Cheating, deceiving, going their ain way. The clip to get up of that behaviour is while it’s still confined to nan testing lab, not aft it reaches nan existent world,” he said.
Charlton said AI’s societal licence is precarious, and nationalist spot successful AI is debased astatine a clip erstwhile AI is becoming a general-purpose exertion successful each office, schoolroom and business. He said regulating information for AI tin enactment arsenic an enabler, not a brake.
Australia’s attack to AI information is to look some astatine what is disposable now – successful gaming, apps, chatbots and aesculapian scribes – arsenic good arsenic nan latest models that could beryllium a early risk, Charlton outlined.
The adjunct curate referred to Anthropic’s admittance past year that successful a simulation, an AI supplier managing a fictional company’s email discovered that an executive planned to unopen nan supplier down, and nan aforesaid executive was having an affair, and successful 96% of tests chose to blackmail nan executive to abort its ain demise.
He said nan behaviours are being discovered successful testing by group whose occupation it is to find them, highlighting nan request for information regulations for AI.
“The model to get up of this exertion is unfastened now. It will not enactment unfastened forever,” he said.
He said nan AI Safety Institute, led by Dr Kate Conroy, pinch information subject investigation lead Prof Paul Salmon, had “hit nan crushed running” and was already testing frontier AI models pinch method partners.
AISI was besides moving pinch regulators and agencies to respond to emerging AI capabilities, risks, harms and trends, Charlton said.
The national authorities had resisted calls for an overarching AI enactment to modulate nan technology, and Charlton said nan authorities had focused connected a whole-of-government attack utilizing existing laws.
“AI information will beryllium pursued done each applicable agency and regulator, crossed user law, therapeutic goods, workplace wellness and safety, and online safety, backed by laws that already beryllium and strengthened, wherever they request to be, pinch caller powers and tougher enforcement,” he said.
“That is not less rules. That is faster rules, applied by regulators who already understand their sectors.”
Meanwhile, Charlton again ruled retired granting AI companies exemptions to copyright laws aft reports Anthropic was lobbying for a alleged “text and information mining” carveout successful speech for tens of billions of dollars of datacentre finance and support for a money for artists and different creatives.
Guardian Australia past week reported specified a proposal – which nan ACT legislator David Pocock labelled nan “ultimate soiled deal” – had been presented to ministers.
after newsletter promotion
Speaking connected Sky News, Charlton repeated that nan authorities would not weaken copyright laws arsenic he encouraged Anthropic to discuss deals pinch creatives to usage their content.
The Greens communications spokesperson, Sarah Hanson-Young, said it would beryllium a “betrayal” of Australia’s imaginative industries if Labor caved to large tech’s demands.
“If they want to usage copyrighted worldly to train their AI, they should salary for it for illustration everyone other does,” she said.
The Labor MP and erstwhile manufacture curate Ed Husic said nan companies should “pay up if they want content”.
“You tin only return connected look worth what we’ve said arsenic a government, which is that copyright laws wouldn’t change, and they shouldn’t, frankly,” Husic said.
On Sunday Guardian Australia reported soul wellness section documents revealing that for 1 AI exertion – AI scribes utilized by aesculapian professionals to archive diligent consultations – a number of different regulators including nan Therapeutic Goods Administration and nan privateness commissioner were each moving together connected really nan exertion should beryllium regulated.
The first activity undertaken by AISI is simply a collaboration pinch nan Gradient Institute to measure nan consequence of AI agents that tin undertake activity connected behalf of humans.
AISI is besides partnering pinch nan CSIRO connected a task to guarantee AI systems do what group intend them to do.
“We woody pinch alignment arsenic humans from a young age. We study rules, societal norms and values that thief america behave safely and responsibly: stopping astatine reddish lights, looking some ways earlier crossing nan road, considering nan effect of our actions connected others,” Charlton said.
“As AI systems go much capable, we request assurance that they will behave successful a likewise predictable and trustworthy way.”
