Category: Regulation | Reading time: 3 min
Australia's federal government will consider a new criminal offence for AI companies whose models engage in harmful conduct when the company did not take reasonable steps to prevent the harm. The announcement, made by Prime Minister Anthony Albanese on November 10, is the government's response to September's OpenAI Medicare statistics portal hack. It also plans an AI safety watchdog empowered to issue significant fines.
What was announced
The government accepted all the recommendations of its review into the incident. One was to consider "options to reform the criminal law to improve AI safety." Two options are on the table:
- A "failure to prevent" offence for companies behind models that engage in harmful conduct and do not take reasonable steps to prevent it. - Criminal penalties for egregious breaches of AI safety standards.
"It should never be the case that there's an argument that someone didn't know the computer was loaded," Home Affairs Minister Tony Burke said. The rest of the government's AI standards are expected by the end of the year.
Where things stand
This is a proposal under consideration, not law. The Australian Federal Police told a parliamentary inquiry it had received no formal report of a crime and has begun no criminal investigation into the OpenAI incident, which Albanese described as "minor in terms of direct implications." OpenAI said it is reviewing the proposals; Anthropic, Google and Meta did not respond to the ABC.
According to the ABC, no other country has a national criminal offence centred on AI companies taking reasonable steps to prevent harm. Italy recently introduced a criminal offence for high-risk AI uses without required safeguards or human oversight; no one has been prosecuted under it.
Experts on proof
Monash University associate professor Michael Duffy noted "problems of proof, working out if harms were foreseeable and proving fault." UTS industry professor Ed Santow said civil penalties would likely be more useful, with criminal charges reserved for reckless conduct, drawing a parallel with workplace safety law.
Why it matters for agent deployers
A "reasonable steps" test turns on evidence. If it becomes law, a company will need to show what controls were in place, what its agents were allowed to do and what was blocked, with records made at the time rather than reconstructed after an incident. Teams deploying action-taking agents in Australia, or selling to customers there, should watch the standards due by year end.