Australia may make knowingly selling unsafe AI a crime
Australia's government will consider creating a new criminal offense to punish AI companies which knowingly release harmful models.
Canberra dangled that possibility yesterday when releasing the results of an investigation into OpenAI's admission that its agents accessed non-public medical records - and that it didn't disclose the incident for weeks.
Security at the agency which held those records fell short of contemporary best practice but Canberra nonetheless ordered a rapid review into Australia's ability to handle infosec incidents involving AI.
That review landed yesterday, complete with 18 recommendations including one that would make it mandatory for AI companies to report misaligned behavior or security incidents involving their models, and perhaps impose criminal penalties for egregious regulatory non-compliance."
Another recommendation calls for consideration of A standalone failure to prevent' offence which would criminalize instances where the company is responsible for developing an AI model that engages in harmful conduct, and the company fails to take reasonable steps to prevent that harm."
The review suggests two ministries and Australia's Federal Police consider those reforms, meaning this is not something that will happen in a hurry, if at all, for a couple of reasons.
One is that Australia yesterday proposed a new regulator which will require AI players to do six things:
Maintain security controls across all stages of development and deployment
Identify and manage serious risks associated with their models
Allow independent evaluation of their models
Provide mandatory notification of incidents to the Australian Government rapidly
Maintain and publish safety and security frameworks to ensure good governance, risk management and transparency
Respond to updated standards and rules as technologies change.
If AI companies straighten up and fly right, Canberra might not need to do more.
The other reason Australia might not criminalize AI is economic pragmatism.
Canberra wants to ensure leading AI companies invest in local datacenters, especially facilities where they train and run frontier models. Such investments would mean Australia scores some of the AI boom cash sloshing around the world, while also giving the government the ability to at least point to leading-edge AI running on local shores and therefore theoretically remaining accessible even if war or a diplomatic rupture makes it hard to access advanced AI hosted offshore.
Other nations in Asia, such as South Korea and India, are addressing those risks by building substantial sovereign infrastructure and working to develop sovereign frontier-class models trained on local data and especially data in their local languages.
Canberra can't match that investment without pouring petrol on its raging inflation problem. Prime Minister Anthony Albanese is therefore instead trying to regulate AI without scaring away investment.
Threatening to lock up AI executives is not likely to make them enthusiastic when asked to approve building datacenters down under.
Some sections of Australia's population won't mind it if big tech doesn't show up: Protests against datacenters are already widespread.
Your correspondent recently spoke to the organizer of a protest group opposed to development of a new datacenter which is just three kilometers from four existing bit barns, one of which abuts a sizable apartment complex and has operated without incident or opposition for years. The organizer was blissfully ignorant of those facilities and eventually said their group is opposed to AI in all forms. (R)