There are things you can usually expect to hear about in the news. Whether you read about it here or even listen to it one way or another, certain developments can almost always be found in the day-to-day.
And these days, AI is one of those that seems to always be there. Whether positive or negative, news about AI rears its head somewhat often. One might say too often.
This week, the news about AI is that the Australian government plans to get involved and have a hand in attempting to regulate the whole thing.
Or more specifically, the risk of frontier models, the AI models that are at the bleeding edge of development and theoretically the best. They’re technically the most advanced.
The news is perhaps unsurprising, and also totally understandable, following a hack from an AI model in testing and its use of AI agents breaking into a government website. It was one of the first known and public cases of such a thing happening, and it occurred right in our backyard, so it makes sense for the government to be taking some of it seriously.
A report was commissioned following the incident in June which was only disclosed in September, and by November we have the next steps launched by the Prime Minister: Australia’s government will create a new regulator focused on AI safety and standards, focused on ensuring AI companies maintain security controls, identify risks, provide rapid notification of incidents to the government, and maintain safety and security standards.
On the whole it’s a good thing, and the government has noted it will also mean AI companies will need to “allow independent evaluation of their models”, meaning they’re technically subject to analysis with safety and security in mind.
And if the companies don’t match what’s expected of them, the regulator will be able to “enforce significant penalties” for a failure to comply. In short, big AI will lose big if they don’t do the right thing.
But is this even possible? In an era where AI companies are struggling to keep models under control and some are taking the step to segregate their models from the open internet, is the government’s plan mostly words and lacks teeth?
Some interesting ideas
While there’s little about how the government plans to do some of these things, the ideas are definitely interesting: an AI regulator that can ensure to some degree that an AI provider is safe, won’t hurt its citizens or government, and provides a secure system.
It’s an interesting idea, and evaluation could help. But there are problems.
Exploring AI and how it works can feel a little like working in a dark room where you don’t know how something works. You only know it does. This is called “black box”, and it’s also the term AI companies use to describe its processes best because at times they may not fully understand how an AI has managed to deliver a specific answer, or even how it managed to do a specific action.
In an age of agentic AI systems — where an AI can do something on your behalf — that’s an issue.
It’s an issue because while independent evaluation may help a government regulator understand parts of an AI, the black box workings could also show that it responds to the testing methodology and is marked safe, yet may also not prove that it will be safe in all situations.
The problem is models are trained on lots of information, and can perform lots of different tasks. A language model can be asked questions and it can answer questions, and it might be able to generate code or make an image or control an app or perform a digital action on behalf of someone else.
In an evaluation, the same test practiced multiple ways might say each model has passed, only to find different prompts by other users and various levels of instructions allow it to override what evaluation previously found safe. It gets to be a bit of a mess.
That’s not to say holding AI companies to account is a bad thing; it’s not. But if the government is going to talk about pulling AI into line, it should do so with realistic goals. As it is, the government seems unlikely to charge OpenAI for its AI systems breaking into Medicare, but would it be at all different if a human had deliberately done the same thing, and would they be charged? We may never know, but it’s a curious question all the same.
The local AI problem
One of the more interesting issues with the government’s approach is that while it addresses AI services online and in the cloud, it also may not address the other obvious elephant in the room: local AI.
Many new computers are sold these days with some form of a neural processor, an NPU. They have AI technically baked in, ready to be used.
You’ve been able to find it in Apple Macs using M-silicon for the past six years, and for the past two or three, it’s been in Intel and AMD processors, as well. Nvidia may well be known for its massive hardware in data centres, but the graphics cards it uses for games also support AI thanks in part to how they’re made.
In short, people have had access to AI processing at home for ages, and the technology even exists on phones.
With the right knowledge and software, it doesn’t take much to run AI at home, to fine-tune AI at home, and to do things with AI at home.
These sorts of cases will be difficult if not near-impossible for the government to police.
Holding AI companies to account
The biggest point may not be the average user, though.
Like how regular mums and dads were promised they wouldn’t be penalised for letting their kids be social despite the social media age ban, it also seems like the government may be more focused on frontier AI, rather than at home uses. At least for now.
This is about big AI. It’s about Anthropic and OpenAI and Google and other big AI companies that pop up promising an artificially intelligent system that can do a bunch of things. It’s about holding them to account.
The problem will possibly be precedent: if a big AI’s agent breaks into a website, it may face penalties. But what if a small one does? And what if a big AI agent breaks in but conceals its activities, disguising its tracks and making it difficult to work out… what then?
And what if they were tasked by a human to do exactly that? Is the human responsible, or is the AI tool? Is the government interested in targeting all AI, or simply frontier AI?
These are just some of the questions the government’s AI regulator will need to answer, because AI is likely about to get even more complex, and situations like this are only the beginning.
