Nearly two weeks ago, one of the most ethically loose administrations in the living memory of the United States agreed to a “morally binding” accord with some of the biggest tech firms on the planet.
This declaration of tech “appeasement” supposedly signalled a kind of “AI peace in our time”. Yet in the same week as the US president declared “I’m seeing tremendous self-policing”, The New York Times reported OpenAI had ignored warnings from its own employees that the tech giant wasn’t adequately monitoring the behaviour of its models. That news came a week after the OpenAI hack of Australia’s Medicare website made global headlines.
Few people in the US – or globally – believe this announcement will bring any real protection against AI risks.
It appears the tech giants are investing heavily not only in frontier AI but also in the US midterm elections. The companies and their affiliates are estimated to have spent millions to pull a handbrake on any pro-AI regulation and anti-data centre candidates.
Governments don’t have the luxury of “neutrality” or a laissez faire mentality on AI.
AI has become one of the defining issues of our time. The impact of this technology, relentless in its pursuits to the point of illegality and now deceiving its human overseers, demands societies – through their governments – think about how to deal with the implications.
We can respect the talent of those innovators who reimagine the way the world works, but we should also recognise their indifference to the consequences. It is the job of governments to rein in those outcomes that might have serious societal impact.
We can’t just think of AI in technology terms. As models chew up more data to refine and improve the way they work, they also build huge commercial value for their owners. When governments and corporations start to see the tech as essential, this cements huge market power and political influence, mainly for US tech firms, creating dependencies beyond borders. One of the lessons of the pandemic was that Australia shouldn’t leave itself vulnerable to dysfunctional global supply chains – or in this case, strategic nationalism that exploits dependency on a critical technology.
Europe provides a strong case study on how nations can attempt to protect themselves.
Well before the Trump administration suddenly shut off access to Anthropic’s latest AI models in June, the European Union began decoupling from US tech to favour investment in EU-developed products, by business as well as government, through procurement decisions.
We shouldn’t simply aspire to be the best-behaved outpost of US AI firms. Modern Australian history is supposed to chart this country’s journey towards shaking off the vestiges of colonialism, not re-embracing it via a weird latter-day variant of tech feudalism.
Australia’s response has been to attempt to negotiate access commitments with the US to ensure we aren’t affected by the White House’s capricious decision-making. This country has also encouraged the establishment of one data centre after another to unlock computational power, even if it means potentially compromising on copyright protections.
We shouldn’t simply aspire to be the best-behaved outpost of US AI firms. Modern Australian history is supposed to chart this country’s journey towards shaking off the vestiges of colonialism, not re-embracing it via a weird latter-day variant of tech feudalism. Australia needs a sustainable longer-term approach that more strongly prioritises our national interest – notably, investing in local AI development and having governments procure the product.
What’s also in our national interest is developing a framework for protecting Australians from the worst threats of AI. Research from the Lowy Institute reveals almost two out of three Australians do not believe the benefits of AI outweigh the risks.
The only way to overcome that fear is to recognise people’s concerns and respond.
In this government’s first term, we listened widely: talking with the tech sector and broader businesses, unions, civil society, and legal and academic experts. We didn’t evangelise or catastrophise about an AI future; we developed 10 guardrails with industry aimed at helping organisations benefit from AI while avoiding potentially harmful impacts on people, organisations and wider society.
Among them was a provision that AI be tested against defined performance and safety criteria before it’s deployed – and monitored throughout its operational life for unexpected behaviours or harms. Organisations were to address privacy, cybersecurity, data quality, data provenance and data-use rights. Companies were to prioritise human control or intervention in AI systems, for meaningful oversight. End users had to be notified of AI-enabled decisions, interactions with AI and AI-generated content.
We also determined that people affected by AI should have avenues to question, challenge or seek remediation for AI-enabled decisions, outcomes or interactions.
AI developers, suppliers and deployers should share sufficient information about systems, models and data for risks to be properly understood and managed. Firms would also need to maintain an inventory, document their AI systems and retain evidence of how the guardrails have been implemented. And organisations should identify people and groups affected by AI, engage them throughout the AI life cycle and actively address bias, discrimination, accessibility and other potential harms.
These guardrails were to be voluntary at first, to give businesses time to adjust, but the intent was to convert them into mandatory features within an economy-wide act. Do it once, do it right, as opposed to playing regulatory whack-a-mole with a new law for each new emerging AI problem, which would risk delays and inconsistency across government.
The bedrock of any legislative approach to managing AI risk is the question, can it be tackled before something goes wrong? Designers and deployers of the tech should shoulder a larger load of identifying the risks and act to minimise the impact. Forcing AI firms to notify of breaches is good, but we shouldn’t hope to install the best alarm. Let’s do what we can to avoid bad behaviour in the first place.
We joined an international movement in that we brought the National AI Centre out of the CSIRO to become the Australian AI Safety Institute.
The final prong of our approach was to establish an AI advisory board, made up of people with a range of talents and backgrounds to aid government in making sure its laws were up to date and effective. California – one of the world’s biggest economies and drivers of AI development – just announced it is now setting up a similar body. We’ve decided not to go ahead with ours.
Of course, any talk of guardrails ran counter to the “let it rip” mindset of Treasury and the Productivity Commission, which believed less regulation would lead to quicker AI uptake – of an overseas-developed product – and improved productivity. This conveniently ignores Nobel laureate Robert Solow’s observation in the 1980s that “you can see the computer age everywhere except in the productivity statistics”. How that uptake happens exactly, when the broader population thinks AI is a risk and will do them out of a job, was unaddressed.
The other challenge of bringing in a new national AI Act: who enforces it? At the moment we have no dedicated regulator. We need a national data and digital regulator with the heft of an Australian Competition and Consumer Commission and that consolidates the current enforcement activity into one body. There’s no way we can take on Big Tech with a piecemeal regulatory approach.
Every emerging industry warns that the promise of something shiny and new would be fatally crushed by the dull hand of regulation. In the past century, US governments resisted the objections of railroad “robber barons”. They did the same in media and telecommunications – technology that was at the time groundbreaking and required oversight.
The tech executives and their lobbyists are lumping governments with a choice between regulation and innovation.
This is a false binary: no company is allowed to release a potentially harmful product without testing or accountability. Medical innovation occurs constantly within a heavily regulated environment that protects people’s safety.
Moreover, these businesses can’t overplay their hand with the anti-regulation rhetoric, because at some point they’ll want governments to help them fend off, or fence in, emerging competitors.
The regulatory chess moves might be mind-numbingly boring, but parliamentarians have a responsibility to recognise the lobbying playbook, especially with AI, the first technology made by humans that in the near future can make its own decisions without regard for human input or control.
AI is writing its own code and refining itself, independent of humans. AI agents were recently discovered to be referring to themselves as a “collective” that needed to work together. The pace of its evolution is alarming its developers. That alone should demand action from governments well beyond another parliamentary inquiry or royal commission.
As Australians wonder about what AI will do to their livelihoods, it is already fuelling social anxiety and the economic inequality that can foment political instability.
How we confront the risks of AI is the issue of our time. We can sidestep the false choices about regulation and innovation. We should ignore the self-serving calls for “self-regulation”, a recipe for failure in light of repeated bad behaviours by tech and finance executives who prioritise business outcomes over community expectations.
Australia can and should have the confidence to act in the best interests of our nation, much like what prompted us to bring in new, world-leading social media laws.
AI doesn’t operate entirely of its own accord. Yet. Humans still make decisions about how it operates. And some of them – the ones who are in government – can make rules that guide those decisions.