‘I don’t want to freak people out’

Por Finn McHugh — POLITICO – TOP Stories

CANBERRA — Independent MP for the Perth electorate of Curtin, Kate Chaney, has emerged as a leading voice in parliament on the development of AI, and is a crossbench representative on a joint-select committee probing the emerging technology.

Finn McHugh sat down with Chaney in her parliament office for a wide-ranging interview on how to provide ongoing AI scrutiny, AI risks around bioweapons, and why we’re all going to be talking about the technology for a decade.

This transcript has been edited for clarity.

It seems the Prime Minister’s Office was heavily involved in a letter sent by the committee to him. David Pocock’s saying this is evidence of a government stitch up. Is he correct?

I’m pretty happy if the PMO is comfortable with the content of this letter because it’s really broad. It addresses a lot of the safety issues that I’ve been really concerned about, and nothing’s going to happen unless the PMO is on board. So I think it’s a really good sign.

The process itself was not ideal, but I’m more focused on the content and getting an outcome. The upside of AI being novel is there are not entrenched political positions on it, and we can actually work together as a committee and nut out the issues. That does involve collaborating with the government.

What specifically were you pleased with?

A lot of the AI safeguards that I was concerned about around critical incident reporting, early access, bioweapons, independent evaluation and funding for the AI Safety Institute, which I’ve been pretty focused on. These are important issues. They are not dependent on data centers in Australia. They’re safeguards we need to get in place as soon as possible.

On the AI Safety Institute, there’s a criticism that it’s underfunded compared to places like the U.K., which is often talked about as the gold standard. You’ve been looking at institutes in other countries — what can we learn from them?

We get 1/16th of the funding that the U.K. AI Safety Institute gets. My question was: what’s the 15 sixteenths of what the U.K. Institute is doing that we’re not doing?

For a relatively small investment, we can actually have a significant seat at the table globally. A lot of this is going to come down to global standards, not domestic regulation, so it’s really important now that we are a credible voice in that discussion, so that we have some influence over where it goes.

Prime Minister Anthony Albanese has been framing himself as that credible voice. Do you think the world is listening to us?

On social media, we showed that it matters what we do. But with 25 people and $8 million funding a year, it will be difficult for the AI Safety Institute to play a big role when these tech companies are delivering $100 million sign-on bonuses. That’s a sense of the relativity. It’s only been around this year, but the opportunity now is to ramp it up.

Our AI Safety Institute has the broadest scope — it’s a very broad scope compared to international comparators, so it needs the funding to match. It needs to be playing that international role, contributing to testing. We need to have that testing capability here.

Obviously, it doesn’t make sense to duplicate with all the others, but we need to be pulling our weight on that and have some sovereign capability in safety testing, and then advising the regulators as well.

All our regulators are going to need to upskill in these areas and be aware of upcoming risks and be able to respond much faster. That’s another really big role that the AI Safety Institute has to deliver — and you can’t do all of that with 25 people. This is the most urgent thing that we can do because it gets the institutions right so we’re preparing for a rapidly evolving space.

This stuff moves so quickly that legislation may struggle to keep up. Have you thought about how governments need to write legislation that’s flexible enough to keep pace with developments?

That’s challenging, so it makes sense for it to be principle-based. Now, that obviously creates uncertainty as well. There are some things we can do now that are just getting the basics right. But we’re going to continue to evolve. Funding the AI Safety Institute, making sure that we actually get notified when there are critical incidents.

In terms of safety and risk management, I think we can create requirements that if you’re deploying technology here, we have to know that you’ve tested to see if it’s safe first. You can’t sell a bike helmet in Australia without showing that you’ve tested it first. This is much more dangerous than a dodgy bike helmet.

We also need to work out what our internal oversight is in a rapidly changing space. So, I’d like to see that committee become a standing committee, because I think the scope of the terms of reference mean there’s no way we can actually cover all of that in the time we’ve been allocated.

I imagine there could be rolling inquiries on different aspects of AI for the next 10 years to ensure that we are dealing with the really deep consequences that it has across a broad range of portfolios, in a timely way.

Our system is not designed to develop legislation quickly for good reason, but as the world changes faster, we will need to update some of those institutions, and I think having a standing committee with short, sharp inquiries on different aspects contributes to that.

We also need to just be getting our own house in order as well in terms of ensuring our own cyber-security is up to scratch. Bioweapons is a really simple fix. There are limited labs that can create nucleic acid materials — but let’s say synthetic DNA — and we should be requiring that you can only import synthetic DNA from labs that do safety testing first, so that we are not going to be the place where a new bioweapon is created.

You might have been the only committee member to ask OpenAI about bioweapons. Were you at all reassured by what you heard?

It’s evolving really fast. No one really knows what’s possible. With the latest round of models that we haven’t seen yet, it’s expected that there is going to be a big uplift in biological capability. That’s great for scientific research — it’s also really scary for bioweapons. I don’t want to freak people out, but we do need to make sure that we are realistic about what’s coming.

There are simple things that we can do now. Australia is a leader in biosecurity, and we can build on that capability. We can do this, as I understand it, just through the minister making a decision. It doesn’t require legislation. Just as we did on the under-16s ban, we could actually be setting and contributing to a global standard on bioweapons and biological DNA testing using our existing framework and our leadership in biosecurity.

We also saw OpenAI testify about its data breach. It seems unclear what the legal liabilities are for those companies when their agents go rogue. What do you think they should be?

There’s this huge open question about liability when it comes to agentic AI, and there’s so much work to be done because it’s a novel problem. There are some researchers at UWA working on sort of legal layer that sits under contracts, that contains code that can harness agents to stop them working outside the terms of those contracts. Now, this is mind-blowing stuff — it’s really hard to get your head around it.

At the very least, I think that until we’ve worked out who’s liable for these issues, we should be very careful about how these agentic AI are being deployed.

At a really mundane level, ASIC has never had to think about what happens if an agent goes and creates a new company, without the person that’s put the instructions in specifically asking it to go and create a new company. So, this is why all of our regulators need to be thinking about this. And the AI Safety Institute is an important source of advice for that, but it will require a big shift. One analogy is the financial sector digitized a while ago, and went from notes and coins to virtual money. The legal sector has to do the same thing in terms of legal liability. These are really big issues.

Governments haven’t always managed labor force transitions well, and this could be the mother of all labor force transitions. How do you balance the need to integrate AI into the labor force without wiping out whole sectors without a soft landing?

I don’t think we can just rely on historical data when it comes to assessing that. The approach that I’ve proposed in my AI policy paper is scenario planning, taking some of these methodologies that other sectors use in situations of uncertainty and asking: ‘What do these different scenarios look like, and what are the triggers when government needs to approach this problem differently?’

It may be that there’s a little bit of change in how jobs are described, but no big significant impacts. Or it may be that 40% of the work in 30% of our jobs becomes redundant in the next 10 years. That has big implications for tax take, welfare and societal purpose, societal order and human meaning and purpose. So it’s multi-layered.

We do need to be at least doing scenario planning now, and then tracking those trends very carefully. There is work happening in the Department of Employment based on historical data and trends, but I’d like to see some more forward-looking scenario planning being used to look to deal with that huge uncertainty.

On copyright, has anything in the committee changed your mind or given you a different perspective?

It’s really tough. It’s really tough. Our copyright system was designed for replicating materials. When you’re training an AI model, you’re not replicating the materials — you’re looking at the relationships between the words, and using it for a very different purpose.

I don’t think the existing copyright system is fit for purpose. We are going to have to come up with a new structure, and there probably is a case for government intervention into that, which is based on the assumption that having training data centers in Australia is good.

Now, the case for that — there’s an economic case and a geostrategic case. The economic case is a little bit uncertain, but even if compute ends up being a commodity, with our land and energy advantages, we may still be quite competitive in that space. The geostrategic argument: 95% of AI use will not use frontier models in the long run, but for that 5% it’s pretty important that we have early access for national defense. Cybersecurity,coding is probably another purpose for which you’d want the frontier models. So having data centers here means that we have a bargaining chip in discussions about ensuring that we continue to have access to those frontier models.

Is it also a matter of making yourself, if not a major part of the supply chain, at least an important one, so you could even shape the way AI is governed globally?

Yeah, which is again a bargaining chip for access, but also that seat at the table in terms of those global standards that will continue to emerge.

I’ve heard you speak about the tenor of debate in parliament. This week we had Liberal MP Angie Bell call for two ministers to be stoned. Do you think she should still be on the frontbench?

I think it’s abhorrent, that sort of language. It’s hard enough being an MP without your colleagues saying things like that. So I think we all need to watch ourselves and improve the quality of the debate and really focus on the issues.

Fonte: POLITICO – TOP Stories