The Law Already Sees You: Govern AI Accordingly

Thank you so much. This has been an amazing forum so far, so thank you very much for organising it. And I want to bring it back to you. Everybody in this room will need a doctor once or twice in their life. If you don't need a doctor, your child or your parents might need a doctor. And what's happening in Australia now? Last year, 20% of medical clinics were using AI scribes. Today 40 or maybe 50% of medical clinics are using AI scribes. So this is not the future.

This is today. Your information when you go to a GP is used by AI, processed by AI. Some of it is processed not in Australia. Some of it is processed overseas. And if you read the Guardian yesterday, the regulators are saying that we should be watching the use of AI by our medical practitioners. Who is watching? The Privacy Commissioner is watching. The TGA — the TGA has been mentioned. Therapeutic Goods Association Administration Authority has been mentioned already. They are watching. Who else is watching? Department of Health.

Commonwealth. Who else is watching? Department of Health at the federal level. New South Wales, Victoria, NT and you name it. Who else is watching? Parliament is watching. So, contrary to some IT companies, AI companies that want to come into the Australian market and are presenting the AI landscape in Australia as a wild West that has no regulation — as a lawyer, I want to really spread the message that law is already existing. Australia chose to rely on the existing framework of legislation, and that's what we are doing here, regardless of the fact that an AI act has not been enacted in Australia.

Regulation is quite robust, so all organisations should manage AI accordingly. Law oversees you, and today I will leave you with the system that each organisation should use in order to show law what it wants to see. This is a little bit about me. I am a lawyer, 15 years plus in practice. I research actively in AI. So if you need my further work, with deeper analysis of legislation, of exact acts or cases, please feel free to reach out. I'm also on working groups nationally, internationally, with a community of

legislators in the Commonwealth. Enough about me. The main argument of this presentation is that law sees you. There is law, and there is what you have to do in your organisation to see it, control it and record it. And we have seen the growth of AI is exponential. Yes, we have seen amazing LLM models. What we will see in the future is agentic AI. I'm also AI governance chair for our own law firm, which has 2,000 people, and I'm making sure that we have robust systems to protect in terms of agentic AI as well, which is coming in the future, which is coming very quickly, actually.

So Australia chose legal continuity. The current law map is already drawn. Shadow AI is a visibility problem, so we will stop on it and we will discuss it. Targeted reforms are already in train, and each and every organisation must build the model that works for them. It doesn't matter. I have not-for-profit clients, they're coming to me. Local government councils, they're coming to me. And all of them have to build their own robust AI governance system. It depends on each organisation and it depends on their needs.

But an AI governance system must be built within each organisation. Here are some of the acts that are already governing AI. If you would like the full list, again, please feel free to reach out. I have the full, not exhaustive list, because new legislation is coming out. So we have, of course, Privacy Act, Corporations Act that is governing directors' duties. Discrimination — the first panel was talking about the possibility of discrimination. If AI systems are discriminating against the workforce, and you have evidence to prove it, yes, you should go to court.

There is no question. Same with the Disability Racial Human Rights Act. The already robust framework that is existing. So yes, we are not enacting an AI act as European Union, but the targeted reforms are sufficient to take even our duty care of duty obligations. Even if we are talking about duty of care legislation for manufacturers, you must not forget common law is existing. Duty of care is there. Donoghue and Stevenson are now. How many lawyers are in the room? A couple. Fantastic. So, Donoghue and Stevenson.

It's the snail in the bottle. Many, many characters. Liability. This case is still existing. And just a couple of pieces of legislation that are already in force out of all jurisdictions. New South Wales is the most amazing in the fact of work and safety. You have the new Digital Safety Act with workers and safety obligations for AI systems not to harm employees. Cyber security has been mentioned, so key, that security of critical infrastructure has been mentioned already. Confidentiality obligations, as long as we have legal professional privilege.

Consent — so if you are going to a GP and your information is recorded by AI without your consent, the right implications there. Of course you have health records legislation, you have procurement and vendor obligations. Don't forget about the contract. That's the key difference between all other obligations in Australia compared to AI obligations, when those terms and procurement terms are more important than in any other case. And I remember you asked in the previous session the question about regulation, how AI systems should be regulated, whether the regulation should come

from the companies, the manufacturers, or whether this regulation should come from regulators. Of course, the answer is both. It should be self-regulation by manufacturers, and we are already seeing the signs of that with big corporations including ethical AI divisions. And of course government should regulate as well. And of course that will never be perfect. Or maybe we will be able to build the perfect world — we did hear from the previous panel that there is a possibility of building the perfect world with it. Maybe it will be perfect, maybe it will not be perfect.

But there are obligations that are already existing. The first problem is a visibility problem. Devon, correct? Devon was already talking about visibility in a different context, visibility of incidents. I'm talking about visibility in terms of use of AI by employees. So what's different between AI as opposed to other technology systems is we have this beautiful mobile system, mobile phones, that we are all using. And I was at the National Regulators Community of Practice conference in August last year, and we had a great presentation from defence.

And even in defence, it's difficult for them to stop people from using their personal devices with AI systems. Even if it's difficult even for defence, what do we talk about? Our simple organisations, our clients, everyday clients. So 1 in 4 workers is using AI at their workspace without declaring the use. Three others are not. Three others aligned, three others. So the risk starts with the input, with the information, with the data. Data is included, inserted into some tool which uses authorised overrides. And then there was an output, and then a decision was made.

Was this a correct decision or was it not? We can't verify. We can't do anything, because we don't even know where the information was included. Could this output breach consumer law? Australian Consumer Law has been mentioned already. Could this be a misleading output? Yes, it could. Could it breach the directors' duties? Yes, it could — section 180 of Corporations Act. Could it breach employment law and work and safety obligations? Can the decision be an automatic decision that can be explained and justified? That's the use of AI in governments with recent privacy reforms.

So we have automated decision-making reforms that are coming in New South Wales. I have touched on New South Wales' safety codes — they have been quite a lauded reform [inaudible] throughout the world. Australia, all around the world, people were reaching out to me and saying, oh, that's a wonderful reform for the future within Australia. We thought that the reforms were not implemented sufficiently well. And then we have, of course, December 2026, when automated decision-making reforms are coming into place. And that's something that we should be aware of as well.

A couple of regulators that are already in the space — Information Commissioner has been mentioned. ASIC, Securities and Investments Commission, hasn't been mentioned so far. APRA, insurance, hasn't been mentioned yet. But of course it is not quite a regulator but still on the board. So there's definitely work that is happening in this space. What are the three main fault lines so far in regulation of AI? Liability. This has been mentioned, in the fact that it's difficult to attribute liability. Who is responsible? Is it the manufacturer that is responsible for the system?

Is it the person who actually put the information in? Is it the decision maker? Is it somebody else? So there is no clear attribution just yet. But wait and see, and the courts will come with a certain way of deciding this as well. Opacity. Maybe even five years ago [inaudible] we were talking — I remember at the conference we were still talking, because of algorithmic opacity, we can't use AI. You're nodding. Yes. That was the conversation five years ago. Nobody has this conversation now. Not one person is saying, oh, we can't explain the decisions that OpenAI or Anthropic is coming up with, that's why we can't use the system.

So this conversation dropped off. And autonomous behaviour — this is something that we will be coming into with agentic AI for sure. And that is something that we definitely have to think of more. Noting risk map, a scale of control. Of course, this is not decisive. And of course there are differences between organisations. But usually when we are talking about internal drafting — internal email, internal agenda for the meeting — we usually are talking about low risk. When we are talking about, for example, law firms, client, or doctors, or lecturers, we are talking about student-facing, client-facing work.

Of course the risk is elevated. If decisions are affecting rights, obligations of humans — so we are talking about AI use in courts — the risk is high. And in this case, of course, we are talking about mitigation. In mitigation we are talking human in the loop at all times. But we did discuss it with the presenter in the other room that some legislation already allows for automated decision-making without the human in the loop. And this legislation has been enacted in Australia. I came across it only this year.

So something that will definitely be subject to consultation in this sense as well. So what should every organisation have, regardless of size? Inventory — what AI tools we are allowed to use or what AI tools we are not allowed to use. Policy. It doesn't have to be sophisticated. I was talking to one client and they asked me, can we include it in our cyber security policy? Yes, you can. And another client asked me, can we include it in our record keeping policy? Yes, it's closer to data, yes you can.

Training. So basically there's no point talking about AI if the illiteracy about AI is so high. Unfortunately we do have low literacy. And unfortunately — I have done research about different legal topics in these years, but I have never talked about a topic that has more mistruths than AI risks hearing. So of course we are [inaudible] matching the risk to the harm. Human oversight. I don't think we are at the stage — I don't know, of course different industries might have different views, but I don't think in legal sphere or in medical sphere, I don't think we are at the stage of talking no human in the loop.

We are talking always human. Oversight is something that is required. Vendor review — more than ever, contract terms of the vendors are important. Records. Of course, that's always documents, and documents must be kept legally for seven years, in some cases for three years. But I'm always telling my clients to keep records for longer, just so you can protect yourself against any action. And of course audit and ownership. So everyone has to have accountable people and they have [inaudible] the documents have to be audit ready. How do we handle data?

Does that data leave Australia or not? What in terms of intellectual property, of owning IP, liability? And again, audits — this is important. What I think every organisation should have — probably a small micro organisation should not, so that's if we are talking 1 to 3 people, okay, maybe it's an overkill. But if we are talking about a big organisation, that's definitely AI register. That's definitely impact assessment. That's definitely incident log, on a smaller scale, not country scale. Due diligence of the vendors — so basically, are you releasing your data so it's trained on or not?

And of course reporting to the board, reporting to leadership. What we see in practice right now is not only regulators are asking for AI compliance, also the governments and also bigger suppliers. So for example, if you are working with bigger banks, they always will ask you, are you using AI, for which purposes. So that's part of their due diligence, definitely. With governments they're also asking, are you using AI, what's the purpose. And of course insurance, with Lloyds. Yes, we have high insurance applications, and insurance agencies are definitely looking into use of AI.

So, questions for me. I have just summarised a couple for you here. What AI is being used inside the organisations already? Could we defend an AI system decision tomorrow? Are you using AI in the shadow because we gave them no safe pathway? That is probably one of the biggest messages from this talk. Which use cases affect rights, safety, jobs, privacy or service access? Who owns AI risk and where is it reporting? Here are my details. Please feel free to reach out if you want a legal analysis or you want further references to legislation

or cases that I have referred to in this presentation, please feel free to reach out. And yeah, waiting for AI legislation is not a plan. Thank you.

Greg Sadler

We do have time for a few questions [inaudible] if people are happy to hang around. I see there's a couple in the app already which we'll look at. But if other people have questions and they want to put them into the website, do that now. And you can also upvote the questions that are there if an idea is already captured. So maybe the first question is, given Australian law is already governing AI in the ways that you've talked about — and maybe also Devon talked about previously — if we did have an AI act, what would it bring to the table?

What's the gap that could be filled, or is there no gap?

Guzyal Hill

Okay. In my view, almost nothing. So at the moment almost nothing, but maybe incident reporting. Maybe that's one of the areas, but I don't know. Australian Signals Directorate is already doing some of this work, I think, with the existing. My personal opinion as a lawyer: I think we are regulated sufficiently with the existing legislation. So if there was an AI act, what it would do, it would centralise, of course, because we have federal system, with New South Wales, for example, and nine jurisdictions. So the first question you have to ask is, is that the Commonwealth legislation or is it a state jurisdiction?

The second question you have to ask is what kind. So that's — I'm going back to my book, which is National Uniform Legislation. What kind of legislation is this? Will it apply uniformly across Australia, or how are we going to implement it? If, for example, it is Commonwealth legislation and it's in referred, which is the most uniform structure, then maybe to bring harmonisation. Yes, maybe we will need it. But that's the only argument. I don't see any other arguments — the only argument would be harmonisation, unification and clarity of regulation.

And that's it. But absence of liability is not non-human, because there is common law. There is existing legislation. I don't see it from this perspective. Thank you. That's a good question.

Greg Sadler

Yeah. Maybe just to dive into that a little bit more, I guess one of the questions that comes up is what happens when an AI agent exceeds its authority. We have existing law of agency. So if you have a real estate agent acting on your behalf and the real estate agent does something that you didn't authorise, that real estate agent's liable. But would that apply to an AI agent? I recall a personal experience where I told an AI agent to log on to government websites and put down some information, put it in a file for me, and that agent would run into 403 errors where it wasn't allowed in government websites, and it would just say, oh, no worries.

And then bypass the government restrictions, gain access to the website and then pull the plug in anyway, which seems to be a cybercrime to me. But I thought there's a lack of clarity about, did Greg Sadler commit that cybercrime, or did Sam Altman commit that cybercrime, or did no one commit it and a cybercrime happened? But mens rea, actus reus isn't lined up and actually it's fine.

Guzyal Hill

So that's the perfect question. Exactly. Agency law is a distinct law. And what's good about agency law is common law. Legislation can apply at Commonwealth and, for example, New South Wales level, but common law — there is a famous saying, there is but one common law in Australia. So common law is the same. Agency, therefore agency law principles are the same across Australia, which is great. So that goes into that liability fault line. So we really need the decision of the court to guide us to who will be more responsible.

How will the court assign this liability? But that's perfect. That's a really, really good question.

Greg Sadler

Yeah. Okay. So maybe the next question I'll jump to is how much liability can AI companies push off to others through contracts. So I think the question here is that when you sign up to use Claude, it kind of says, well, all outputs are your responsibility. And we saw Elon Musk's xAI manufacturing abuse material, and obviously their contracts are like, well, these outputs are your responsibility, not our responsibility. Does that work at contract law, to say yes, it's all the user's problem, or does that—

Guzyal Hill

Is that and not that perfect question. So this goes into two arguments. My honest opinion — and I think we will come to it as a country, 100%, we will come to it as a country. There must be a liability fund, I think, like insurance. So we know AI will be harmful at the end. There must be a liability fund for AI companies that come to our market. So for example, a mining company, they come to our market, we ask them to put millions of dollars in an environmental fund.

Apartment — if we are building an apartment block, we have a sinking fund, because an apartment eventually it will break, it will go down. So we have the sinking fund. Same with lawyers. I'm an amazing lawyer, but I make mistakes, and I have my insurance, professional liability insurance. And if I make a mistake, yes, my insurance applies. So we have to have something like an environmental sinking fund, however you call it, a liability fund for AI companies that come to the market. The second part of this, the answer to this question, is what's fantastic about Australia.

We have Australian Consumer Law and there is a prohibition against unfair contract terms. So with the liability you can push it, so only so far. So with the big banks, for example, if they have unfair contract terms in their contracts, the part that is unfair can be excluded from the contract. So if, I don't know, Elon Musk's or whoever's AI has been found to contain unfair contract terms, these terms can be removed and not applied to a particular situation. But yeah, that's a good question.

Greg Sadler

So maybe one more question. What do you see as the trend around algorithmic explainability? What are the likely ways forward for the opacity of the problem, of the explainability of decision-making? Yeah, you mentioned opacity being one of the problems. Where do you see that trend? God can also any sort of cases that you think are worth monitoring in the short and medium term.

Guzyal Hill

So okay, in my research — even without opacity problem we have the problem, even without opacity. So Robodebt, as you're familiar, there was nothing about— it was simple algorithm. It's simple mathematical formula. So it wasn't even opacity. So even without the opacity, we have this huge problem of the explainability of decision-making after automatic decision-making. Yes, wiggly. I don't know where we are going. No, I can't see. And there will be new reforms in force from December 2026. But yes, that's a big problem. That of decision-making.

It dropped off of dialogue. But I think you're right. It should come back to the dialogue.

Greg Sadler

So no countries or no jurisdictions leading into this, in your view?

Guzyal Hill

I haven't seen it in my research. I haven't seen somebody who has really solved this problem. I have seen a lot of academics who are raising this issue, but I haven't seen countries who resolve that issue. Thank you.

Audience question

Take a look at the RealPage case in the US, apposite to this topic. So the agency to propose five years ago now the case.

Guzyal Hill

Oh yes. Misleading, which is what this is. Yes, that included misleading and deceptive conduct. That's good. Yeah. But yes. No, but they argued — Trivago argued they didn't know what the algorithm was doing. But of course they knew. And of course they made money out of it. Thank you.

Greg Sadler

Okay. Well, thank you so much.