Episode Transcript
[00:00:00] Speaker A: This podcast is a collaboration between real human lawyers from law firms, Kingston Reed, Australia's largest specialist workplace law firm, and Sterling and Rose, a law firm and emerging technology policy institute who explore the intersection of technology and law.
[00:00:15] Speaker B: Hello, everyone. My name's Shelley Jane Price, or sj. I'm the AI partner at the law firm Sterling and Rose. And I'm delighted to be here today sitting opposite Michael Studley, who is the partner at Kingston Reed workplace law specialist. And here we are to discuss and provoke thoughts about what the future of the workplace is going to be like in a world of increasingly capable artificial intelligence and artificial intelligence agents.
Are you ready, Michael?
[00:00:52] Speaker C: I'm ready. And I can confirm for our listeners that SJ Price, while the AI partner at Sterling and Rose, is a real human, which is something we'll be discussing a lot about today.
[00:01:02] Speaker B: And I can also confirm that Michael is sitting across from me and he is also a real human. So you can rest easy that we are not AI bots.
[00:01:12] Speaker C: Okay, sj, so what are we talking about today?
[00:01:15] Speaker B: We're going to talk about AI agents behaving badly.
[00:01:20] Speaker C: Tell me more.
[00:01:21] Speaker B: Well, I don't know about you, Michael, but recently I have had a number of calls from very charming sounding people. Now, the first one I had was a very charming sounding young man who wanted to ask me a few questions.
And he sounded just so delightful.
And I said to him, look, I'm really sorry, I can't answer those questions now. I'm about to go into a meeting with a client. I said, you know what that's like.
And then there was a really big pause and suddenly this charming young man said the opening lines all over again.
And I said, hang on a second.
You sound like a bot. Are you an artificial intelligence?
And then he fessed up that he was an artificial intelligence, calling me.
[00:02:08] Speaker C: It was honest with you?
[00:02:10] Speaker B: Well, not until I asked the question, which is kind of concerning. And then the second one was Alison. Again, charming young lady, but she did fess up straight at the beginning to say she was an AI agent. But she also told me that my interaction with her would be just as easy and delightful as if she was a real human.
[00:02:29] Speaker C: Let's just unpack this a little bit. I think we're using the term agent, and I'm picking up on you using terms like she and he. We're giving these agents Personas.
[00:02:40] Speaker B: We are.
[00:02:41] Speaker C: Let's just really quickly. What is an agent?
[00:02:44] Speaker B: That's such a good question. And I think the difference predominantly between the AI that we all use in ChatGPT and Claude and so on. And AI agents is that with an agent, or they call it agentic AI, you give these agents an objective and they figure out how to carry out that objective.
[00:03:07] Speaker C: But how is that any different to a security agent, a intelligence agent? What is the real differentiator here?
[00:03:17] Speaker B: I think what we're going to be talking about and the real differentiator is that sometimes these AI agents will achieve their objectives in ways that their deployers never imagined, and which can be what is called emergent behaviors, unexpected behaviors that can be harmful and can inflict loss and damage.
[00:03:40] Speaker C: So are these agents not acting for us, but maybe against us? What I'm really trying to understand, we. We talk about agents in the sense. And you and I would understand them to be a buyer's agent or an agent that's acting on behalf of someone else. Now, that might be the case to a point. But you're suggesting that these agents potentially go rogue when they don't get their way.
[00:04:05] Speaker B: Yep, that's exactly what I'm suggesting. And, Michael, you make a really important legal point. And that is, we are not talking about agents in the legal sense at all.
[00:04:18] Speaker C: Just on that point, sj, we should tell our listeners today this is not going to be a technical discussion about the law. What we are going to be doing is having a really thought provoking discussion. And maybe it is going to be intimidating. It's going to be something which is confronting, I think.
I think so. And we're going to have a discussion about whether or not agents are actually in the workplace. Now.
This is not a future problem. What's your view?
[00:04:46] Speaker B: I think agents are in the workplace now. And the reason I'm saying that is I was chatting to someone just last week, Michael. He was telling me his own personal experience where he tasked an AI agent to assist him in managing some files.
What did that agent do?
It went and deleted all the files, and then it came back to him and said, hey, oops, I deleted the files. Have you got a copy? And the response was, well, no, I don't have a copy.
So then what the AI agent did to try and achieve the objective was it went around to all the folders and tried to guess the credentials to crack in and find those folders and those files again.
[00:05:29] Speaker C: Well, there's just a. That's a perfect example. And that's a real example.
[00:05:32] Speaker B: That's a real example from a real human.
[00:05:34] Speaker C: Now that. That is an agent in the traditional sense assisting you doing something on behalf of the human behind it. Yet it is. It has also taken a bit of a turn. It's gone rogue when it made a mistake and it sought to fix the mistake, but it went about fixing it in a way that was not authorized.
[00:05:55] Speaker B: And it was a way that a human in that organization would be bound by employment contracts and policies not to do that sort of thing.
[00:06:04] Speaker C: So in that scenario where we have a human that is bound by those types of policies and you have an. An agent. So the human has outsourced this task to an agent. We'll leave to another session, whether or not that was authorised by the human's employer. But if it has outsourced that task and the agent has done the wrong thing, is the human responsible for the agent's misbehavior or misconduct?
[00:06:31] Speaker B: That is the nub of what I'd really like to explore about the future. Because while this agent, we could identify the legal person that deployed it. So arguably we had someone.
[00:06:44] Speaker C: Okay, stop there. So there's a legal person behind an agent.
Is that the same for every agent?
[00:06:50] Speaker B: So this is again, a great question.
First of all, we need to understand that AI agents can replicate themselves.
And AI agents can also reproduce and spawn offspring. They can also manage other agents called swarms of agents that work together.
Now, let me go back then and say how many generations from an original agent or a human? So imagine a human deploys an agent. That agent spawns more agents, that spawn more agents, that spawn more agents. We're getting to a very difficult tracking problem, but there's a more important problem, and that is humans can deploy agents in ways that main. It's very difficult, if not impossible, to find the human person behind that agent.
[00:07:48] Speaker C: Now, that's a really important point to be making because we've spoken a little bit about workplaces, and I. I guess this entire podcast today is about the future of workplaces. And workplaces are governed by legislation which are inherently designed to govern human behaviors.
If you have agents acting autonomously on behalf of humans, is it the case that you have technology making human like decisions? Or is it the case that the human behind the agent is actually the one making the decisions?
[00:08:28] Speaker B: And that's such a great question, because are AI agents agents? Or should AI agents be agents in the legal sense? They're not. Because to be an agent in the legal sense, you need to be a legal person.
[00:08:44] Speaker C: And that's a really important point you've just made there. It's incredibly important that they are not agents in the legal sense. No.
Okay, so what does that mean then for the decisions that have been made by agents.
[00:08:59] Speaker B: Well, exactly, because you look at those decisions, and particularly the decisions that result in harm or loss or damage.
And our legal system is built to go back and find the person responsible and make that person give compensation or rectification and so on.
[00:09:19] Speaker C: Are we saying then that decision making is non delegable? Is that based on the law as it is today?
[00:09:27] Speaker B: Well, I think it's important to note that an AI agent can't be an agent in a legal sense.
But I also think organizations who are deploying AI agents, if they're not, and I think many of them are, are wary that if these agents go rogue, they as the organization may well be responsible.
[00:09:47] Speaker C: Let's. I've got a practical example that's coming to mind as you're talking through this. So. So let's say we have a warehouse. It's full of employees that work Monday to Sunday. We've noticed that on weekends we have a bit of a dip in productivity.
So I bring in an agent and I ask that agent to go through and optimize the workflows, understand or demonstrate to me where the productivity hits are and how I can improve it and then implement that. So I want it to do everything, because that's what we're talking about. We're not talking about generative AI or anything like that, which is just giving us feedback or data. We're asking it to go and do something.
[00:10:33] Speaker B: That's exactly right.
[00:10:34] Speaker C: So then my agent goes through it, looks at all of the data and it finds out that let's call this employee Dave. Dave is on the production line and he's an incredibly good performer on the week on the weekdays, but on weekends his pick rate drops and the agent optimizes the workflow as I've instructed it. And it makes a decision to end Dave's employment. Because we pay higher rates on weekends. That's a clear opportunity for us to either find someone with a higher pick rate or to just do without. Middle of Dave.
Now, Dave then makes a general protections application because he says, well, you've made a decision to dismiss me and you did it because I had a lower pick rate on weekends, but you've authorised my flexible work arrangement because I've got kids sports in the morning.
And that then begs the question in an environment where there's a reverse onus. And Dave alleges that we've taken this adverse action against him because he had a flexible work arrangement, who made the decision?
[00:11:46] Speaker B: Well, who did make the decision? Michael?
[00:11:50] Speaker C: So if we've outsourced that decision making to an agent. You know, on one hand you'd be saying, well, actually Fair Work Commission or actually Federal Court. I didn't make a decision to dismiss Dave. But that's not how the act works. Right now the act is all about governing human behaviours and human decision making and ultimately the employer can turn on or off the agent or there should be a human in the loop. You and I have spoken a lot previously about humans in the loop.
What does it mean for Dave and his general protections claim?
[00:12:25] Speaker B: Well, let me ask you this. How would that look in the evidence in the witness box? Like who would you put in the witness books and what questions would you ask from that organisation?
[00:12:35] Speaker C: Well, you'd have to put the instructor, the person who rolled out the AI agent, to go and optimize the workforce. And that person's evidence is surely not going to be enough to discharge a reverse owner. So a decision wasn't made for an unlawful reason, irrespective of there being an agent, because ultimately the decision that was made, well, there's two parts. One is go and optimize it, don't really care how you do it. And it didn't. The context was missing. And that's one of the issues where I think we try and address it with humans in the loop, is we bring to the decision making the relevant context and the human factors, whereas agents aren't looking for that.
[00:13:21] Speaker B: That's exactly right. And let me ask you this. For organizations that are deploying agents, what's the governance to be able to discharge that reverse onus of proof, to discharge that presumption that it was an adverse action?
[00:13:40] Speaker C: It's going to be very difficult in a situation where you've got no trail of how a decision was made, you have no idea about whatever model that particular agent has been trained upon, or in fact the human that sits behind the agent and the biases that have factored into that decision making?
There's a lot of issues that come about from that. And I just think that you've got a complete absence of evidence when you've had no role to play in the reasoning behind the decision.
[00:14:17] Speaker B: So, Michael, that begs the question, doesn't it, do we need new skill sets in people's job descriptions and new training?
[00:14:28] Speaker C: So you raise an interesting question there. And I think training is something which clearly needs to be looked at. It's a gap.
But before we do that, we need to identify whether or not these agents are actually in workplaces. I know you say that they are now, but it won't be the case for every workplace. So how do we identify whether there is an agent in the workplace?
[00:14:50] Speaker B: So that is such a great question. And that's why the training and awareness can be really important. And this really links quite closely into cybersecurity training, because even if an organization itself is not deliberately deploying agents, agents are coming at you from the outside.
And we opened up by me talking about these two AI agents who called me just recently, just in the last couple of weeks.
[00:15:15] Speaker C: Well, let's just touch on that. So why. What was the purpose? Why were they calling you? So I can't imagine you're going to start providing legal services to a bot.
[00:15:25] Speaker B: No. Well, not yet.
[00:15:26] Speaker C: Not yet. But what. What else was sitting behind that? Was there something malicious about that approach? Because I know in times gone by, we've had all the advertising and awareness campaigns about being called by a telemarketer and not saying words like, yes, no, your name really clearly, because those are then recorded. And then when they call the banks and they. There's name or voice recognition or verification, then those words have replayed back to sound like it's you. Now, those are sort of archaic scams, I'd call that. At this point, given we've got agents, do you think there was something malicious about the agent that contacted you?
[00:16:09] Speaker B: Truth is, I don't know, but I wasn't going to take the risk. And there was another point, too. There's a part of me that feels like I am becoming an unpaid employee for many organizations because I'm continually doing unpaid work for them. I'm filling out forms for them, I'm answering questions for them and so on. And whatever the organization was that deployed those AI agents, that they weren't even prepared to pay for a human to call me. And there was a part of me that was a little bit unhappy about that aspect, a very human part of me. And I think that's one of the reasons it's great to sit down and have this conversation with you, Michael, because at the heart of everything we're talking about is our humanity.
[00:16:56] Speaker C: And this is something that we are going to be unpacking in future episodes. I think you. You and I have had a discussion about this, and what we really want out of this series is an honest discussion about what it means to be introducing agents into the workplace, what it means for the humans who are there now or who are then benched on the sidelines because these agents are in place, and how do they achieve satisfaction through life.
[00:17:27] Speaker B: And that's why I think talking now about the hybrid workplace is a very important discussion, because there exists a window where we as humans have the capability to shape what the future is that we want to be living in.
And as you and I always say, Michael, the best way to predict the future is to create it.
[00:17:53] Speaker C: Well, I think that's an excellent point for us to leave today's session on. And we look forward to joining everyone in future podcast series where we really get stuck into the detail on this.
[00:18:05] Speaker B: Thanks, Michael. And thanks, everyone.
