Insuring AI - EP1: Liability for AI harms

WEBVTT
00:00:12.240 --> 00:00:19.760
Hello and welcome to the first episode of an exciting new podcast series from our insurance and professional risks team here at HSF Kramer.
00:00:20.160 --> 00:00:23.120
I'm Greg Anderson and I'm a partner in the team.
00:00:23.440 --> 00:00:32.159
This new series, Insuring AI, focuses on an issue that's getting more and more airtime, almost more than anything else at the moment.
00:00:32.399 --> 00:00:34.240
Artificial intelligence.
00:00:35.039 --> 00:00:49.359
AI isn't a niche topic, not a passing trend, it's a genuinely multifaceted phenomenon that's reshaping the way we live, the way businesses operate, and critically the way risk is created, assessed, and insured.
00:00:49.600 --> 00:00:54.640
That's why we felt now was the right time to dedicate a focus series to it.
00:00:55.920 --> 00:01:15.359
Over the coming episodes, we're going to be exploring a broad range of questions that AI is throwing up for the insurance world, including who bears responsibility when an AI system causes harm, whether parametric insurance has a role to play, data centres, which are the physical backbone of the AI economy, and the particular risks they present.
00:01:15.599 --> 00:01:19.120
And we'll be asking one of the questions we hear most from our clients right now.
00:01:19.439 --> 00:01:22.640
Is AI actually covered under existing insurance policies?
00:01:22.719 --> 00:01:25.680
And if so, under which ones and in what circumstances?
00:01:26.239 --> 00:01:34.560
These are live issues landing on the desks of both us and those in the insurance market every day, and our aim in this series is to help you navigate them.
00:01:34.959 --> 00:01:36.879
So now to our first episode.
00:01:37.120 --> 00:01:46.560
We couldn't think of a better place to start this series than with a document that landed in early July and has already generated significant discussion across the legal and insurance communities.
00:01:46.719 --> 00:01:55.840
I'm referring, of course, to the legal statement and liability for AI harms that have been published by the UK Jurisdiction Task Force, or UK JT for short.
00:01:56.159 --> 00:02:05.359
So at its core, that's a statement that attempts to set out a clear legal framework for how liability is determined when AI causes harm.
00:02:05.599 --> 00:02:19.840
And in particular, it grapples with one of the thorniest questions in this space: whether and in what circumstances those who have not set out deliberately to cause harm may be liable for harms resulting from the use of AI in any event.
00:02:20.080 --> 00:02:24.479
Now I should warn you, it's not a short document, 130 pages in fact.
00:02:24.639 --> 00:02:28.560
It's detailed, it's comprehensive, and it covers a lot of ground.
00:02:28.879 --> 00:02:38.159
Fortunately, I'm joined today by Will Gibson, a senior associate here at HSF Kramer, who has in fact read it cover to cover and is very well placed to take us through it.
00:02:38.319 --> 00:02:39.360
Will welcome.
00:02:39.680 --> 00:02:40.240
Thanks, Greg.
00:02:40.479 --> 00:02:41.199
Great to be here.
00:02:41.360 --> 00:02:44.560
Yes, I have read it, but I won't pretend it was a quick afternoon's reading.
00:02:45.120 --> 00:02:45.759
I'm sure.
00:02:45.919 --> 00:02:47.439
So let's start at the beginning.
00:02:47.520 --> 00:02:50.000
Uh, what's the document and where does it come from?
00:02:50.400 --> 00:02:58.960
Yes, so the legal statement on liability for AI harms, to use its full name, is a publication, as you said, uh, by the UK jurisdiction task force.
00:02:59.120 --> 00:03:07.199
And I think it's worth spending a moment on what the UKJT actually is because it helps explain why this document carries the weight that it does.
00:03:07.439 --> 00:03:20.639
So the UKJT is an expert, industry-led body with clear mission, and that's to promote the use of English law and the UK's legal framework is the jurisdiction of choice for technology and digital innovation.
00:03:20.960 --> 00:03:31.120
It operates under the umbrella of LawTech UK, which is a government-backed initiative supporting the modernisation international competitiveness of the UK's legal sector.
00:03:31.360 --> 00:03:47.759
So there's a very deliberate policy and commercial objective behind this document, and that is to ensure that English law keeps pace with technological change and that market participants around the world should continue to choose English law to govern the most complex and cutting-edge transactions and disputes.
00:03:48.240 --> 00:03:51.680
It's done this kind of work before, hasn't it, in other areas of tech?
00:03:52.240 --> 00:03:53.120
Um, exactly, yes.
00:03:53.199 --> 00:04:02.960
So the UKGT has previous form and space, it's published legal statements on crypto assets, for example, uh distributed ledge technology, smart contracts, digital securities.
00:04:03.039 --> 00:04:13.520
So each of these times stepping into areas of genuine legal uncertainty and hoping to provide and providing, in fact, a considered framework for how existing legal principles of English law are applied.
00:04:13.680 --> 00:04:16.639
And naturally, AI is the next frontier.
00:04:17.040 --> 00:04:24.240
And I think it's really important to be clear about the status of these legal statements because they're sometimes a misconception.
00:04:24.639 --> 00:04:30.399
They're not legislation, they don't create new law, but they are quite powerful in their own right.
00:04:30.560 --> 00:04:45.199
They're authoritative statements on how existing law applies to emerging technology, drafted by people with the seniority, expertise, and critically the judicial involvement to ensure that the courts, market participants can draw real confidence in their conclusions.
00:04:47.439 --> 00:04:50.800
And so when the UK JT speaks, people should listen, right?
00:04:51.120 --> 00:04:52.800
Yes, yes, Greg, they really should.
00:04:52.879 --> 00:05:01.279
And in a fast-moving area like AI, where waiting for legislation or litigation to catch up is simply not an option for businesses making real decisions today.
00:05:01.439 --> 00:05:06.399
And this kind of document provides something close to an authoritative legal roadmap.
00:05:06.560 --> 00:05:11.600
It's not binding in the way the judgment is, but it carries genuine persuasive authority.
00:05:11.920 --> 00:05:12.959
How did it come about?
00:05:13.920 --> 00:05:20.000
Well, there was a proper consultation process, which adds further to the document's credibility.
00:05:20.160 --> 00:05:24.879
The UK JT didn't sit and write sit down and write this in in isolation.
00:05:25.040 --> 00:05:34.639
Instead, it engaged with a wide range of stakeholders across law, technology, industry, and academia, seeking input on the key questions before arriving at its conclusions.
00:05:34.879 --> 00:05:38.160
And for example, we at HSF Crayum are fed into this document.
00:05:38.560 --> 00:05:42.000
And can you give us a bit of a sense of the terrain that it covers?
00:05:42.959 --> 00:05:43.759
Yes, briefly.
00:05:43.839 --> 00:05:51.040
So the statement's obviously a substantial piece of work and it's helpfully structured in the way that takes one on a logical journey through the issues.
00:05:51.199 --> 00:06:02.639
Um, it begins, as you would expect, with an introduction on what AI is, seeking to define it, and that exercise matters more than you might think, um, given it encompasses such a wide spectrum of systems.
00:06:03.120 --> 00:06:10.639
From there, the document moves to the heart of the matter, as you've outlined, liability for non-deliberate harm caused by the use of AI.
00:06:11.040 --> 00:06:12.639
And that framing is important.
00:06:13.120 --> 00:06:18.720
It's focused on the common scenario where AI causes harm, not as a result of anyone's deliberate wrongdoing.
00:06:18.959 --> 00:06:26.079
And as we can all appreciate, and as we all have experience, unintended harm is arguably the defining feature of AI risks.
00:06:26.240 --> 00:06:34.800
And as we all know, the technology has a well-documented propensity for hallucination, producing outputs that are confidently stated but entirely wrong.
00:06:34.959 --> 00:06:38.959
And we're still at the early stages of people truly understanding its limitations.
00:06:40.000 --> 00:06:49.120
Having covered this, the statement then focused on liability arising out of negligence, defective products, and false statements, stepping through each of these requirements for liability in turn.
00:06:49.279 --> 00:06:56.800
On negligence, it works through the building blocks, so duty of care, breach causation, and we're going to look at them shortly.
00:06:57.120 --> 00:07:08.240
On defective products, it examines how existing product liability frameworks designed with physical goods in mind translates to a world where the product may be a software model or AI generated output.
00:07:08.480 --> 00:07:20.959
And on false statements, it tackles the increasingly live question of what happens when an AI system, so a chatbot, for example, produces information that is simply wrong and someone relies on it to their detriment.
00:07:21.199 --> 00:07:26.879
Each of these areas raises its own distinct legal questions that are grappled with in this statement.
00:07:29.279 --> 00:07:35.519
So as I've understood it, the focus is on uh tortuous and statutory claims arising out of AI.
00:07:35.680 --> 00:07:38.160
But where does contractual liability then fit in?
00:07:38.240 --> 00:07:41.519
Why does the UK JT focus its attention elsewhere?
00:07:42.000 --> 00:07:44.720
And yes, Greg, that's exactly the right question to ask.
00:07:44.800 --> 00:07:50.240
And it goes to the heart of why document's useful to the insurance market specifically.
00:07:50.720 --> 00:07:55.759
So, as you say, contract really is the foundation for most commercial relationships involving AI.
00:07:55.920 --> 00:07:59.680
And the statement acknowledges that up front and very clearly.
00:08:00.000 --> 00:08:13.199
For the most part, participants in the AI supply chain, and we're talking here about data providers, model developers, application developers, and end users, and the contract will be the primary method of risk allocation.
00:08:13.439 --> 00:08:21.279
How liability is apportioned between those parties will, in most cases, be determined by whatever contractual arrangements they put in place between them.
00:08:22.240 --> 00:08:25.680
I mean, my experience is the contract's the first and last protocol.
00:08:26.959 --> 00:08:28.480
For most claims, yes.
00:08:28.639 --> 00:08:32.960
If something goes wrong, the starting point will be um what does the contract say?
00:08:33.120 --> 00:08:35.519
Who's accepted responsibility for what?
00:08:35.679 --> 00:08:38.159
What limitations and exclusions have been applied.
00:08:38.480 --> 00:08:52.879
And that's where the analysis will begin and often end, which is why getting those contractual arrangements right in the first place is so important for everyone in that chain, from the developer building the underlying model through to the businesses deploying it and the end user relying on it.
00:08:53.600 --> 00:08:59.759
But, and this is where the statement's focus on tort is so significant: contract is not the whole picture.
00:09:00.000 --> 00:09:07.519
And from an insurance perspective, in particular, the distinction between contractual and tortuous liability is critically important.
00:09:08.480 --> 00:09:19.360
As you know, many liability policies contain exclusions that carve out cover for liability, which has been assumed by the insured under the terms of a contract and would not have otherwise arisen in tort.
00:09:20.799 --> 00:09:31.759
In other words, if you have contractually accepted responsibility for something in circumstances where you would not otherwise be liable, your liability insurance policy may not respond to that.
00:09:32.000 --> 00:09:37.200
And what the UKJT is mapping out is the liability exists independent of any contractual arrangement.
00:09:37.360 --> 00:09:40.159
The liability that arises simply by operation of law.
00:09:40.240 --> 00:09:46.000
And it's precisely that kind of liability that a liability insurance policy is designed to cover.
00:09:46.240 --> 00:09:56.000
So for those in the insurance sector, this statement is in many ways a guide to how liability insurance claims arising from AI are going to be framed both now and in the future.
00:09:56.879 --> 00:10:04.240
So with contracts setting a commercial backdrop, why don't we turn to the real complexity issues, the tortuous claims themselves?
00:10:04.480 --> 00:10:07.919
And we're going to focus for the rest of the podcast on negligence claims.
00:10:08.080 --> 00:10:09.360
Can you set the scene a bit?
00:10:10.240 --> 00:10:10.879
Of course.
00:10:11.039 --> 00:10:18.960
And I think it's worth being clear up front that the statement's not attempting to identify every possible scenario in which AI might give rise to a tortuous uh claim.
00:10:19.120 --> 00:10:25.279
Um, it's focused, deliberately so, on situations where liability is most likely to arise in practice.
00:10:25.519 --> 00:10:31.279
And in the context of negligence, the statement focuses on liability for physical damage and economic damage.
00:10:31.519 --> 00:10:44.879
Think negligently designed AI, robotics on a factory floor causing physical damage, doctors being liable for clinical negligence for using AI systems, or accountants relying too heavily on AI and giving negligent advice to their clients.
00:10:46.240 --> 00:10:49.600
You said earlier the statement doesn't create new law.
00:10:50.080 --> 00:10:55.840
So what's the value in working through those familiar principles in an AI context?
00:10:56.159 --> 00:10:58.080
Yeah, that's a very important point.
00:10:58.320 --> 00:11:05.919
Um, but what the statement does, and it does very well, is demonstrate that the flexibility and adaptability of the English common law system.
00:11:06.240 --> 00:11:10.879
It doesn't need Parliament to pass legislation before the courts can get to grips with these claims.
00:11:11.120 --> 00:11:17.200
The existing English law principles are capable of doing the work, and they just need to be applied to a new factual context.
00:11:17.360 --> 00:11:20.320
And that's precisely what this statement is illustrating.
00:11:20.799 --> 00:11:39.360
So stepping through the requirements of a negligence claim on the first duty of care, the statement reaffirms the established three-part framework that practitioners will know very well: foreseeability of harm, proximity between the parties, and the question of whether it is fair, just and reasonable to impose a duty in all the circumstances.
00:11:39.600 --> 00:11:41.840
None of these requirements have changed.
00:11:42.399 --> 00:11:47.679
And as such, in many AI-related scenarios, the analysis will be relatively straightforward.
00:11:47.840 --> 00:11:54.320
For example, a manufacturer incorporates AI into a product, clearly owes a duty of care to the end user who purchases it.
00:11:54.799 --> 00:12:01.600
A professional who deploys AI in the course of delivering advice or services to a client clearly owes their client a duty of care.
00:12:01.840 --> 00:12:03.840
Again, uncontroversial.
00:12:04.320 --> 00:12:17.120
However, where things do get more complicated and where the statement is careful to flag genuine uncertainty is in relation to foundation model developers and application developers further up the supply chain.
00:12:17.360 --> 00:12:22.960
Do they owe a duty of care directly to end users who have no contractual relationship?
00:12:23.200 --> 00:12:25.200
Well, the statement doesn't rule it out.
00:12:25.519 --> 00:12:33.600
There may be circumstances, particularly where the end user is sufficiently identifiable and foreseeable, where a court could find that such a duty exists.
00:12:33.759 --> 00:12:37.279
But the statement is clear, this is deeply fact-specific.
00:12:37.519 --> 00:12:51.519
The nature of the AI system, the context of its deployment, the degree to which the developer had visibility, how and by whom it would be used, and the nature of the harm suffered, all of these will bear on the analysis.
00:12:51.759 --> 00:12:58.240
Anyone considering these issues should be careful about reaching any easy generalizations and conclusions.
00:12:59.039 --> 00:13:00.559
But what about the standard of care?
00:13:00.639 --> 00:13:06.000
What does reasonable scale and care actually look like in an AI context?
00:13:06.320 --> 00:13:10.960
Yeah, so this is where I think the statement makes one of its most uh important contributions.
00:13:11.600 --> 00:13:15.360
The standard of reasonable care is not fixed, static benchmark.
00:13:15.679 --> 00:13:22.480
It's measured against what reasonable, competent practitioner in the relevant field would do at the relevant time.
00:13:22.720 --> 00:13:30.799
And the crucial words here are relevant time, because the standard involves is technology, knowledge, and practice to evolve.
00:13:31.039 --> 00:13:33.840
And with AI, that creates a real challenge.
00:13:34.080 --> 00:13:40.000
The technology is developing at pace, and many industries and professions are struggling to keep up with that pace.
00:13:40.639 --> 00:13:47.440
What was considered a perfectly reasonable approach to using AI a year ago may not meet that standard today.
00:13:47.840 --> 00:14:01.840
And the statement explicitly recognizes this and places a positive obligation on professionals to stay current, to engage actively with developments in their field, and not simply to assume what they were doing last year is still good enough.
00:14:02.720 --> 00:14:10.000
And the statement's actually usefully specific about what it's required in practice, and it gives a very helpful list.
00:14:10.639 --> 00:14:14.320
It's not simply a matter of using AI and hoping for the best.
00:14:14.559 --> 00:14:24.720
It says that professionals need to understand the limitations of the systems they're using, what they were designed to do, where they are prone to error, where their outputs cannot be relied on.
00:14:24.960 --> 00:14:30.879
It says that professionals need to test and verify outputs rather than treating them as authoritative.
00:14:31.120 --> 00:14:37.600
They need to be transparent with their clients and customers about the fact that AI is being used and what role it plays.
00:14:37.759 --> 00:14:44.960
They need to maintain appropriate oversight, human oversight, rather than treating AI as a substitute for professional judgment.
00:14:45.039 --> 00:14:49.440
And they need to keep records, the evidence all that these steps have been done.
00:14:49.600 --> 00:14:53.360
I mean, good record keeping here is going to be essential.
00:14:54.799 --> 00:15:11.519
If you go back 50 years, lawyers use books, paper, pens, and I wonder if some of our listeners might be thinking um whether if AI creates all this exposure, maybe the safest way is to go back to books and paper and pens.
00:15:11.759 --> 00:15:14.000
But uh, is that a fair option?
00:15:14.399 --> 00:15:16.480
Um unfortunately not.
00:15:16.639 --> 00:15:26.559
I appreciate it's a very uh attractive position for some, perhaps, but the statement rather decisively closes off uh that escape route.
00:15:26.720 --> 00:15:36.639
Because one of the most striking conclusions of the document is that in certain circumstances, the duty of skill and care may actually require a professional to use AI.
00:15:37.039 --> 00:15:42.000
Not just permit it or allow it, but actually require you to use AI.
00:15:43.039 --> 00:15:44.240
So that's disappointing.
00:15:44.320 --> 00:15:46.720
So you can be liable for not using AI.
00:15:47.120 --> 00:15:48.080
Potentially, yes.
00:15:48.240 --> 00:15:51.600
And the statement gives some really striking examples to bring this to life.
00:15:52.000 --> 00:16:01.360
It gives an example of a radiologist who failed to use an AI system that's highly effective at identifying tumours and could have been procured at reasonable top cost.
00:16:01.679 --> 00:16:08.879
Um it gives another example of an auditor who fails to deploy AI to detect anomalies and fraud in the businesses they're auditing.
00:16:09.039 --> 00:16:12.960
And these aren't, as you all appreciate, uh hypothetical examples.
00:16:13.039 --> 00:16:21.840
That they're entirely plausible scenarios arising right now in fields where AI tools are already available, already proven and increasingly expected.
00:16:22.399 --> 00:16:26.000
So a really important message for professionals and their insurers.
00:16:26.159 --> 00:16:31.679
Um, we've talked about duty of care and the standard of care expected of professionals using AI.
00:16:31.840 --> 00:16:38.639
But even if you can establish a duty and a breach, you still have to prove the breach actually caused harm.
00:16:38.799 --> 00:16:44.799
And I understand that causation is one of the areas where AI creates some particular thorny legal questions.
00:16:45.120 --> 00:16:46.080
Um yes, it does.
00:16:46.159 --> 00:16:49.759
And it's one of the areas I find most interesting in the statement.
00:16:50.159 --> 00:17:00.799
But let's start with what is hopefully a reassuring point for listeners, and that is as with duty and breach, the legal principles governing causation haven't changed.
00:17:01.039 --> 00:17:04.559
The statement applies existing law to a new factual context.
00:17:04.640 --> 00:17:11.920
And for a significant proportion of AI-related negligence claims, that application is actually reasonably straightforward.
00:17:12.240 --> 00:17:18.079
Um, so as listeners would be familiar, the causation test and the negligence claims remains the butt-for test.
00:17:18.319 --> 00:17:22.160
Would the claimant have suffered harm but for the defendant's breach of duty?
00:17:22.480 --> 00:17:30.160
The answer is yes, if the harm would have occurred regardless of what the defendant did or failed to do, then causation is not established and the claimant fails.
00:17:30.480 --> 00:17:37.839
The answer is no, if the harm would not have occurred but for the breach, then causation is made out and you move to the question of damages.
00:17:39.759 --> 00:17:44.240
And so in many AI cases, that analysis is likely to be relatively clear, isn't it?
00:17:44.480 --> 00:17:45.119
Um of course.
00:17:45.200 --> 00:17:50.160
So in many cases, and the statement gives some examples, uh, it's going to be very obvious.
00:17:50.400 --> 00:18:00.240
Take an example uh uh of an architect who uses AI for a task which was actually simply not appropriate to use it for, and as a result, a third party suffers harm.
00:18:00.400 --> 00:18:02.640
The butt for analysis is actually quite tractable.
00:18:02.880 --> 00:18:12.160
You ask what the architect exercising reasonable skill and care should have done had they not used AI at all, and and that wouldn't have resulted in any loss and um causation.
00:18:12.240 --> 00:18:14.480
It's very straightforward in that instance.
00:18:14.799 --> 00:18:16.559
Why does it get complicated then?
00:18:17.279 --> 00:18:23.200
Well, it gets more complicated as you move up the AI development chain, um, as I outlined earlier.
00:18:23.359 --> 00:18:27.599
And and I think this is the point that has real practical significance for the insurance market.
00:18:28.079 --> 00:18:39.200
So, so far we've been talking about um liability for those at the end of the chain, so the manufacturers that incorporated AI into their products, the professionals who are using AI tools in practice.
00:18:39.440 --> 00:18:55.279
But what about the foundation model developers at the top, the anthropics and open AIs of the world, the organizations that built these AI systems we're all using in the first place, or the application developers who use these AI systems and built software lay on top of them?
00:18:55.359 --> 00:18:57.680
And they could potentially be drawn into litigation.
00:18:57.839 --> 00:19:10.720
For example, the manufacturer professional at the end of a chain is not solvent, or via a contribution claim where a defendant seeks to spread liability along the supply chain up to the foundation model developers or application developers.
00:19:11.200 --> 00:19:14.480
Yeah, I can see causation might be tricky in cases like that.
00:19:14.880 --> 00:19:29.119
Yes, because if you're trying to establish that a foundation model developer was negligent, perhaps through inadequate training of the model, insufficient controls, or failure to put in place appropriate guardrails, you're faced with a profound evidential challenge.
00:19:29.279 --> 00:19:35.759
And that is to satisfy the but four test, you need to show what AI would have done absent the alleged negligence.
00:19:35.920 --> 00:19:40.640
And here you run headlong into what is commonly known as the black box problem.
00:19:41.119 --> 00:19:43.599
I'm sure everyone's desperate to know what that is.
00:19:44.400 --> 00:19:56.640
Well, it's the fact that the internal workings of sophisticated AI models, particularly large language models and other deep learning systems, are simply not transparent, even to the people that built them in the first place.
00:19:57.200 --> 00:20:07.920
Um you can observe what goes in, what comes out of these systems, what happens in between, the process by which the model arrives at its output is in many cases genuinely opaque.
00:20:08.480 --> 00:20:13.839
There's no clear chain of reasoning you can trace in the way that you might trace a calculation on the spreadsheet.
00:20:13.920 --> 00:20:23.039
And the model's been trained on such vast quantities of data and the relationships it's learnt that no human could can fully interrogate them.
00:20:23.279 --> 00:20:30.079
So when you ask the question, what would this AI model have done but for the negligence, this legal question we're all so familiar?
00:20:30.400 --> 00:20:37.440
You're asking a question that the current technology may simply not be able to answer or it finds extremely challenging.
00:20:37.920 --> 00:20:43.440
That sounds like it could be quite a significant obstacle to establishing liability against developers, right?
00:20:43.680 --> 00:20:46.640
Yes, it is a significant obstacle.
00:20:46.799 --> 00:20:52.000
Uh, but the statement makes the important point that it's not necessarily an insurmountable one.
00:20:52.160 --> 00:20:58.160
And this is where English law and English courts' existing approach to causation becomes really relevant.
00:20:58.400 --> 00:21:13.680
So the statement suggests it may be possible to establish what AI would have done uh but for the negligence uh through through expert evidence, um, evidence from AI specialists, data scientists, or others, those with deep technical knowledge of the systems in question.
00:21:13.920 --> 00:21:18.799
And English courts are, as we all know, accustomed to relying on expert evidence to establish facts.
00:21:19.039 --> 00:21:23.200
There's no reason in principle why AI systems should be any different.
00:21:23.440 --> 00:21:30.160
It will be extremely challenging for the reasons I've outlined earlier, but not necessarily utterly impossible.
00:21:30.640 --> 00:21:43.279
In my experience of the English courts, is that typically they try to apply a common sense approach, and historically at least, have taken a flexible approach in cases where causalation is difficult to establish with precision.
00:21:43.680 --> 00:21:47.440
Yes, they have, and the statement draws some important precedents here.
00:21:47.599 --> 00:21:59.200
Um English courts have, in a series of cases, recognised that the strict application of the butt four test can produce unjust results where there are evidential difficulties, as we've outlined here.
00:21:59.359 --> 00:22:07.200
Um What you know, where the the evidence is not as a result of the claimant's um failings, but inherent in the nature of the harm that was caused to them.
00:22:07.519 --> 00:22:12.240
In those circumstances, the courts have been prepared to adopt a more generous approach.
00:22:12.480 --> 00:22:27.680
One example uh was a Court of Appeal case um flagged in the statement, and that's involved an employee who suffered hearing loss, but could not show that they suffered repeated exposure to excess noise levels uh because their employer had breached its underlying duty to monitor those levels in the first place.
00:22:27.839 --> 00:22:31.359
There weren't the documents to show that those noise levels had actually been exceeded.
00:22:31.759 --> 00:22:44.319
Here the Court of Appeal held that the court should judge the claimant's evidence benevolently and the defendant's evidence critically, without reversing the burden of proof on the basis that the defendant's own breach had created the this evidential difficulty.
00:22:45.440 --> 00:22:46.720
And then there's Fairchild.
00:22:47.759 --> 00:22:48.720
Yes, Fairchild.
00:22:48.799 --> 00:22:58.400
And that's the landmark House of Lords case from the early 2000s, which will be familiar to a number of our listeners for its knock-on consequences for the wider insurance market.
00:22:59.119 --> 00:23:04.559
This is the case the statement highlights as potentially having the most significant implications for AI liability.
00:23:05.519 --> 00:23:15.920
As a reminder, Fairchild involved mesophilioma claims brought by workers who'd been exposed to asbestos by multiple different employers over the course of their working lives.
00:23:16.720 --> 00:23:25.119
The scientific difficulty here was that mesophilioma is caused by a single asbestos fibre, triggering a mutation of the body's cells.
00:23:25.599 --> 00:23:37.359
But where workers have been exposed to asbestos throughout their working lives for numerous different employers, science couldn't determine which employer's asbestos, which particular fibre had caused the triggering event.
00:23:38.160 --> 00:23:51.039
On a strict application of but four, no individual employer could have been shown to have caused disease, even though one of them must have, resulting in potentially resulting in no recovery for terminally ill claimants.
00:23:52.160 --> 00:23:53.759
And we come back to common sense again.
00:23:53.839 --> 00:23:56.640
The House of Lords refused to accept that by recollection.
00:23:57.119 --> 00:23:57.440
Yes.
00:23:57.680 --> 00:24:05.680
And instead, the House of Lords held that each employer who had materially contributed to the risk of the claimant developing metathelioma could be held liable.
00:24:05.839 --> 00:24:16.880
And that's notwithstanding the fact that it's impossible to prove which exposure had in fact caused the disease, with liability assigned in proportion to the probability that the defendant had caused the injury.
00:24:17.359 --> 00:24:24.720
It was a deliberate and principled departure by the English courts from strict causation doctrines, all in the interests of achieving justice.
00:24:30.240 --> 00:24:40.559
Indeed, and the fair child demonstrates that the English common law is capable of evolving and developing solutions to where systemic problems or causation would otherwise produce systemic injustice.
00:24:41.200 --> 00:24:53.920
Here, the scientific impossibility of establishing the correct counterfactual because of the black box nature of AI I outlined earlier is precisely the kind of evidential challenge that may lead to the courts to find a different route causation.
00:24:54.160 --> 00:24:57.839
But what that route looks like in practice remains to be seen.
00:24:58.480 --> 00:25:05.119
So I would have thought the black box isn't necessarily a get out of jail free car that developers uh might otherwise hope it is.
00:25:05.359 --> 00:25:07.119
Uh no, it might not be.
00:25:08.240 --> 00:25:11.519
And that's from an insurance perspective, why it's significant, right?
00:25:11.680 --> 00:25:18.880
Because it means that even the upstream players in the AI supply chain uh carry quite significant potential exposure.
00:25:19.200 --> 00:25:21.680
Yes, potentially very substantial exposure.
00:25:21.839 --> 00:25:29.839
And that's something that anyone underwriting technology liability, errors and emissions, or product liability risk in this space needs to be considering now.
00:25:30.079 --> 00:25:38.079
The scope of the defendants and AI liability claim is potentially much, much wider than the conventional professional negligence or product liability case.
00:25:38.720 --> 00:25:42.799
So, for my part, it's a reminder that this isn't an academic exercise.
00:25:43.039 --> 00:25:53.440
These are live underwriting decisions that are having to be made right now against policy wordings that were never designed with any of this uh in mind, however nascent one might think it is.
00:25:53.680 --> 00:25:57.519
And what's driving about the statement for me is just how much ground it covers.
00:25:57.839 --> 00:26:00.400
And we've actually managed to cover just some of that today.
00:26:00.720 --> 00:26:04.559
We've focused on negligence and causation, but there's a great deal more in there.
00:26:04.799 --> 00:26:21.279
And the question of whether a developer can be liable when a bad actor deliberately misuses their AI, a whole section on liability implications of AI chatbots producing false statements, and an important discussion on product liability under the Consumer Protection Act 1987.
00:26:21.440 --> 00:26:24.480
And we can come back to that in a future episode, I would have thought.
00:26:24.799 --> 00:26:27.440
Yes, we really um have only scratched the surface today.
00:26:27.519 --> 00:26:31.839
That as you've outlined, there's a great deal more in the statement than we've been able to cover.
00:26:32.000 --> 00:26:36.079
And we'd encourage everyone listening to um read the document in full.
00:26:36.160 --> 00:26:40.319
As I said, it's a very clear uh statement if a long one.
00:26:40.480 --> 00:26:45.440
Um and but failing that, we've produced on our uh litigation notes blog a short summary.
00:26:45.759 --> 00:26:50.559
If you want to just see the headline points in one place, it's it's a great place to start.
00:26:51.039 --> 00:26:58.559
That really does bring home what makes AI so fascinating and at sometimes uh daunting uh area for everyone in the market.
00:26:58.640 --> 00:27:04.880
It's not a single risk, it's not a single product line, it's a force that's reshaping the nature of risk itself.
00:27:05.039 --> 00:27:10.640
And that's true across every class and every sector, and the pace of change is only accelerating.
00:27:10.799 --> 00:27:20.480
The legal framework the UK JT has set out uh seems to me to be an important milestone, but it's very much the beginning of a conversation rather than the end of it.
00:27:20.640 --> 00:27:21.599
Uh so thanks, Will.
00:27:21.680 --> 00:27:26.880
There's been a lot to explore, and we're looking forward to unpacking more of the issues in the future.
00:27:27.119 --> 00:27:31.680
If you are listeners, have found today's episode uh useful, please do share it with colleagues.
00:27:31.759 --> 00:27:33.599
We're grappling with the same questions.
00:27:33.759 --> 00:27:38.079
We hope this series can be a useful resource as the landscape continues to evolve.
00:27:38.160 --> 00:27:40.240
So until next time, thank you.