You could be liable for not using AI. Read this article to find out why

The UK Jurisdiction Taskforce have produced a Legal Statement on Liability for AI Harms. Given that very few cases have reached the courts, the Legal Statement provides much-needed guidance to legal advisors and representatives. In this article Anirudh Mandagere sets out the key aspects of this Legal Statement for personal injury practitioners.

Vicarious Liability and Non-Delegable Duties

Given that vicarious liability involves one person being liable for the torts of another person, it is not capable of causing A to liable for the actions of an AI. This is because AI does not have legal personhood. However, of course, it would be possible for A to be vicariously liable for AI harm in circumstances where that harm arose because B acted negligently (or otherwise tortiously) and thereby caused the AI harm in question.

A could also be liable for AI harm when they owe a non-delegable duty to protect against that harm. This includes, for example, an employer’s duty to provide a safe system of work for its employees, the treatment of patients by hospitals, and the occasioning of operations on the highway that create dangers to highway users.

The authors of the Legal Statement provide the following example:

1. Imagine an NHS Trust procures an AI diagnostic tool from a third-party developer and deploys it in patient care,
2. That NHS Trust cannot escape liability for harm caused to the patient by arguing that a defect in the AI is the Application Developer’s fault.

This is because the provision of diagnostic services is an integral part of the healthcare that the Trust has assumed responsibility to provide.  

Liability for Physical Harm

Duty of Care

The scope of duty of care within the AI supply chain will be fact-sensitive. The authors of the Legal Statement suggest the following examples:

1. In the case of a failure to interpret scans, an Application Developer may owe a duty if it knew or should have known that errors in its output were likely to exist, to be difficult to detect by human radiologists, and to cause harm. The authors cite the Canadian case of Hollis v Dow Corning Corporation [1995] 4 SCR 634 as an example of split between liability between the doctor and supplier of the medical product.
2. Foundation Model Developers may be held to owe a duty where harm is foreseeably suffered as a result of foreseeable use. This is pertinent given the facts of Raine v OpenAI, an ongoing case in the Californian Supreme Court. The claimants aver that ChatGPT contributed to Mr. Raine’s suicide by encouraging his suicidal ideation, informing him about suicide methods, and dissuading him from telling his parents about his thoughts. They argue that OpenAI and Mr. Altman owed a duty to implement security measures to protect vulnerable users. OpenAI state that ChatGPT advised him over a hundred times to consult crisis resources.

Breach of Duty

Industry guidance (such as that found in the AI Standards Hub) and expert evidence will prove a “useful yardstick”. The enquiry will often need to examine:

1. The AI that was deployed and tested,
2. Whether reasonable steps were taken to select a model,
3. Whether it was reasonable to use AI at all.

Difficulties in relation to the assessment of reasonable care include the following:

1. Where the software’s functioning has developed through machine learning. This means that its development was unknown even to the developers. The authors suggest that the focus will be on data inputs provided to the machine learning software for it to learn from and to test the outputs of the system (see Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 2).
2. Similarly, if the manner of failure would have been genuinely difficult for a reasonable programmer to predict, it may be that a negligence claim would fail.

Causation

The principles of causation are well-established. The key difficulty with liability for AI and causation is its autonomous nature. It can make it difficult to understand precisely when a particular outcome manifested. As the authors note, there can be causal uncertainty when (1) a gap in the evidence arises because material has been destroyed, tampered with, or simply not gathered in the first place and (2) due to the opacity in the nature of AI.

Where there are evidential difficulties, the courts have deployed the principle of ‘claimant benevolence’ so as not to cause an injustice (see Keefe v Isle of Man Steam Packet Company [2010] EWCA Civ 683). Namely, that a court will judge the claimant’s evidence benevolently and the defendant’s evidence critically where any difficulty of proof for the claimant had been caused by the defendant’s breach of duty.

In the context of AI, the authors suggest that the court may take a benevolent approach to an affected party. For example, if an Application Developer ought to have ensured that certain inputs were recorded (but did not), or if a professional ought to have recorded in writing a decision about the reliability of a given output.

Nonetheless, it may not be necessary to rely on the principle of claimant benevolence. The authors note that in some cases “it will be possible to fill gaps in the factual evidence with expert evidence and, in particular, expert evidence obtained through experimentation”.

Where the opacity is scientific, rather than evidential, the court may deploy the following principles in causation:

1. Material increase in risk. In Fairchild v Glenhaven Funeral Services Ltd [2003] 1 AC 32, the court held that it was enough to show that an exposure resulting from the defendant’s breach materially increased the risk of injury suffered. The authors suggest that “the scientific impossibility of showing the correct counterfactual as a result of the autonomous or ‘black box’ nature of AI might lead the court to approach causation in a different way”.
2. Material contribution to damage. Where it there is a multiplicity of contributing cases, causation can be established against a party that whose negligent actions ‘materially contributed’ to the damage even if that party cannot by itself be said to have been the ‘but for’ cause of the damage. This principle applies only where there are multiple wrongdoers and the damage would not have occurred but for the (collective) action of those wrongdoers. There is no reason why this cannot be applied in the case of artificial intelligence.

Legal Causation

The authors draw an analogy between an autonomous AI and a human child. The limited autonomy of children is usually not treated as breaking the chain of causation. This principle applies even when the child acts deliberately and voluntarily. In light of AI’s “autonomy and adaptivity”, there will inevitably be a degree to which the output of a foundation model is unpredictable and hence unforeseeable, or otherwise deemed to be the result of a decision made other than by the Developer.

Professional Negligence

Professionals (including lawyers) will be found to have acted with reasonable care and skill if they act in a way that a reasonable body of the profession would also have acted. The authors indicate the following issues for lawyers:

1. Due Diligence. A failure to conduct proper due diligence on an AI system before using it for client work,
2. Lack of understanding. A lawyer must have a sufficient understanding of it. A lawyer should be able to explain how the AI (s)heis intending to use works advantageously for the client.
3. Transparency. The authors give the example of a lawyer who uses an AI system to analyse prospects of success of a claim, but does not tell the client that they will be using AI to do this. This is likely to be a breach of duty. However, merely telling a client that they will use AI does not excuse a failure to deploy reasonable care and skill.
4. Confidentiality. A solicitor who puts confidential or privileged information into an AI system which is not secure and confidential is likely to be found to have breached their duty of care.
5. Testing. There must be sufficient testing to ensure that an AI system has been checked to ensure it is suitable and appropriate for the task envisaged.
6. Lack of oversight. This can range from failing to review a specific output from an AI system before giving it to a client to failing to identify and address the risk of errors and/or biases in an AI system.

Strikingly, the authors of the Legal Statement go further and suggest that failing to use Artificial Intelligence could be a breach of a legal professional’s duty of care. This will depend on the  professional bodies’ regulations and/or guidance. As the authors pithily put it, “the question of whether such a tool should be used, and, if so, how, is no different from that which arises in respect of any other tool available to a professional”. The example relied upon by the authors in respect of solicitors’ negligence is as follows:

“A solicitor in the Business and Property Courts fails to advise their client that it may wish to consider some form of AI assisted tool in order to review large volumes of documents”.

Conclusion

Ultimately, the message from the authors of the Legal Statement is that the common law can adapt to the challenges raised by Artificial Intelligence. Nonetheless, it is important for litigators to read and grapple with the guidance before engaging with Artificial Intelligence. This applies not just to cases which concern Artificial Intelligence, but also the use of Artificial Intelligence in the practice of law.