“Danger” is a loaded term. It facilely conjures an emotional response that is analytically caustic, especially when it comes to AI. Yet, the term is casually thrown around, in similar vein to the careless tagging of an application as being AI when it really is just an automation of some task. Rather than launch into a definition-centric analysis, it is an interesting exercise to think through what it means, from a legal perspective (as opposed to a pedestrian one) to say that an AI application is “unreasonably dangerous.”

As is often the case in law, lawyers look to other applications in which the term is used. For this we turn to product liability and here we see the “unreasonable per se” theory. A product is deemed to be just that if a reasonable person would conclude that the danger-in-fact of the product, whether foreseeable or not, outweighs its utility. Replace the word “product” with “AI” and we can begin to size this definition, see how it fits.

Well, at first glance, the fit appears to be poor. For starters, the use of the “reasonable person” standard seems much too awkward when it comes to AI. There are (at least) two main reasons for that: (1) AI is not a platform that is conducive to a traditional reasonable person prism, it is much too complex, and (2) AI’s illusive definitional quality dissolves the cohesive nature of this standard, which functions well when it comes to a tangible product. It is premature, however, to jettison this theory altogether, but these reasons highlight the conceptual difficulty in using it. So if the unreasonable per se theory is not suitable to AI, a different one is necessary. Let’s take a look at what that looks like.

The foundations of the definition I propose here are based on four variables:

(1) Application – the intended use/vertical;

(2) Complexity – computational power/resource consumption/reliability;

(3) Sophistication – of the end-user; and

(4) Presence or absence of explainability – outcome-predictability and overall reliability enhancing features.

Some applications can impact human life and death, such as those that are used in healthcare or in the military. The more complex these AI applications are, the more important it becomes that the end user be sophisticated, licensed (as a matter of operation), and that the AI have an explainability interface. Once the AI application is slotted into such a sensitive operational segment, the other factors either expand or contract its risk profile. With this we can see, for example, that an “unreasonably dangerous” title would properly attach when a given AI application belongs to a life and death type use/vertical, is complex in that it is computationally intensive and its reliability, for instance, must be carefully supervised by a sophisticated, licensed end-user, yet it lacks an explainability feature.

These variables can be used with other applications, and the result will be instructive as to whether the AI in question is or isn’t “unreasonably dangerous.”

***Postscript***

March 18, 2022: The OECD recently released a “Framework for the Classification of AI Systems.” It is designed to examine AI risks that are deemed “typical” of AI and centers on bias, explainability and robustness. My view is that this framework is much too high level and falls short of its aim to help the development of AI policies and regulations. In the context of aiding in effective management of unreasonably dangerous AI, for example, it would fail.