It is the year 2023, and for the first time, a self-driving car navigating city streets strikes and kills a pedestrian. A lawsuit is sure to follow. But exactly which laws will apply?

I am putting this question since lately in Malta we have been hearing, reading and viewing much information about artificial intelligence and its use in many spheres of society. Thus, the dilemma I have started with above is bound to crop up sooner or later.

First, it is important to establish what one is to understand by the term “artificial intelligence”. There are researchers in the AI field who consider anything that mimics human intelligence, by whatever method, to be “artificial intelligence”; there are others who think that the only “artificially intelligent” programmes are those that mimic the way in which humans think. There are also those in the field of information systems who would classify many “artificially intelligent” programmes as being complex information systems, with “true” artificial intelligence being reserved for the meta-level decision-making that is sometimes characterised as “wisdom”. But in order to have a holistic and exhaustive understanding, any computer system that is able to recognise a situation or event, and to take a decision of the form “IF this situation exists THEN recommend or take this action” should be taken to be an artificially intelligent system.

It is interesting to ponder how, and whether, artificial intelligent entities might be held criminally liable. Criminal laws normally require both an actus reus (an action) and a mens rea (a mental intent). Actus reus consists of an action or of a failure to act. Mens rea requires knowledge or being informed, but it can also require only negligence (“a reasonable person would have known”). There are then strict liability offences, for which no mens rea needs to be demonstrated – just the action or actus reus.

There could be another, yet rare, legal model by which offences committed by AI systems might be considered. This is what I like to refer to as the “innocent agent model”. Let us suppose this innocent agent is another human being, or another living creature, or even a non-living moving object, and such innocent agent is instructed by another person to do an act which in itself amounts to a criminal offence or civil tort (for example, if the owner of a dog instructed his dog to attack somebody), then the instructor is held criminally or civilly liable as the case may be. According to this model, AI programmes could be held to be an innocent agent, with either the software programmer or the user being held to be the perpetrator-via-another.

Now let us suppose that part of the AI programme which was intended for good purposes is activated inappropriately and performs a criminal action. Indeed, there was an actual case a few months ago in which a Japanese employee of a motorcycle factory was killed by an artificially intelligent robot working near him. The robot erroneously identified the employee as a threat to its mission and calculated that the most efficient way to eliminate this threat was by pushing him into an adjacent operating machine. Using its very powerful hydraulic arm, the robot smashed the surprised worker into the machine, killing him instantly and then resumed its duties.

So users or (more probably) programmers might be held legally liable if they knew that a criminal offence was a natural, probable consequence of their programmes/use of an application. The application of this principle must, however, distinguish between AI programmes that “know” that a criminal scheme is underway (that is, they have been programmed to perform a criminal scheme) and those that do not (they were programmed for another purpose, but may still be liable for civil tort if something goes wrong). It may well be that crimes, where the mens rea requires knowledge, cannot be prosecuted for the latter group of programmes (but those with a “reasonable person” mens rea or strict liability offences, can).

It is relatively simple to attribute an actus reus to an AI system. If a system takes an action that results in a criminal act or fails to take any action when there is a duty to act, then the actus reus of an offence has occurred. Assigning a mens rea is much harder and so it is here that the three levels of mens rea become important. For strict liability offences, where no intent to commit an offence is required, it may indeed be possible to hold AI programmes criminally liable. Considering the example of self-driving cars, speeding is a strict liability offence; so if a self-driving car is found to be breaking the speed limit for the road it is on, the law may well assign criminal liability to the AI programme that was driving the car at that time.  This possibility raises a number of other issues, including defences (could a programme that is malfunctioning claim a defence similar to the human defence of insanity or if it is affected by an electronic virus, could it claim defences similar to coercion or intoxication?) and punishment (who or what would be punished for an offence for which an AI system was directly liable?)

As LeCun, one of the computer scientists behind modern artificial intelligence, once said: “Artificial intelligence itself should not be targeted by regulators, but how it is used; AI systems need to be secure, that they don't put people in danger, that they respect privacy."

Be that as it may, AI systems have the potential to go haywire and wreak havoc.

Dr Mark Said is an advocate