European Commission proposes AI liability rules: Introduction

Background

The current liability rules, i.e. the harmonised product liability framework, and the non-harmonised (often fault-based) rules governing AI-related damage-compensation, are often taken to form an obstacle rather than an incentive for innovation. This is for instance expressed in the AI liability package itself, and argued in academic literature, including my own PhD research. For instance, where fault and causalitymust be proven in order to establish liability under those rules, victims would need extensive knowledge of the underlying self-learning (often very complex and ever-changing) algorithms in order to assess whether or not these cause an AI-application to inflict damage. This can be prohibitively difficult. Take for example future fully Autonomous Vehicles (AVs): these will have to be equipped with many sensors and communication devices to them to navigate on the road, determine driving behaviour and to prevent collisions. That driving behaviour will have to be improved and adapted continuously, through the use of self-learning algorithms. Should nonetheless an accident occur, under current rules a victim would need to prove either a defect in a product, or a fault in the driving behaviour, as well as the causal relationship between that defect or fault and the damage that occurred. As technology becomes increasingly complex as a result of the use of self-learning technology and an increasing amount factors that could play a role in the origination of an accident, this will become a herculean task.

At best, proving norm violations and causality will become expensive and time consuming, and might stand in the way of effective remuneration of damages. In turn, this results in unjustified risks for victims, and could negatively impact consumer’s (including victims) trust in AI – as I argue in my thesis. The uncertainty resulting from the application of fault-based liability rules, is seen to negatively impact the willingness of innovators to invest in AI.

Similar problems follow from the Product Liability Directive (PLD), although it is sometimes questioned whether software and algorithms qualify as ‘products’. Under the PLD, victims for instance need to prove defectiveness of a product to establish liability. Furthermore, victims need to prove the causal relationship between a defect and damage. This cannot be achieved without thorough analysis of ever more complex AI-products, and data processed therewith. At the same time, there are ample opportunities for producers to escape liability on the basis of for instance the “later existence defence” (when a defect originated after market-introduction of the product) and the “development risks defence” (when a defect could not have been discovered by the producer at the time of market-introduction). The consequence is that AI-victims are not as well protected as originally foreseen in the PLD, because also here the risks are allocated to victims rather than being fairly distributed between them and the producers. This would likely negatively impact consumer trust in the preventive and reparative capacities of the PLD.

Against this background, the EC strives with the introduction of the AI-liability package to achieve “an economic incentive to comply with safety rules and therefore contribute to preventing the occurrence of damage”, and to “promote trust in AI […] by ensuring effective that victims are effectively compensated if damage occurs […]”, by taking “a holistic approach in its AI policy” through the adaptation of the PLD, and the adaptation of national (fault based) liability rules (quotes taken from the Explanatory Memorandum of the proposed AI Liability Directive (AILD), p.3.