Publication alert: The Explanations One Needs for the Explanations One Gives
Ljupcho Grozdanovski
Ljupcho GROZDANOVSKI's (long) study titled ‘The Explanations One Needs for the Explanations One Gives – The Necessity of Explainable AI (XAI) for Causal Explanations of AI-Related Harm: Deconstructing the ‘Refuge of Ignorance’ in the EU’s AI Liability Regulation’, initially published in the JUST-AI Jean Monnet Research Papers Series is also published in the spring issue of the International Journal of Law Ethics and Technology (pp. 155-262).
Conceptually speaking, this article canvasses the features and functions of the generic concept of explanation, as well as the features and functions of two of its variants: explainable AI (XAI) and causal explanations i.e. narratives on the causal links between AI systems/outputs and harms suffered. An overview of the available – mostly North-American – caselaw in the field of AI liability (namely Loomis and Pickett) revealed emerging trends on the content of explanations and evidence that litigants and courts flag as necessary for the purpose of providing plausible causal explanations.
Three lessons should be highlighted from this caselaw:
- Explaining how a system made a specific decision (post hoc explainability) matters for the general understanding of the disputed facts;
- Evidence on the accuracy of a given AI output is - when available - required for the purpose of informed judicial decisions;
- If the accuracy of an AI output cannot be proven, providers and deployers should state the reasons why they considered that output to be accurate.
Against this backdrop, the article goes on to analyze the systems of evidence in the upcoming AI Liability Directive (AILD) and Revised Product Liability Directive (R-PLD). The paper is critical of the assumption shared by these instruments that, respectively, fault and defectiveness are consequences of non-compliance with the safety standards listed in the AI Act and applied to high-risk systems. The procedural implication is that, under both instruments, the evidentiary debates (and causal explanations) will focus on showing that providers or deployers had (or had not) complied with applicable safety requirements. Those debates will not prima facie include explanations on how a specific system caused harm in a specific case… But isn’t that the point of XAI when causality is debated before a court?
Driven by a needs-based approach to procedural justice, the article is critical of the ‘evidentiary hermetism’ of the AILD and R-PLD. If litigants are called to causally explain a harm suffered, they ought to first receive explanations on the AI system’s functionalities (ad hoc) and the decisional process having yielded a harmful outcome (post hoc). It remains to be seen if, in future caselaw, the equality of arms principle (under Article 47 Charter) and the right to explanation (under Article 86 AI Act) will support this understanding of explainability.
