Responsibility Attribution for AI-Mediated Damages with Mechanistic Interpretability
Lena Kästner; Johann Cordes; Herbert Zech · 2026 · AISoLA 2024 (LNCS 16032, Springer), pp. 187-202 interlocutor high priority coded
Main argument
Thesis: legal responsibility for AI-mediated damages should be attributed by conceptualizing causation as difference-making (Woodward-style interventionism, matching the legal conditio sine qua non), distinguishing three types of difference-makers - (i) inputs, (ii) post-training functional organization, (iii) system history (training data, design decisions) - and matching each to an identification strategy: classical XAI for (i), mechanistic interpretability (MI) for (ii), lifecycle analysis for (iii). Argument type: conceptual analysis applied to legal doctrine. Practical upshots: liability for input-attributable damages splits deployer (misuse) vs provider (intended use); MI, where it gives deployers genuine understanding of a system's functional organization, SHIFTS liability from providers to deployers - responsibility tracks epistemic access; history-attributable damages sit with providers; distant actors (data subjects) should be legally indemnified. Recommends MI as gold standard where high risk meets high stakes. Acknowledges causation is necessary but not sufficient for responsibility - sometimes 'no (human) actor is actually in command of the difference-maker in question' - and points to game-theoretic/multi-agent methods without resolving that case.
Why it matters here
Philosopher (Kästner, philosophy of science/causation) + legal scholar (Zech) jointly working the responsibility-attribution problem with real legal machinery (EU AI Act, AILD, PLD). Shows what the LEGAL half of the responsibility question looks like when done rigorously - the institutional-liability counterpart to the dissertation's moral-responsibility analysis, and a live demonstration that interpretability tools redistribute responsibility.
Reading notes
Full close read completed. 16pp chapter, extracted from AISoLA 2024 proceedings. Note: explicitly brackets moral responsibility (fn2: liability analysis 'explicitly brackets considerations of moral responsibility') and explicitly defers generative AI to future work - two gaps the dissertation occupies. Funded by Volkswagen Foundation 'Explainable Intelligent Systems' project.
Kästner, L., Cordes, J., & Zech, H. (2026). Responsibility Attribution for AI-Mediated Damages with Mechanistic Interpretability. In B. Steffen (Ed.), AISoLA 2024 (LNCS 16032, pp. 187-202). Springer. https://doi.org/10.1007/978-3-032-01377-4_10
Close reading — 14 coded units
“By AI systems we mean, following the current version of the AI Act, any 'machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments' (art. 3(1) EU AI Act).”
“an unambiguous attribution of responsibility is crucial to ensure legal certainty: for one thing, those affected should be able to assert their rights effectively; for another, various stakeholders should be provided with a clear legal framework to guide their activities. Yet, it remains unclear (a) what conception of causation liability law relies on, (b) how this conception can be utilized to attribute responsibility when human actions rely on the use of opaque AI systems, and (c) how liability for AI-mediated damages should be handled in practice.”
“The core idea of interventionism is intuitive: causal explanations embody a what-if-things-had-been-different conception of explanation. [...] the idea is that if C causes E, then C will (provided certain conditions are met) be a difference-maker for E. [...] the intuition it embodies [...] matches the intuition behind the legal scholars' conditio sine qua non.”
“modern AI systems are becoming—and will remain—increasingly opaque not only to their users but also to deployers and providers. Besides, there are many actors potentially involved in building and using AI systems [15] and different stakeholders prioritizing different norms in different contexts [27]. Thus, identifying relevant difference-makers, and the humans in command of them, presents a serious challenge.”
“(1) If we are trying to determine who is liable for damages attributable to inputs to an AI system, we are effectively asking about the relevant type (i) difference-makers [...] (2) If, by contrast, we seek to find out who is liable for damages attributable to a system's overall functional organization, we are interested in type (ii) difference-makers (e.g., certain components, units, or circuits within the system's functional architecture). (3) Finally, if we wish to identify who is liable for damages attributable to a system's history, we are seeking to uncover type (iii) difference-makers (viz. features in a system's history, such as the training data or design decisions [...]).”
“First, suppose the system works properly but is used incorrectly [...] In this case, we submit, the deployer is liable for the damages that potential misclassifications incur. If, by contrast, the system is used as intended, it is up to the providers to ensure the proper functioning of AI systems, and thereby prevent inputs (type (i) difference-makers) incurring damages.”
“provided that the system's components work as anticipated, liability for damages attributable to type (ii) difference-makers may actually shift from system providers to deployers; at least insofar as MI gives deployers a certain grasp of how a given AI system works and role specific components play in its overall functioning. [...] if deployers understand the functional organization of a system, they can reasonably infer where and how to use it safely; if they consciously deviate from that use and this incurs damages, they should be held liable even if providers might have specified a different scope of safe use.”
“This suggestion relates to worries about humans in the loop becoming scapegoats [40]: When deployers work with black box AI systems they do not understand, and just (blindly) confirm or follow AI recommendations, they run the risk of becoming mere scapegoats for AI-mediated damages; they are held responsible although they do not have any understanding or control over the damages incurred. With our analysis in mind, scapegoating can be avoided: for if the deployer utilizes an opaque AI system in the ways it is intended to be used, they are not responsible for potential damages [...]; if, by contrast, the system is rendered interpretable, they might be responsible but no longer unknowingly so.”
“Our suggestion is in line with Article 14 of the EU AI Act which states that adequate human oversight in high-risk scenarios demands an overseer to 'properly understand the relevant capacities and limitations of the high-risk AI system and be able to duly monitor its operation, also in view of detecting and addressing anomalies, dysfunctions and unexpected performance' and be able 'to intervene on the operation of the high-risk AI system'.”
“it seems straightforward that difference-makers in an AI system's history will usually not be influenced by deployers but providers. Thus, liability for damages attributable to an AI system's history will generally lie with providers. [...] AILD narrowly focuses on linking an actor's fault directly to the AI system's output, neglecting earlier causal factors such as flawed training data or design choices, which we seek to integrate. PLD, by contrast, stipulates that the liability encompasses the entire causality chain.”
“Though causation is considered a necessary condition of responsibility and liability (Sect. 2), it may not be sufficient. For instance, identifying relevant difference-makers may not automatically reveal the actor in command of them (see also Sect. 4.3); or we might find that no (human) actor is actually in command of the difference-maker in question. To this end, multi-agent models, Markov-models and game-theoretical methods might be extremely useful [2,52].”
“Second, we suggest requiring MI as a gold-standard for complex AI systems, at least under certain conditions and where systems are not already inherently interpretable. [...] We propose to require MI specifically where high risks and high stakes come together.”
“Third, and finally, we suggest to create legal rules indemnifying (excusing) distant actors. Paradigmatically, data subjects are only remotely involved in model building by providing training data. [...] the individual influence of any data subject on any given AI-mediated damage is negligible compared to that of system designers.”
“While our discussion has not explicitly taken generative AI (genAI) into consideration, we believe it is only natural to extend our views to this technology. To be sure, MI is already hard and costly to achieve for non-generative AI [...] But that remains the project of another paper.”
Synthesis-matrix row
complicates
T6-RESPONSIBILITY-UNALLOCATED
supplies allocation tools (difference-making, epistemic condition) but brackets moral responsibility and defers genAI
supports
T7-AGENTIC-BREAKS-FRAMES
genAI explicitly deferred by the responsibility state-of-the-art
Memos (5)
theoretical · unit #11
Unit 11 is the responsibility gap (Matthias 2004) surfacing inside a rigorous legal-doctrinal analysis: after all the difference-making machinery is deployed, the authors concede there may be 'no (human) actor actually in command of the difference-maker' - and their answer is a pointer to game-theoretic methods, not a normative account. Contrast with Luke's move (assign responsibility to the rational AI): Kästner et al. never even consider the AI as a responsibility-bearer; the option space is humans, institutions, or formal reallocation. This is strong evidence for the dissertation's framing: the legal literature has sophisticated tools for DISTRIBUTING responsibility among humans but no answer for the residue case - which is exactly where agentic systems make the residue larger.
theoretical · unit #7
RL-EPISTEMIC (units 7-8) is the legal analogue of the knowledge/epistemic condition on moral responsibility (Fischer & Ravizza reasons-responsiveness; Aristotelian ignorance excuses). The philosophically significant consequence: interpretability TOOLS redistribute responsibility - a technical artifact (MI) changes who is morally/legally on the hook. This mirrors and extends Gabriel 2020's tech/normative interdependence thesis (GABRIEL_2020 unit 3): there, architecture constrains which values can be encoded; here, explanation infrastructure determines who bears responsibility. The dissertation can unify both under one claim: technical choices are covert normative allocations.
thesis-link · unit #8
The scapegoat-in-the-loop passage (unit 8, citing Ranisch 2024, American Journal of Bioethics - medical AI) is direct Health-chapter material for the AI Scribe case: a clinician blindly confirming AI-generated clinical notes is the paradigm scapegoat. And the structure of the worry - formal human sign-off without genuine understanding or control functioning as a responsibility-transfer ritual - is Augustine's comps 'consent laundering' concept operating on the deployer side rather than the data-subject side. Worth developing as a named parallel: consent laundering (data subjects) and scapegoating (human overseers) are two faces of the same illegitimate-responsibility-transfer pattern.
comparison · unit #14
Cross-source pattern now visible across all three coded sources: GABRIEL_2020 anticipates autonomous default decision-making but builds for the dyadic tool case; FISCHLI_2026 handles agentic personal assistants but defers collective/multi-agent effects; KAESTNER_2026 handles multi-actor liability but defers generative AI entirely (unit 14). Each is strong precisely where the others are silent, and the intersection - responsibility allocation for multi-step agentic systems - is occupied by none of them. That intersection is the dissertation's gap, now evidenced from three independent literatures (alignment theory, agent ethics, liability law).
theoretical · unit #1
Note for the regulation strand: the EU AI Act's own definition of AI system (unit 1) already includes 'varying levels of autonomy' and post-deployment adaptiveness - the LAW's definition is agentic-ready while the responsibility doctrines applied to it (AILD's narrow output-fault link, unit 10) are not. The mismatch between definitional scope and doctrinal machinery is itself a citable finding for the governance chapter.