Bart Kubiak (ERA Fellowship) has posted The Verifiable Responsible Agent Framework: Making AI Agents Liable For Their Mistakes on SSRN. Here is the abstract:
Autonomous AI agents now deal in stocks, currencies and commodities. They do it at a speed and volume that defeats human intermediation. Yet, when one of them causes loss, no liable subject stands behind it: a model is not a legal person, and the operator who deployed it is often too far removed to be held liable. This article argues that the missing subject can be supplied without resolving questions of machine consciousness. It proposes granting agents a limited legal capacity made conditional on attested architectural constraints and bonded to insurance at the moment of issuance. The effect is that the legal form and the technical guarantee cannot drift apart — because enforcement requires a subject. It develops this Verifiable Responsible Agent (VRA) framework focusing on the horizontal coordination problems of private commercial transactions. Drawing on functionalist jurisprudence, principal-agent theory, and comparative institutional analysis, the article distinguishes what is verifiable today (operational-safety and commercially-reasonable constraints) from the fiduciary properties that remain a research frontier. Through examination of implementation possibilities in Delaware (US), the United Kingdom, and the European Union, it argues that incremental legal innovation can establish foundations for broader standardisation of AI-agent recognition while maintaining meaningful human accountability. The framework’s technical architecture, cryptographic attestation, formal verification methods, open tool-use protocols (e.g., the Model Context Protocol), and standardised agent-agent communication, provides concrete implementation pathways that balance operational autonomy with verifiable compliance.
Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum