ABSTRACT
This article addresses two related questions arising from the commercial deployment of so-called ‘AI agents’. First, it examines whether artificial intelligence systems can properly be regarded as agents, fiduciaries, or trustees in the legal sense recognised by the common law. Drawing on agency law in the United Kingdom, Australia, and the United States – including Bowstead and Reynolds on Agency and the Restatement (Third) of Agency – it argues that ‘AI agent’ is a functional and technological label rather than a legally accurate characterisation. Because contemporary AI systems lack legal personality, they cannot give legal assent, assume fiduciary obligations, hold property, act as trustees, or bear responsibility in their own right. Secondly, the article considers liability when an AI system is hacked, hijacked, jailbroken, or otherwise manipulated by an unauthorised third party and subsequently causes loss. Analysing doctrines concerning intervening criminal conduct, novus actus interveniens, cybersecurity negligence, and product liability, it shows that developers or deployers are ordinarily liable only where their own conduct contributed to the relevant risk, particularly through inadequate security or failures of control. Yet successful hacking events may disrupt causal analysis and undermine reliance on doctrines such as res ipsa loquitur, especially where exclusive control cannot be established. Emerging strict-liability regimes, including the EU’s revised Product Liability Directive, further complicate the landscape by treating software and AI systems as products and cybersecurity vulnerabilities as potential defects. The article concludes that existing doctrines leave a remedial gap for victims of hacked AI systems, supporting consideration of statutory strict liability, mandatory insurance, or comparable reforms.
McCullagh, Adrian, Silicon Agents: Machines, Artificial Intelligence, and the Boundaries of Agency, Fiduciary Positions and Tortious Liability (September 18, 2026).
Leave a Reply