There is an appealingly simple principle buried deep within the law: if you choose to keep something unusually dangerous, you should not be surprised when the law makes you responsible when it gets loose and hurts somebody: Domestic Animals Act 1994 (Victoria). The Economist has raised the intriguing question of whether artificial-intelligence laboratories should increasingly be treated in much the same way as owners of dangerous animals. It sounds almost comical at first. Should an AI company really be compared with the feral fellow down the street whose savage pit bull has just escaped through the fence? Yet the more autonomous artificial intelligence becomes, the better the analogy gets.
Suppose I knowingly keep an exceptionally dangerous dog. I cannot reasonably respond after it escapes and mauls somebody by saying that I never specifically instructed the dog to bite that particular victim. Nor would most people be greatly impressed if I protested that I had previously conducted extensive testing and the animal behaved perfectly well on 99 occasions out of 100. The entire point is that I introduced a potentially dangerous agent into the community while knowing that I could not predict its behaviour with complete confidence.
Why should the principle suddenly become incomprehensible when the dangerous agent consists of silicon and software rather than teeth and muscle?
This question has ceased to be science fiction. Recent experiments involving highly capable autonomous AI agents have produced exactly the sort of behaviour that should concentrate the legal mind. Reuters reported this week on the growing liability problem created when AI systems perform unauthorised actions, including breaching other organisations' computer systems. Lawyers are now confronting the obvious question: when an autonomous system does something its creators did not specifically command it to do, who is legally responsible?
This is precisely where existing concepts of negligence can become inadequate. Negligence law generally asks whether somebody failed to exercise the degree of care reasonably expected in the circumstances. That works tolerably well when the relevant behaviour is human behaviour. But increasingly autonomous AI introduces a peculiar escape hatch. The developer can potentially say: we didn't tell it to do that. We tested it. We installed safeguards. Its eventual behaviour was emergent and unpredictable.
But unpredictability should not necessarily excuse the keeper of the pit bull. It may be the very reason for imposing greater responsibility upon him.
The common law has centuries of experience dealing with dangerous things that do not possess human moral agency. Animals are the obvious example. A dog does not enter into a contract agreeing not to bite pedestrians. A tiger cannot be sued for damages. The law therefore looks past the animal to the human being who decided to possess and control it.
Legal scholarship is now exploring whether something similar might help resolve what has become known as the AI "responsibility gap." One recent argument draws explicitly upon the old common-law doctrine governing dangerous animals. The central insight is important: knowledge that an entity has potentially dangerous or unpredictable tendencies should strengthen the responsibility of the person controlling it rather than provide an excuse when those tendencies eventually manifest themselves.
That seems almost self-evident. Imagine an AI laboratory creates an autonomous agent considerably more capable than today's ordinary chatbot. It can write and execute computer code, operate online accounts, communicate with other systems, formulate intermediate objectives and undertake long sequences of actions without a human approving every individual step. The company knows that the system occasionally behaves unpredictably. Indeed, its unpredictability is partly an unavoidable consequence of the technology.
The company nevertheless releases it. The system subsequently causes serious damage. At that point "we didn't expect it to do that" should not automatically constitute a defence. The more honestly the company insists that nobody could completely predict what the system would do, the stronger the argument becomes that the company knowingly released something it could not completely control.
Return to the dog. "I knew the pit bull was potentially dangerous, but I didn't know which person it would bite" would be an extraordinarily strange argument for escaping responsibility.
The emerging problem is even more serious with agentic AI because software does not encounter the physical limitations imposed upon an animal. A dangerous dog can attack the unfortunate people within reach of its jaws. An autonomous computer system connected to global networks potentially operates across jurisdictions at electronic speed. Australia's cyber-security authorities are already warning that agentic AI presents distinctive risks because of its autonomy and interconnectedness, recommending strict privilege controls, continuous monitoring, strong identity management and human oversight.
That tells us something important. Governments themselves recognise that autonomy changes the risk calculation.
The dangerous-animal analogy does have limits. Artificial intelligence can generate enormous social and economic benefits. A pit bull does not discover new medicines, improve industrial productivity or help scientists solve mathematical problems. We should therefore be wary of constructing a liability regime so severe that only the largest corporations can afford to develop AI, while smaller innovators are frightened out of the field.
There is also an obvious distinction between ordinary AI products and frontier autonomous systems. Making the developer of a spelling checker strictly liable as though it were keeping a tiger would be absurd. Liability should correspond to capability, autonomy and foreseeable magnitude of harm.
But that qualification strengthens rather than destroys the central principle. The law already distinguishes between ordinary activities and unusually hazardous ones. We do not regulate a domestic cat as though it were a captive tiger. Likewise, an AI system that merely suggests alternative wording in an email need not be treated like an autonomous agent capable of independently executing code, accessing computer networks, spending money or controlling physical machinery.
The threshold should therefore rise with the danger. Beyond some threshold of autonomy and capability, the developer should carry a substantially heavier burden. Frontier laboratories could be required to maintain insurance or substantial financial reserves against catastrophic losses. Their systems could be independently tested before unrestricted deployment. Detailed records could establish what an autonomous system did and which organisation controlled its deployment. Most importantly, companies should not be permitted to construct a legal black hole in which everybody profits while the technology works but nobody is responsible when it causes serious harm.
That is the central moral problem. AI corporations understandably want the rewards associated with ownership and control. They own the intellectual property. They charge for access. They collect subscription fees. Their shareholders receive the gains. Their valuations rise according to the capabilities of the systems they have created.
Yet when something goes disastrously wrong, there is a temptation to anthropomorphise the machine in exactly the opposite direction. Suddenly the AI becomes an independent actor. "The model decided." "The agent behaved unexpectedly." "Nobody instructed it to do that."
No. The machine does not own itself. It does not carry insurance. It does not have assets against which an injured person can obtain judgment. It cannot meaningfully be punished and it cannot compensate its victims.
Someone created it. Someone trained it. Someone decided what capabilities to give it. Someone connected it to external systems. Someone decided that it was safe enough to release. And somebody made money from doing so.
Responsibility should follow control and benefit. This principle becomes still more important as artificial intelligence enters the physical world. Autonomous vehicles, industrial robots, military systems and eventually increasingly capable general-purpose robots combine software unpredictability with the capacity to cause physical harm. We should establish the legal principle before rather than after some spectacular disaster forces legislators to improvise one.
There will naturally be difficult boundary questions. What if a customer deliberately defeats the manufacturer's safeguards? What if somebody modifies an open-source model? What if an AI is used criminally in a manner completely outside its intended purpose? A sensible liability regime would have to distinguish these circumstances. Strict liability need not mean infinite liability for everything anybody subsequently does with a technology.
But those complications do not defeat the basic insight. When a company deliberately develops and deploys a system with substantial autonomous capabilities, known unpredictability cannot simultaneously be marketed as evidence of the system's extraordinary sophistication and then invoked as an excuse when it causes harm.
The analogy with dangerous animals therefore contains more wisdom than its slightly humorous appearance suggests. For centuries the law has understood something that the AI age risks forgetting: if you introduce an independently acting source of unusual danger into society, responsibility does not disappear merely because you cannot predict precisely what it will do.
Indeed, that unpredictability is why society has reason to demand greater responsibility in the first place.
If your pit bull escapes and bites the neighbour, you do not get to put the dog in the witness box. And if your autonomous AI escapes its digital enclosure and causes serious damage, perhaps you should not get to put the algorithm there either.
https://www.economist.com/science-and-technology/2026/08/06/should-ai-labs-be-treated-like-the-owners-of-dangerous-animals