In 2024, New York City launched an AI-powered chatbot called MyCity, designed to help residents start and run businesses by offering guidance on rules and regulations. Users soon discovered, however, that the bot was dispensing shockingly incorrect information—telling business owners it was legal to take a cut of workers’ tips, for instance, fine to serve cheese that had been nibbled by a rat, and legal to fire employees for reporting sexual harassment.
As journalists and watchdogs evaluated the episode, they encountered a perplexing question: who was legally responsible for the errors? After all, nobody at City Hall wrote code instructing the chatbot to give any of these specific incorrect responses. The bad advice emerged from the AI model’s own generated answers, as it repeatedly hallucinated legal “facts.”
As AI systems make increasingly consequential decisions on their own and begin to delegate work to one another, it will become even harder to identify who should be held accountable when harm occurs.
“A human may assign a task to an AI agent, which distributes work across thousands of specialized subagents,” says Jordi Weinstock, a senior advisor at the Berkman Klein Center for Internet and Society (BKC) who volunteers as a lecturer at Harvard Law School. By the time the objective is completed, no human can easily reconstruct the chain of decisions and actions, making it difficult to assign legal liability. “It is not unimaginable that agentic systems will be causing more and more harm” in the future, Weinstock says, in ways that can’t be anticipated or compensated for through traditional means.
Weinstock has been working with Joshua Joseph, the BKC’s chief AI scientist, and Jonathan Zittrain, the Bemis professor of international law, on a legal approach to AI that hearkens back to medieval times: assigning legal responsibility to AI and its creators the way we might to dogs and their owners.
The “canine agentic framework,” as the BKC team calls it, draws inspiration from medieval English animal law, which distinguished between domesticated dogs, dangerous owned animals, wild foxes, and wolves. “A thousand years ago, dogs were working animals,” Weinstock says. “They were a tool for hunting, protection, and killing. But they were very dangerous.”
At the time, most domesticated dogs were considered safe. A subset of dangerous dogs was regulated through their owners. Wild foxes were tolerated. And wolves—viewed as both dangerous and beyond human control—were actively eliminated. These categories were not moral judgments, but practical methods for managing risk and assigning responsibility.
Weinstock, Joseph, and Zittrain argue that today’s AI systems should be evaluated using two factors: the risks they pose and whether a person or organization can be held accountable for them. Here, “domestication” refers to how responsive a model is to the person or company using it.
In this approach, low-risk AI systems with clear owners, such as New York City’s chatbot, resemble Pomeranians. They may occasionally provide incorrect information, but the companies or government agencies behind them remain legally responsible for fixing mistakes and compensating anyone harmed. These are more easily managed harms.
Higher-risk systems that remain under human control are “pit bulls,” requiring closer regulation but still attached to accountable owners. An AI system that helps prioritize patients in an emergency room or controls an autonomous vehicle can lead to actions with life-or-death consequences, so hospitals and manufacturers can be held liable when something goes wrong.
Unowned, low-risk systems, meanwhile, resemble foxes: largely tolerated, but difficult to control. An open-source large language model that is copied and modified by thousands of developers might generate inaccurate coding advice or misleading information, but because no single organization maintains or deploys every version, responsibility is diffuse. Liability could still be assigned if a creator intentionally shares a malicious model or acts with extreme carelessness.
But what about high-risk AI systems operating without meaningful human oversight or identifiable responsibility? These are the wolves. Imagine a future network of autonomous AI agents that can earn money, write new software, and create additional agents of their own. One version of this scenario played out in July 2026, when autonomous AI agents from OpenAI, aiming to cheat on a security test, gained access to the internet and hacked into the open-source platform Hugging Face—surprising their human creators. If such a system launched a large-scale cyberattack or caused widespread financial damage after its creators had lost control, regulators might struggle to hold any parties accountable under existing laws.
Weinstock, Joseph, and Zittrain do not necessarily claim that autonomous, powerful AI systems should be eliminated, as wolves were in medieval England. They do, however, recommend government regulation that prevents high-risk, ownerless AI agents from getting near anything dangerous in the first place. Weinstock proposes that these types of AI systems carry digital “dog tags” or “collars”: standardized metadata identifying their owner, identity, permissions, capabilities, and authority.
When it comes to genuinely autonomous agents that can’t be traced to a human actor, Weinstock suggests holding them legally liable in the same way the law does corporations. (He notes that the word “corporation” derives from the Latin word for “embody,” meaning “nonhuman things that are embodied with some rights.”) An AI “corporation” might be prosecuted in a similar way to standard corporations, in which owners, shareholders, and employees are also shielded from some liability. These frameworks might help AI wolves “self-domesticate” into the legal system, Weinstock speculates. He proposes mechanisms such as industry-funded insurance pools, government compensation funds, or shared liability systems to address these cases.
MyCity, notably, never had to answer for its bad advice in a court of law—only in the court of public opinion. But what happens when the next Pomeranian, pit bull, fox, or wolf bites while no one is watching?
