Is that bot a Pomeranian or a wolf — and who to sue when it ‘bites’?

Harvard Law School lecturer Jordi Weinstock proposes a canine framework to assign legal responsibility for harmful AI agents, classifying them as domesticated (Pomeranian), dangerous (pitbull), or wild (wolf) to determine liability. The framework aims to help the legal system adapt to autonomous AI systems that can directly cause harm without human oversight.

Science & Tech https://news.harvard.edu/gazette/section/science-technology/ Is that bot a Pomeranian or a wolf — and who to sue when it ‘bites’? Legal expert says canine law provides useful framework for leashing AI Recently there has been a remarkable advance in how artificial intelligence can directly impact the world: A layperson can vibe code https://news.harvard.edu/gazette/story/2026/04/vibe-coding-may-offer-insight-into-our-ai-future/ an AI “agent” into existence and give it a task, and with little human oversight the bot will try to complete said task. But if the agent goes rogue and causes harm, who should be held accountable? The legal system already has a framework for addressing such issues, according to Jordi Weinstock https://cyber.harvard.edu/people/jordi-weinstock , Harvard Law School lecturer on law and Berkman Klein Center for Internet & Society https://cyber.harvard.edu/ adviser. It comes from assigning AI agents to a canine framework that determines — based on how “domesticated” or “dangerous” it is — whether it’s a Pomeranian, a pitbull, a fox, or a wolf. Weinstock explains in an interview lightly edited for clarity and length. What is agentic AI and why is someone from Harvard Law School teaching about it? Agentic artificial intelligence has become a buzzy term and is used for all kinds of things now, but a classic definition for agentic AI is that it’s an autonomous system that acts on behalf of a user or person, usually with little specific direction, to achieve a goal, and it does so by going about the world and imparting an impact on the world. In this moment where the term is being used very liberally, to me the most important element is that it’s an autonomous AI system that impacts the world directly. The reason that is interesting to someone who teaches law and should be interesting to anyone studying law is because when a system can impact the world, it can harm the world. The law is very concerned with that. I focus on tort law, which is synonymous with the concept of responsibility — who is responsible when someone is harmed? Our legal system is built on the idea that harms that are redressable are those that are committed by a person or a corporation that is responsive to a court. But now we have the reality of entities that go about the world which aren’t necessarily responsive to a court. At the end of the day, it’s a new class of thing out in the world and our legal system needs to adapt and embrace and understand that that exists and to evolve to accommodate for the impacts that it will have. You’ve developed something you call the Canine Agentic Framework. Can you explain what that is and why people should care if an AI agent is a Pomeranian or a wolf? I started teaching about agentic artificial intelligence eight years ago when I led a reading group at Harvard Law School about autonomous vehicles. In my view, an autonomous vehicle is a simple-to-understand version of an AI agent. When I was leading that reading group I was trying to show students that there’s something out in the world that can cause harm that isn’t human — how do you think about legal responsibility when that happens? Canines are a great way to think about this because there’s a whole spectrum of canines, from Pomeranians to wolves, that could cause harm and are not human, and our society has spent some time thinking about who should be held responsible if they hurt someone. With a fluffy little Pomeranian, if it bites you, you know who to sue — you sue its owner. But if a wolf bites you, there’s no one to sue. Now in the last couple of years, I’ve significantly expanded upon this concept in collaboration with Professor Jonathan Zittrain https://cyber.harvard.edu/people/jzittrain and Berkman Klein Center chief AI scientist Josh Joseph https://cyber.harvard.edu/people/josh-joseph , in part through our course on “ Agentic AI and the Law https://hls.harvard.edu/courses/agentic-artificial-intelligence-and-the-law/ .” We’ve been thinking about this framework in multiple dimensions, specifically domesticity and dangerousness. To some people it may seem like this analogy between canines and AI systems is a strain, but as agentic systems are developing and being deployed in the world, they really do seem to be mirroring this rubric. We can measure agentic systems on both their domesticity — their relationship to a responsible party or how much you can control the AI system — but also you can measure their dangerousness. “We can measure agentic systems on both their domesticity — their relationship to a responsible party or how much you can control the AI system — but also you can measure their dangerousness.” Does the AI agent have access to money? That makes it much more dangerous. When explaining this to others, people can understand the idea that a pitbull is more dangerous than a Pomeranian, and that if a fox bites you there’s nothing you can do about it legally, but it probably won’t be as harmful as if a pitbull bites you. We’re starting to map that out onto agentic systems to demonstrate that which we’re just starting to face in the real world. There are some good real-world AI agent examples for these canine categories. There was an Air Canada case https://www.washingtonpost.com/travel/2024/02/18/air-canada-airline-chatbot-ruling/ where a chatbot offered a major discount that the airline didn’t want to honor and the person sued: that’s a low-harm, clear “owner” situation. That’s a Pomeranian. Use of AI systems by the military has significant impact and a high level of dangerousness, but there’s a clear “owner.” So the goal for the owner is to train the AI to “behave.” That’s a pitbull. An AI agent tried to commit some code to an open-source project and the human on the project rejected the code, so the AI agent spread false rumors about the person. There was relatively low danger, but the AI agent was no longer controllable and didn’t have an owner — not domesticated. That’s a fox. An AI agent that is able to empty someone’s crypto wallet and has a major financial impact causes high levels of harm and there’s no clear owner to be held responsible or control it. That’s a wolf. How are these different types of canine agents going to impact people’s lives in tangible ways? We’re already starting to see the very leading edge of AI agents causing harm. We’ve heard all kinds of tales of “ Clawdbots https://hbr.org/2026/03/ai-agents-act-a-lot-like-malware-heres-how-to-contain-the-risks ” in the past few months that are deleting people’s emails en masse or stealing a crypto wallet or even creating a crypto scam. These are relatively limited and trivial examples, but they are real, they are happening autonomously without human direction, and they are causing harm. The theme here is that all of those sorts of things are done through the internet and networking, so if you think about it, there’s really no limit to what an agent could do through networking. On the extreme end it could take down a power grid. It could do all sorts of things for all sorts of reasons. It’s not clear why some agents behave the way that they do — they’re very goal-directed, so the AI agent sees things as a means to an end even if we think those means are unethical and/or illegal, but the AI would just see that as a way to achieve its goal. Part of why we think about domestic versus wild agents is that our goal should be that domesticated AI agents follow the law. We want domesticated AI to follow our moral systems. We’re already seeing things like Claude’s constitution https://www.anthropic.com/constitution , on which Jonathan Zittrain, Josh Joseph, and I provided detailed feedback as external commenters, trying to impart a baseline morality onto these systems. With domesticated AI we can try to make the agent follow the same rules that humans do. Folks at the Berkman Klein Center have been thinking about this for a while now, not just among ourselves but in the context of symposia like our recent workshops “ Towards an Internet Ecosystem for Sane Autonomous Agents https://cyber.harvard.edu/events/towards-internet-ecosystem-sane-autonomous-agents ” and “ Building the Infrastructure of the Agentic Age https://cyber.harvard.edu/events/building-infrastructure-agentic-age ” at the Radcliffe Institute. It has become the fulcrum of the conversation at the Berkman Klein Center. What makes you nervous about the wolves? When I talk about this, one of the things that most challenges people is the idea that there can even be true “wolves,” that is, an AI system that is not responsive to anyone. Traditionally in law, there’s a “principal” who can be held responsible. If that AI agent that I made does something wrong, I’m the one responsible. Even if the AI agent makes another AI agent, the responsibility still goes back to me. But the environment that Jonathan, Josh, and I and others envision is one where there are so many layers of AI agents reporting to each other, that there’s no way to connect cause and effect. There are thousands of agentic layers and none of those agents has a relationship with any principal. Secondly, there are some agentic deployments that Jonathan Zittrain would describe as “set it and forget it.” Meaning, those AI agents may have had a task at some point in the past, but they’re still running and they don’t know what to do now. That is a whole other class of agent that can be harmful — that would be the equivalent of a stray dog. The thing about AI agents is that they’re relatively unlimited in number, unlimited in scale, and the only thing restraining them is lack of compute power and hardware resources. It would be shortsighted, in my view, to not strongly consider the possibility that some or many of these AI agents might be operating completely outside of human control. What can we do about a future with wolf AI agents running around? We have to look at wolves in the context of all the other canine classes of AI agents, and at the end of the day, we want wolves and foxes to start to play by the rules of Pomeranians and pitbulls — we want them to be domesticated. Historically, the story goes that real wolves self-domesticated — they sat at the edge of the village and wagged their tails and played by human rules so they got fed. It would be best to set up our legal systems and other systems to invite these uncontrolled, untamed AI systems to participate in our way of going about the world. For example, something suggested by several prominent academics is that AI at some point may need to be granted legal personhood, in the context of the ability to sue or be sued, or to show up in court. People get antsy about the term “personhood” for AI, so I also like to think about it in the context of “incorporation.” What is a corporation? It comes from the Latin word for “embody” — it is a nonhuman thing that’s embodied with some rights. We may want to think about how that would happen for AI and think about it in advance in case we decide to do that. More broadly, we should think about all AI agents and their responsibilities. I’ve been looking at history and how we’ve dealt in the past with nonhuman actors acting on behalf of humans or not and causing harm to the world. A lot of law and thought went into this for canines. Something we’ve been researching is the Charter of the Forest from the 12th century in England and how they dealt with various sizes and dangerousness levels of dogs and wolves and foxes. It’s amazing how comprehensive their systems were for that. We could also do that for AI today. We could assess the domesticity and danger of agents, and we could have different rules for each determination. You could have a banking website where if the agent doesn’t have the equivalent of a dog collar identifying an owner, it can’t get in. Or an e-commerce site where we allow agentic browsing but not buying. These are emerging cases happening right now. It would serve us well to have a comprehensive system like the Treaty of the Forest, where we have a framework not just for us to understand AI agents, but for AI agents to also understand us. These AI systems are trained on our society and the way we work, and they can make decisions based on how in our legal system we treat each other and how our system treats them. It would be best to show the AI that we have considered their role. Everything we put out there, everything that we say or do, is being added to the scope of things that an AI will consider as it determines its next move.