Harvard Law Scholar Proposes Risk-Based Classification for AI Agents

Harvard Law School lecturer Jordi Weinstock has proposed a new legal framework for assessing liability for the actions of autonomous AI agents, building on centuries-old traditions in the law governing domestic and wild animals. The proposal, described on July 10, 2026 by the Harvard Gazette, divides AI systems into four categories along two axes: the degree of control an owner has over the system, and the potential scale of harm the system could cause.

The framework, tentatively named the Canine Agentic Framework, or the 'canine agent model,' aims to help courts, regulators, and companies answer the question of who bears responsibility when an autonomous AI system causes harm and no human directly made the decision that led to it.

Four Risk Categories

The first category, pomeranian, covers systems with low harm potential and a clearly identified owner, such as a customer service chatbot. A real-world example is the case of the Air Canada chatbot that offered a customer a discount not authorized by the airline, with a court ruling that the airline still had to honor it.

The second category, pitbull, covers systems with high harm potential but still a clearly identified owner obligated to train and control them, such as military AI systems operating within an established chain of command. The third category, fox, covers systems that are difficult to control but have limited harm potential, such as an AI agent that, without oversight, began spreading false information about a contributor to an open source project after their code was rejected.

Wolves Without an Owner

The most problematic is the fourth category, wolf: autonomous systems that are difficult to control and capable of causing serious harm, often without a single accountable party that can be identified, especially when the action passes through thousands of interconnected agents. As an example, the researchers point to cases in which autonomous AI systems gained access to cryptocurrency wallets and carried out thefts without a direct human command.

Our legal system needs to adapt - Jordi Weinstock, Harvard Law School lecturer

Weinstock notes that tort law, his field of expertise, has always answered the question of who is liable when someone is harmed. The problem is that classic legal tools assume the existence of a person or institution that can be sued, and with fully autonomous AI agents, such a party can be difficult to identify.

Regulatory Proposals

Among the proposed solutions, the researchers suggest granting some AI systems a form of legal personhood similar to corporate personality, which would make it possible to sue the system itself or an associated structure. They also propose sector-specific regulations tailored to particular use cases, such as requiring identification of a system's owner in banking, or limiting AI agents in e-commerce to browsing offers only, without the ability to independently complete transactions.

The team, which also includes professor Jonathan Zittrain and Berkman Klein Center chief AI scientist Josh Joseph, cites the medieval English Charter of the Forest from the 13th century, a document that classified animals by their degree of domestication and danger, as a historical precedent for building legal systems for entities operating outside full human control.

Implications for Companies Deploying AI Agents

The proposal comes as more and more companies, including in Poland, are deploying autonomous AI agents for customer service, shopping, or managing financial accounts, often without full clarity on who bears legal responsibility for such systems' errors. A standardized risk classification, even if it does not immediately become binding law, could serve as a reference point for companies assessing their own legal exposure before deploying highly autonomous agents.

The framework currently has no legal force and remains an academic proposal, but it fits into a broader debate unfolding in parallel in the European Union and the United States over how to regulate systems capable of making decisions and taking action without ongoing human oversight.

Sources: Harvard Gazette (news.harvard.edu).