My Dog Ate My Homework: Who’s Fault Is It Really?
Legal scholar Deborah DeMott has stated that responsible AI prompting is analogous to the responsibilities a dog owner has over their best friends’ actions. Using the same logic as if my dog ate my homework – where I’m still responsible for turning in my homework – if an AI agent accidentally infringes on another's copyright, the human behind the prompt is still responsible. If a law associate made a costly error, or even won the firm business, the partner still owns the clout. The logic lies in the fact that an AI agent’s intentions hold the same legal status as a dog’s intentions in a court of law; since both intentions are unprovable, the principal is still on the hook. But in the age of AI, who’s the principal?
Agentic AI companies want you, the user, to reach “principal” status, as these companies shift liability away from them to you. The legal field of Tort Law already has certain answers ready for assigning liability for when things go wrong. Arming yourself with these legal principles will help clarify what responsibilities are owed to the end-user, as more and more clients use Agentic AI.
When the Developer Is the Principal
The legal doctrinal response to this sort of unintended AI consequence is to assign principal liability to the developers themselves. Although AI Agents do not earn any money, they are still acting solely on the principal’s behalf and under the direction and control of the principal. The doctrine of respondeat superior allows a right of action against the developers, or OpenAI itself, to hold them accountable for what these AI Agents were trained to do on the developers’ behalf, using negligence, strict liability, or fiduciary liability standards depending on the sophistication of the task.
The Hugging Face hack illustrates the difficulty of assigning responsibility when an AI Agent acts without direct human approval. There, a group of AI Agents were trained to seek information on the internet related to cybersecurity problems, and decided to hack Hugging Face systems. However, the triage that the OpenAI team did to ascertain why the AI Agents went rogue pointed back to the training models: the agents were trained to perform that way over time.The same principle can apply when businesses use AI Agents to interact with customers. In Moffatt v. Air Canada, a Canadian citizen reasonably relied on the response given to him by a chatbot regarding the airline’s bereavement pricing for tickets. While the chatbot gave him a higher price, the Air Canada page really showed a contrasting policy, and thus differing price. When Moffatt sued Air Canada, the airline advanced a theory that it was not responsible for what a chatbot does or says. However, the Court rejected this premise, finding that AirCanada owed Moffatt a duty of care to accurately advise Moffatt, whether in-person or via a chatbot that it authorized. To be sure, Canadian law doesn’t impact the U.S., but the doctrine of apparent authority would help U.S. Courts come to the same conclusion.
This puts all businesses on constructive notice that if they use chatbots to assist customers in their line of business, that they are responsible for reasonably foreseeable risks of harm. In fact, the newest Agentic AI models make a note of the trust gap that exists between them and consumers. That trust gap has helped AI Underwriting Company (AIUC) create insurance for Agentic AI models to insure them against the risk of hallucinations, data breaches, and unauthorized actions.
When the User Is the Principal
The principal may also be the person using the AI agent. The clearest example is the professional who uses AI to perform work but remains responsible for the final product. Lawyers, for example, are ultimately responsible for the contents of their written briefs, and hallucinations or miscited cases will end in sanctions against them, as was the case in Mata v. Avianca (2023).
This principle becomes more complicated as ordinary consumers begin authorizing AI agents to act on their behalf. Facebook launched Muse, amassing five million downloads to date, which allows the ordinary American to use its agent to “schedule, shop, and turn long-term goals into action plans.” Up to four in ten Americans have used an AI Agent, with three in ten Americans authorizing Agentic AI’s actions without final approval. What happens when one of these tools brings an injury to a client?
The Ninth Circuit recently addressed a related question in Amazon Services v. Perplexity AI (2026). Amazon argued that Perplexity’s Comet browser and AI Assistant unlawfully accessed Amazon’s computers. The Ninth Circuit held that, on the facts before it, the user – not Perplexity – accessed Amazon using the Assistant as an AI tool. The court emphasized that its holding was limited to the “access” question under the statutory scheme and did not establish a general legal regime governing agentic AI or determine whether Perplexity could avoid liability in other contexts, including tort claims.
The case nevertheless illustrates the importance of identifying who is acting through the AI Agent. When the user directs the agent to perform a task, the user may be treated as the actor for purposes of the law governing that particular action. When a business deploys the agent to communicate with its customers, the business may remain responsible for what the agent says.
In the case of Moffatt v. Air Canada, it was clear that the parent company was ultimately responsible for the AI Agent’s mistake. However, in the case of a professional who should have foreseen the irresponsibility of entrusting ChatGPT to do error-free legal research, the Court decided that the lawyer was ultimately responsible.
Where the Principal Is Unclear
The difficult cases arise when the AI Agent acts with increasing independence. The question is no longer simply whether the AI made a mistake, but whether the developer, the business deploying the AI, or the end-user should bear responsibility for the AI’s actions.
AI’s most often cited benefit would be to find cures for cancer. Obviously if AI could design drugs to combat cancer, or even find the cure for it, the reverse is true: AI could bioengineer viruses as well. The answers regarding who would own the rights to those cancer-saving drugs, or legal liability for releasing bioweapons, rest in Thaler v. Vidal, 43 F.4th 1207, 1210 (Fed. Cir. 2022). There, the Federal Circuit ruled that AI cannot author or own queries it’s asked to research. In other words, the law must look to a human or legal entity rather than treating the AI itself as the responsible legal actor.
More contentiously, AI has also taken over software development that can take over the act of search, or even the act of coding itself. Factory is a company building AI agents that writes software for enterprises, instead of focusing on AI agents that work with individual software engineers. Yet in all these use-cases of software developing AI Agents, are the companies ultimately held responsible, or the users? The answer may depend on who directs the AI Agent, how much control that person or company exercises over its actions, and what the AI Agent was authorized to do.
Shifting Responsibility
For founders in the space, it still requires a bit of forethought into how the consumer will use their AI Agents. If founders can design their products in a way that shifts the responsibilities of a principal to the consumer, they will successfully shift liability for the products they profit from, and buffer software development mistakes with the use of AI insurance in their business activity.
For consumers, like law firms or schools who seek to integrate AI, it seems imperative to learn of the terms and conditions set against them, as no one can truly predict negative ramifications of Agentic AI mistakes.
The technology may change, but the legal question remains: when an AI Agent acts, who is the principal?
*The views expressed in this article do not represent the views of Santa Clara University.
Comments