Courts are entering a new phase in AI law as disputes move beyond static models and into the world of systems that can plan, decide, use tools, and act across digital environments. In that setting, the core legal question is no longer only whether a model generated inaccurate or harmful output. It is increasingly whether agentic AI should be treated more like a passive instrument or more like an actor whose conduct can trigger familiar liability doctrines such as agency, tort, product liability, and professional supervision.
That shift matters because liability becomes harder to assign when harm emerges from interactions among developers, deployers, users, and semi-autonomous systems. Recent legal analysis suggests courts are already testing the law of agency in current disputes, while legislators, scholars, insurers, and regulators are building frameworks for a future in which AI agents may independently initiate actions with real-world consequences. The result is a fast-forming debate over whether existing law can stretch to meet the technology or whether new legal structures will be needed.
Why agentic AI creates a distinct liability problem
Agentic AI differs from conventional software because it can do more than respond to prompts. These systems may pursue goals, break tasks into steps, call external tools, interact with users or third parties, and sometimes modify digital environments over time. That expanded operational role makes it harder to describe the system as merely a neutral conduit, especially when its behavior becomes part of the chain of events that causes harm.
Recent legal commentary reflects this concern. Analysts note that courts are starting to confront whether agentic systems should be treated less like passive tools and more like actors for purposes of assigning responsibility. Once that question is asked, traditional legal categories become unstable. A developer may argue it only built a general-purpose system, a deployer may say the user controlled the outcome, and a user may claim the AI acted unpredictably or exceeded instructions.
This is why liability allocation in agentic AI cases is expected to become especially complex. The core issue is not simply fault in the abstract, but how courts should map responsibility across multiple human and machine participants. Current reporting and scholarship consistently identify agency, tort, product liability, supervision, and insurance as the main pathways through which that mapping is likely to occur.
Courts may lean on agency law before creating new doctrines
One of the most immediate legal tools available to courts is agency law. If an AI agent acts on behalf of a company, professional, or user, judges may ask whether the system functioned in a way analogous to an agent carrying out delegated tasks. That does not require granting AI personhood. It only requires a court to determine whether the human or firm behind the system should bear responsibility for acts performed through it.
Recent reporting suggests this line of reasoning is already gaining traction. A Washington Post report from July 13, 2026, indicated that judges may be motivated to impose liability where plaintiffs can show that a chatbot did something that would have been illegal if a human had done it. The report quoted Gabriel Weil of the Institute for Law & AI on the likelihood that courts will seek some form of corporate responsibility rather than allow harmful conduct to disappear into technical complexity.
That practical instinct could be decisive in early cases. Courts often prefer adapting established doctrines before endorsing sweeping legal innovations. If judges conclude that firms cannot avoid accountability merely by inserting an AI system between themselves and the harmful act, agency concepts may become a bridge doctrine. They would allow courts to assign responsibility without first resolving larger philosophical questions about machine autonomy.
Tort law is being pushed to account for nondeterminism
Tort law is another central arena for the liability debate, especially because agentic systems may behave unpredictably even when used as intended. A Yale Law Journal article published in May 2026 argues that tort law should explicitly account for “nondeterminism” in agentic contexts. The article contends that a legal system recognizing nondeterminism, particularly in safety-critical and agentic settings, would better protect users and discourage reckless deployment.
This argument is significant because defendants may try to frame unpredictability as a reason to reduce their responsibility. The Yale analysis points in the opposite direction. If nondeterminism is a known property of the system, then deploying it in contexts where mistakes can injure people may itself support a finding of negligence or another tort theory. In that view, uncertainty is not a defense. It is part of the risk profile that responsible actors must manage.
For future litigation, that approach could reshape how courts evaluate foreseeability, reasonable care, and causation. Instead of asking whether a precise harmful output could have been predicted, judges may ask whether it was foreseeable that an agentic system with open-ended capabilities could create a category of harm. That would make risk management, monitoring, testing, and deployment constraints legally important long before a specific failure occurs.
Product liability and supervision claims are likely to expand
Product-liability doctrine may become especially relevant where agentic systems cause user harm through design defects, inadequate warnings, or unsafe deployment choices. An April 2026 IMF note observed that legal liability for agentic AI remains unclear but specifically highlighted product-liability issues where these systems cause harm. That matters because product liability can sometimes bypass the need to prove the same level of individualized fault required in negligence cases.
At the same time, professional supervision duties are becoming harder to ignore. A June 2026 Thomson Reuters legal guide warned that “agentic systems acting externally on client matters” create special risks of unauthorized commitments and insufficient oversight. In legal practice, for example, an AI system that contacts opposing counsel, files drafts, or gives advice without proper review could trigger malpractice or ethics-related claims even if no one intended the specific outcome.
These supervision concerns show why courts may not need entirely new rules to begin assigning responsibility. Existing doctrines already penalize professionals and firms that delegate sensitive functions without adequate review. As AI agents move from advisory roles into external action, plaintiffs will likely argue that the failure was not only in the model’s output but also in the human decision to let that system operate with too much autonomy.
Real cases are beginning to test AI harm theories
What was recently a theoretical debate is now entering active litigation. Bloomberg Law reported in July 2026 that Nippon Life Insurance Company of America filed a lawsuit in March 2026 alleging that ChatGPT engaged in the unlicensed practice of law by giving legal advice to a pro se litigant. Whatever the eventual outcome, the case illustrates how plaintiffs are starting to frame AI harm claims using familiar legal categories rather than waiting for AI-specific causes of action.
The significance of such disputes lies in the framing. If a chatbot performs conduct that would have been regulated or prohibited had a human done it, courts may become more willing to impose liability somewhere in the commercial chain. That does not mean an AI system itself must become a legal person. It means the legal system may refuse to let automation sever the connection between harmful conduct and accountable human organizations.
Early cases also function as experiments in proof. Litigants must show what the system was designed to do, how it was deployed, what instructions it received, what safeguards existed, and how the alleged harm followed. Those evidentiary fights will shape whether future courts view agentic AI failures as isolated misuse, foreseeable design problems, or manifestations of insufficient supervision by firms that benefited from the technology.
Scholars are proposing new frameworks without embracing AI personhood
Even as courts test old doctrines, scholars are designing new conceptual tools for more complicated human-AI chains. One 2026 paper proposes an “Algorithmic Corporation” or “A-corp,” a framework that would treat some AI systems as legally distinct entities for counting and responsibility purposes. The proposal is not simply about symbolism. It aims to help courts and lawmakers assign liability where multiple humans and AI agents jointly contribute to decisions and actions.
Another line of scholarship argues for “permeable” legal fictions rather than full AI personhood. A 2026 paper on “Operational Agency” says courts need an ex post evidentiary framework to trace causal interactions among individuals, firms, and AI systems. The point is to map culpability through the chain without pretending the AI has the moral or legal status of a human being. In practice, this could help courts identify where responsibility should attach while preserving a human-centered legal order.
These proposals reflect a broader recognition that ordinary attribution rules may struggle in dense, hybrid systems. When developers create base models, deployers fine-tune them, companies integrate tools, professionals supervise outputs, and users trigger workflows, liability can become diffuse. New frameworks seek to prevent that diffusion from becoming a loophole, while still preserving doctrinal flexibility for judges who want to rely on existing law where possible.
Legislators and states are signaling that AI is not a liability shield
Courts are not acting alone. State and federal policymakers are beginning to signal that the use of generative AI should not function as a defense against ordinary legal duties. Utah Code § 13-77-102, effective May 6, 2026, states that it is not a defense to a consumer-protection violation that generative AI was used. That provision is notable because it directly rejects a likely argument that automation should somehow dilute accountability.
Congress is also exploring statutory routes. The AI Accountability and Personal Data Protection Act, S. 2367, would create a civil cause of action for misuse of covered data. While that proposal is broader than agentic AI alone, it shows that lawmakers are considering explicit liability mechanisms alongside judge-made doctrines. If enacted, such statutes could supplement tort, agency, and product-liability claims with clearer paths for plaintiffs.
These developments matter because they reduce the chance that legal uncertainty will be interpreted as legal immunity. Legislatures appear increasingly aware that if courts move too slowly or inconsistently, statutory interventions may become necessary. In that sense, the biggest open question is not whether liability will exist, but whether it will emerge mainly through adaptation of existing doctrines or through more targeted legislative design.
Insurance may become the practical backbone of accountability
Liability rules do more than assign blame after harm occurs. They also shape how risk is priced before deployment. A June 2026 paper on “Insurance of Agentic AI” argues that autonomous planning, tool use, and persistent modification of environments create a risk landscape that may require coordinated cyber, errors-and-omissions, product-liability, and AI-specific coverage. That view treats insurance as part of the governance architecture, not merely a financial afterthought.
Insurance markets can influence behavior by demanding documentation, controls, auditability, and human oversight as conditions of coverage. If underwriters conclude that agentic systems create novel or poorly bounded exposures, they may require firms to adopt stricter governance practices. In effect, insurers could become private regulators helping translate abstract legal concerns into concrete operational standards.
This insurance dimension also underscores why courts’ early decisions will matter beyond the courtroom. The more judges recognize claims based on agency, supervision, product defects, or foreseeable nondeterministic harms, the easier it becomes for insurers to model exposure. Over time, that may support a more stable ecosystem in which firms cannot externalize the cost of risky autonomy onto users, clients, or the public.
As courts weigh liability for agentic AI, the central challenge is likely to be preserving accountability in systems where control is shared, delegated, and sometimes obscured by technical complexity. The legal system already has tools that can address much of this problem, including agency law, tort principles, product liability, supervision duties, and consumer-protection rules. What remains uncertain is how aggressively judges will apply them when AI behavior appears novel or partially unpredictable.
The emerging consensus across scholarship, litigation, legislation, and insurance is that autonomy should not become a liability vacuum. Whether courts adapt old doctrines or lawmakers build new ones, the direction of travel is clear: firms and professionals that design, deploy, or rely on agentic AI will increasingly be expected to anticipate its risks, supervise its actions, and bear responsibility when those systems cause harm.
