The AI Responsibility Chain

Most AI liability debates become confused because they treat “the AI company” as if one company controlled the entire system. Modern AI is usually a supply chain. A foundation-model developer may create the model architecture and training process. A cloud provider hosts inference. An application developer adds system instructions, retrieval, tools, memory and workflow logic. A business deploys that application inside a specific process. A manager decides the scope of authority. An end user gives instructions. The agent may then call third-party software, access financial accounts, send communications or control physical equipment.

Responsibility therefore has to be mapped to control. The foundation-model provider controls model design, some training choices, evaluation and disclosed limitations. The application builder controls how the model is prompted, which tools it can use and what guardrails exist. The deployer controls the business context, user population, permissions and operating environment. The human principal controls at least some goals and authorizations. The downstream service may independently control whether a transaction is accepted. Different legal doctrines can attach at different points.

Why the Machine Is Not the Defendant

Legal responsibility requires more than causal involvement. A storm can cause damage but is not sued for negligence. A hammer can cause injury but does not owe a duty of care. Software can be part of the causal mechanism without becoming the legal bearer of duties. Current U.S. law does not generally grant an AI model independent legal personality merely because it can generate language, make recommendations or initiate actions.

This matters because phrases such as “the AI decided” can obscure legally relevant human choices. A company chose to deploy the system. Someone decided which data it could access. Someone established whether a human would review the output. Someone selected the model provider. Someone accepted contractual terms and configured transaction limits. The legal system may eventually create special rules for highly autonomous systems, but current doctrine usually asks which recognized person or entity should bear responsibility for those decisions.

Corporate law already demonstrates that legal personality is a policy choice rather than a measure of intelligence. A corporation has legal personality because statutes and centuries of doctrine assign it rights and obligations. An AI system would require similar legal architecture: ownership rules, capital, liability, representation, jurisdiction, service of process, taxation and remedies. Merely being capable of conversation or planning does not create those institutions.

Intent When Software Acts

Intent is particularly difficult because many legal rules distinguish intentional misconduct from negligence or strict liability. Criminal law may require a mental state such as knowledge, recklessness or purpose. Fraud requires legally relevant knowledge or intent. Discrimination law may turn on statutory standards that differ by context. A software system does not automatically possess the legally recognized mental state of a human actor simply because its output appears purposeful.

The better approach is to examine human and organizational intent around the system. Did a person deploy the model to accomplish an unlawful objective? Did management knowingly ignore evidence that the system produced prohibited outcomes? Did a company instruct an agent to conceal facts or make misleading statements? Did an employee use AI to scale intentional wrongdoing? In those cases, the software may increase capability without changing the underlying human mens rea.

Harder cases arise when no human intends the specific harmful output. A generative system produces defamatory text, a recommender selects a dangerous option or an autonomous agent takes an unanticipated step. The legal analysis then shifts toward foreseeability, negligence, product design, supervision or statutory responsibility rather than pretending the model formed a human mental state.

Negligence and Foreseeability

Negligence is likely to become one of the most important general doctrines for AI because it is built around risk rather than intentional wrongdoing. In simplified terms, negligence asks whether a defendant owed a duty, breached the applicable standard of care and thereby caused legally cognizable harm. AI complicates each element but does not make the structure disappear.

The central question becomes reasonable deployment. Was the system appropriate for the task? Were known limitations disclosed? Was independent verification required for high-consequence outputs? Were permissions broader than necessary? Were foreseeable failure modes tested? Did the deployer monitor performance? Could a reasonable human intervention have prevented the harm? The standard of care will evolve as industry practice, professional rules, technical standards and regulatory expectations mature.

Foreseeability matters because advanced AI systems can behave unpredictably at the individual-output level while still having predictable classes of failure. A company may not foresee the exact false citation a model generates, but hallucinated citations may be a known category. A company may not foresee the exact unauthorized purchase an agent attempts, but excessive tool permissions may be a recognizable risk. “We could not predict that exact output” is different from “we could not reasonably predict this kind of failure.”

Product Liability and Design Defect

Product liability becomes difficult because software has historically occupied an awkward position between products and services under different state doctrines. Modern AI makes that boundary even less comfortable. A foundation model can be delivered as cloud service, embedded in a medical device, incorporated into a robot, packaged into enterprise software or modified continuously after sale. The same model family may therefore participate in legal relationships that look like services in one context and product components in another.

Product-liability analysis typically distinguishes manufacturing defects, design defects and inadequate warnings or instructions. AI maps imperfectly onto each category. Training or model generation is not manufacturing in the traditional factory sense, but Congress is already debating proposals that explicitly define AI “design” to include training data selection, testing, auditing and fine-tuning. The proposed AI LEAD Act would create a federal products-liability structure for AI systems and allocate potential liability between developers and deployers. That proposal is evidence that lawmakers see existing doctrine as uncertain; it is not current law.

The hardest design question is whether a model’s known statistical failure modes should be treated like defects, inherent limitations or risks to be managed through deployment. A language model that occasionally produces false statements may be functioning exactly as its statistical architecture predicts. Liability may therefore depend less on proving that one output was abnormal and more on whether the system was reasonably designed, represented and used for the particular context.

Agency Law and Autonomous Agents

AI developers increasingly use the word “agent” technically: software that can pursue a goal, use tools and take multiple steps. Legal agency is different. An agent in law acts on behalf of a principal under doctrines that determine when the principal becomes responsible for the agent’s acts. The linguistic overlap can create confusion because an AI agent is not automatically a legal agent merely because software engineers call it one.

Yet agency principles offer a powerful analogy. A business authorizes employees to act within defined authority. Third parties can sometimes rely on apparent authority. A principal can become bound by authorized acts and in some circumstances ratify unauthorized ones after the fact. AI systems similarly operate under permissions, credentials and delegated objectives. The legal challenge is to decide when those technical permissions should count as legally attributable authority.

The safest architecture is explicit delegation. An organization should be able to identify who authorized the AI system, what transactions it may initiate, what financial or legal limits apply, when human approval is required and how authority can be revoked. That technical architecture mirrors the legal architecture of agency rather than leaving attribution to be reconstructed after harm occurs.

When an AI Agent Makes a Contract

Autonomous contracting sounds futuristic, but federal law anticipated automated transactions decades ago. E-SIGN defines an electronic agent as a computer program or automated means used independently to initiate action or respond to electronic records without contemporaneous individual review. It also states that a contract cannot be denied legal effect solely because electronic agents participated in formation, provided the action is legally attributable to the person to be bound.

That does not mean every purchase made by an AI agent automatically binds its owner. Attribution still matters. Contract doctrines involving assent, authority, mistake, fraud, unconscionability and capacity continue to operate. If an agent exceeds its permissions, the dispute may resemble unauthorized employee action, credential compromise or software error depending on the facts.

Modern AI raises the stakes because the agent may negotiate language dynamically rather than selecting from predefined options. One agent may converse with another agent, compare offers, revise terms and execute a transaction faster than a human can review each step. That increases the importance of machine-readable authority limits. Contracting systems may eventually need technical equivalents of corporate signature authority: maximum transaction value, permitted counterparties, prohibited clauses and mandatory human gates for specified commitments.

Professional Malpractice and AI Tools

Professional responsibility provides one of the clearest current answers to AI attribution: the professional remains responsible. ABA Formal Opinion 512 says lawyers using generative AI must continue to satisfy duties of competence, confidentiality, communication, supervision, candor and reasonable fees. The model does not become co-counsel in the legal sense and does not absorb the lawyer’s professional duty.

Federal courts are reinforcing the same principle procedurally. Courts have issued standing orders and notices warning lawyers and litigants that AI-generated filings remain subject to ordinary certification and accuracy requirements. A 2025 notice from the U.S. Bankruptcy Court for the Northern District of Indiana expressly warned that unverified AI-generated research can implicate Federal Rule of Civil Procedure 11. Several courts now require disclosure or certification concerning generative-AI use.

The principle will likely generalize to medicine, accounting, engineering and financial advice. Professionals can use sophisticated tools to improve performance, but licensed duties are not transferred automatically to software. If a surgeon uses AI decision support, the legal standard may still ask whether the professional used the system reasonably. If an investment adviser uses an automated allocation engine, fiduciary and disclosure duties remain with the adviser or regulated entity.

Employment Decisions

Employment law demonstrates how existing statutes can reach AI-assisted decisions without waiting for a comprehensive AI code. In February 2026, DOJ’s Civil Rights Division settled allegations that Elegant Enterprise-Wide Solutions posted job advertisements generated by an AI tool that unlawfully restricted jobs based on citizenship status. The settlement is unusually instructive because the prohibited text was machine-generated, yet responsibility remained with the employer that used the advertisement.

The lesson is broader than one immigration statute. An employer that uses software for recruiting, screening, ranking or promotion still operates inside employment law. The exact legal standard can depend on the statute, jurisdiction and theory of liability, and current federal doctrine regarding disparate-impact liability is itself contested and changing. The stable point is simpler: automation does not create an employment-law exemption.

This should affect procurement. Employers need to know what an automated hiring system measures, what data it uses, how performance is validated and how outputs are reviewed. Contract language allocating responsibility between vendor and employer cannot necessarily eliminate statutory obligations owed to applicants or employees.

Credit and Black-Box Decisions

Credit regulation offers an unusually direct response to AI opacity. CFPB guidance under the Equal Credit Opportunity Act and Regulation B states that creditors must provide specific reasons for adverse credit actions even when decisions are made using complex algorithms. A creditor cannot simply say the model is too complicated to explain. The legal duty to provide accurate reasons remains.

This matters because explainability is often treated as a technical preference. In regulated credit, it can become a compliance requirement. A model that produces strong predictive performance but cannot support legally required reasons may be unsuitable for the workflow regardless of its benchmark accuracy.

The principle is significant beyond lending. Regulation can convert interpretability, logging and traceability from desirable engineering features into legal infrastructure. A company cannot satisfy a statutory duty by saying the model’s internal complexity prevented compliance if the law requires a human-understandable explanation.

Finance, Fiduciary Duty and Robo-Advice

Automated financial systems have existed long enough to demonstrate another principle: automation does not erase conflicts of interest. In March 2026, the SEC announced a settled order against Ally Invest Advisors involving disclosure failures related to robo-advisor accounts. The case concerned the adviser’s cash-allocation methodology and an undisclosed conflict, not generative AI, but that is precisely why it matters. The regulatory duty attached to the investment adviser regardless of automation.

SEC officials are now openly considering a future in which investment firms use autonomous AI agents to explain funds, provide information or interact with investors. A February 2026 SEC speech raised questions about whether such agents might constitute marketing, require registration or trigger supervision obligations. Those questions remain open, but they reinforce that autonomous interfaces will be analyzed through existing securities-law categories before law invents entirely new ones.

Financial AI therefore has at least three layers of responsibility: whether the model works as represented, whether the regulated entity supervises it appropriately and whether the business model creates conflicts that must be disclosed. An accurate algorithm can still participate in unlawful conduct if the surrounding institution violates its duties.

Consumer Protection and AI Claims

Consumer protection may become one of the fastest-moving enforcement areas because Section 5 of the Federal Trade Commission Act already prohibits unfair or deceptive practices. Companies that market AI accuracy, objectivity, privacy or performance can therefore face ordinary substantiation requirements. In May 2026, the FTC announced settlements with Cox Media Group, MindSift and 1010 Digital Works over allegations involving deceptive claims about an AI-powered “active listening” advertising service.

The FTC also finalized an earlier order against IntelliVision involving allegedly false or unsubstantiated claims about facial-recognition accuracy and bias performance. The cases illustrate a basic rule: calling a product “AI” does not lower the evidence required to support advertising claims. If anything, AI’s opacity makes documentation more important because users may be unable to independently evaluate how the system works.

Current enforcement policy is also politically contested. In late 2025 the FTC reopened and set aside its prior Rytr order, concluding that the original theory had not satisfied the legal requirements of the FTC Act and had burdened AI innovation. In July 2026 the Commission proposed a new policy statement addressing alleged suppression of accuracy in AI systems. The direction of enforcement can change with administrations, which is another reason to distinguish durable statutory duties from temporary agency policy.

Cybersecurity and Autonomous Action

Agentic AI converts cybersecurity from protection of information into protection of authority. A conventional chatbot may reveal confidential information if compromised. An autonomous agent may possess credentials that let it send money, modify cloud infrastructure, delete files, approve requests or control physical devices. A security failure can therefore become an authorization failure.

Legal responsibility will depend heavily on reasonable security. Did the organization give the agent credentials broader than necessary? Were transactions rate-limited? Were high-risk actions separately approved? Could credentials be revoked quickly? Were tool calls logged? Did the system distinguish trusted instructions from untrusted content retrieved from the web? These are technical questions with obvious negligence and compliance implications.

This is where Article 003’s principle—authorize before the tool call—becomes a legal design principle. If a company cannot establish what the agent was allowed to do at the moment of action, it becomes harder to distinguish authorized behavior, software error, credential compromise and employee misuse after the harm occurs.

Robots and Physical Harm

Physical AI makes the responsibility chain tangible. A humanoid robot can knock someone down. An autonomous vehicle can collide. A warehouse system can damage property. A medical robot can affect a patient. The immediate cause may be a motion-planning decision made milliseconds before impact, but legal analysis will look backward through hardware design, sensor performance, software, maintenance, operating instructions, human supervision and the environment.

Traditional product-safety concepts remain relevant because physical systems have identifiable manufacturers and integrators. But continuously learning or remotely updated software complicates the point at which responsibility should attach. A robot may leave the factory safe, receive a software update months later and behave differently. A customer may modify the model, attach a new tool or operate the system outside intended conditions. The EU AI Act directly recognizes similar value-chain problems by shifting provider obligations when another party substantially modifies a high-risk AI system.

The likely future is layered responsibility rather than one defendant for every accident. Hardware manufacturer, autonomy provider, system integrator, owner and operator may each control different causes. Courts and regulators will need technical evidence capable of separating them.

The Causation Problem

AI systems make causation difficult because the harmful output may emerge from interactions rather than one identifiable defect. A foundation model contributes general capability. System instructions shape behavior. Retrieval supplies external information. User prompts provide context. Tool APIs determine possible actions. A downstream system accepts the request. Which element “caused” the harm?

Law routinely handles multiple causes, but AI increases evidentiary complexity. Plaintiffs may need logs showing model version, prompt history, retrieved documents, tool outputs, approval events and software updates. Defendants may need the same records to show that another actor materially altered the system or ignored warnings. Without those records, causation becomes a battle of reconstruction and inference.

This makes logging more than a debugging convenience. It can determine whether responsibility is provable. An AI system that leaves no reliable record of why, when and under whose authority an action occurred may create legal uncertainty for every participant in the chain.

Why Auditability Becomes Legal Infrastructure

Traditional software can often be reconstructed through deterministic logs. Generative systems are more complicated because outputs may vary, context windows change and model providers update underlying systems. A defensible enterprise architecture therefore needs to preserve enough information to reconstruct the legally relevant event even if exact token-for-token reproduction is impossible.

Useful records may include the identity of the human principal, the agent instance, model version, system instructions, permissions, retrieved sources, tool calls, approval checkpoints, transaction results, risk flags and post-action review. Retention has to be balanced against privacy and security because storing every prompt indefinitely creates its own liability.

The EU AI Act’s documentation, human-oversight and post-market monitoring requirements demonstrate how regulation can formalize this concept. In the United States, sector-specific rules often reach similar outcomes indirectly through recordkeeping, explanation, professional supervision and ordinary discovery obligations. The engineering principle is the same: if an autonomous system can create legal consequences, the organization should be able to reconstruct the authority and evidence behind the action.

The American State-Law Patchwork

State AI regulation is changing rapidly, which makes date discipline essential. Colorado originally enacted a broad Artificial Intelligence Act in 2024. In May 2026, Senate Bill 26-189 repealed and reenacted the core automated-decision provisions with a new framework covering developers and deployers of automated decision-making technology used in consequential decisions. The Colorado Attorney General says the new law takes effect January 1, 2027 and is now in rulemaking preparation.

The significance is not that Colorado has solved AI liability. It is that state law increasingly distinguishes between developer and deployer responsibilities, requires documentation and creates consumer rights around consequential automated decisions. That allocation mirrors the broader responsibility-chain concept developing internationally.

Other states are pursuing different models around employment monitoring, school surveillance, chatbots, privacy and automated decisions. The result is legal fragmentation. National companies may have to map not only what an AI system does, but where the affected person lives, which sector is involved and which state law applies.

The European Value-Chain Model

The European Union’s AI Act is the most comprehensive enacted cross-sector AI regulation, but its implementation schedule changed materially in 2026. Regulation (EU) 2026/1744 amended the AI Act after delays in standards and national implementation infrastructure. Major high-risk obligations for Annex III systems are now scheduled to apply from December 2, 2027, while high-risk systems tied to certain regulated products under Annex I are scheduled from August 2, 2028.

The amendment also sharpened value-chain cooperation. Article 25 requires written agreements in relevant high-risk integrations specifying information, technical access and assistance needed for compliance. If a distributor, deployer or other party substantially modifies a high-risk system or effectively becomes the provider, regulatory responsibility can shift along the chain. The original provider may then have cooperation and documentation duties rather than remaining the sole responsible actor.

This is not the same as tort liability for damages. The AI Act primarily establishes regulatory obligations, conformity, documentation, oversight and enforcement. But the architecture is important because it rejects the idea that one model provider necessarily owns every downstream risk. Responsibility follows role and control.

The Federal Liability Debate

Congress has not enacted a single comprehensive federal AI tort code. Instead, competing proposals reveal different philosophies about where responsibility should sit. The proposed AI LEAD Act would create a federal products-liability framework with separate standards for developers and deployers and defines AI design broadly enough to include training, testing and fine-tuning. The proposed RISE Act takes a different approach, offering conditional developer immunity for certain AI errors in professional services when developers provide specified transparency materials.

These bills are not law. Their value for analysis is that they expose the unresolved policy choices. Should developers be liable for downstream professional errors? Should professionals absorb responsibility because they chose to rely on the tool? Should developers receive safe harbors for transparency? Should federal law preempt state tort rules? Should open-source developers be treated differently from commercial providers?

The debate will intensify as AI moves from recommendation to action. Liability rules create incentives. If responsibility always falls on deployers, model providers may underinvest in safety. If responsibility always falls on developers, customers may deploy systems recklessly. A durable framework will likely need responsibility proportional to control, knowledge and ability to prevent the specific harm.

Should AI Ever Have Legal Personality?

Giving AI systems legal personality is sometimes proposed as a solution to the attribution problem. The analogy is usually the corporation: corporations are artificial entities that can own property, enter contracts and be sued. Why not give an autonomous AI the same status?

The analogy fails unless the surrounding institutions exist. Corporations have owners, directors, assets, capital requirements, registered offices, governing documents and legal representatives. A judgment against a corporation can be enforced against corporate assets. An AI model has none of those features inherently. Declaring the software a legal person without requiring assets or accountable human governance could create a liability sink: responsibility moves from solvent humans and companies into an entity incapable of paying for the harm.

Legal personality may eventually make sense for narrowly defined autonomous economic structures if law deliberately creates capitalization, insurance, registration and governance requirements. But personhood should not be used as a rhetorical shortcut for machine sophistication. The immediate policy goal should be preserving traceability to the people and organizations that design, authorize and benefit from autonomous systems.

The SURVXCOM AI Liability Test

AI accountability should begin before litigation. Organizations deploying autonomous or decision-making systems can ask twelve questions that map technical architecture onto legal responsibility. The objective is not to predetermine liability in every jurisdiction, but to make responsibility reconstructable before harm occurs.

What to Watch Next

Agent contracting disputes. As AI agents begin placing orders, negotiating subscriptions and executing financial transactions, courts will need to determine when technical permission constitutes legal authority and when an agent’s mistake can unwind a transaction.

Federal AI-liability legislation. Watch whether proposals such as AI LEAD, RISE or successor bills move toward a durable allocation between developers and deployers. The critical issues will be preemption, safe harbors, open-source treatment and causation standards.

Colorado rulemaking. Colorado’s rewritten automated-decision framework takes effect January 1, 2027. Watch how the Attorney General defines documentation, developer/deployer obligations and consumer rights in final rules.

EU Article 25 implementation. The 2026 AI Act amendment makes value-chain cooperation more explicit while pushing back high-risk implementation dates. Watch provider contracts, technical documentation and substantial-modification disputes.

Professional malpractice cases. Bar guidance already makes clear that lawyers retain responsibility for AI-assisted work. Watch analogous standards harden in medicine, accounting, engineering and financial services.

AI audit evidence in court. Litigation will increasingly turn on whether parties preserved model versions, prompts, permissions, retrieved evidence and tool-call logs. Expect discovery rules and enterprise retention policies to evolve.

Autonomous-agent insurance. Insurers may become de facto standard setters by pricing risk according to tool permissions, human gates, audit logs and model-validation procedures.

Physical AI accidents. As humanoids and autonomous machines move into workplaces, product liability and occupational-safety law will confront systems whose behavior can change after software updates.

Consumer AI enforcement. FTC policy can change across administrations, but false or unsupported claims about AI performance, privacy or objectivity will remain a recurring legal risk under durable consumer-protection statutes.

Legal personality. Watch whether any jurisdiction seriously proposes capitalized, registered autonomous legal entities. The key test should be whether personhood improves accountability rather than allowing human actors to externalize liability.

The Machine Does Not Break the Chain

The most important fact about AI law may be that the legal system has seen versions of this problem before. Companies act through employees. Financial markets act through algorithms. Contracts form through automated systems. Manufacturers distribute products they cannot control after sale. Professionals rely on complex instruments. Corporations separate ownership from immediate human action. Law has spent centuries building doctrines that connect distributed action back to responsibility.

Artificial intelligence makes those doctrines harder to apply because the machine contributes more discretion. A traditional software system executes predetermined instructions. A modern model interprets context. An agent chooses among tools. A robot reacts to an environment. The causal chain becomes probabilistic, adaptive and sometimes difficult to reproduce. That technological change is real, but it does not logically require abandoning human accountability.

The stronger response is to redesign systems so that law can see the chain. Authority should be explicit. High-consequence actions should have defined human gates. Model and application versions should be identifiable. Known limitations should be documented. Tool calls should be logged. Deployers should know what systems they are putting into consequential workflows. Professionals should retain independent judgment. Contracts should allocate technical information and support without pretending private agreements can erase statutory duties to third parties.

The ultimate policy mistake would be creating an “autonomy gap” in which responsibility decreases as machine capability increases. That would invert the incentive structure. The systems capable of causing the greatest consequences would become the systems with the least identifiable accountability.

Article 003 framed the technical solution as bounded agency: identity, authority, scope, duration, human gates, revocation and evidence. Article 025 adds the legal reason those controls matter. They do not merely improve cybersecurity. They preserve attribution. They allow organizations, regulators, courts and injured parties to identify which actor controlled which decision at which point in the chain.

That principle will become more important when AI agents handle money, robots enter homes and workplaces, autonomous laboratories choose experiments and neural systems mediate human communication. The machine may become more capable, but capability does not have to become a shield between consequence and responsibility.

The law of autonomous machines will ultimately be judged by whether technological delegation can expand without allowing accountability to disappear into the software. A mature legal architecture should let organizations delegate more capability while making authority, causation and remedy easier to trace rather than harder.