California AB 316 and the EU AI Act create binding AI liability for developers and deployers; SEC AI-washing penalties reached $400,000 in 2024.
AI Governance and Risk Liability: Research Report
The Innovation Attorney | Michael Kimball | mjk@theinnovationattorney.com | June 26, 2026
AI liability in 2026 is not theoretical: it is an active enforcement and litigation category via State, Federal and EU laws and regulations. California AB 316, effective January 1, 2026, eliminates the defense that an AI system acted autonomously, making developers, integrators, and deployers jointly exposed across the entire supply chain. The EU AI Act imposes penalties of up to 35 million euros or seven percent of global annual turnover for prohibited AI practices, with its full high-risk obligations enforceable from August 2, 2026. At the federal level, the SEC has already extracted $400,000 in civil penalties from two investment advisers for AI-washing, and no coherent federal preemption of state AI law currently exists. Every organization that develops, integrates, or deploys an AI system, including autonomous agents and generative AI workflows, must treat legal risk as an operational reality, not a future concern.
A. Executive Summary
The legal and regulatory framework surrounding AI liability and risk has moved from preliminary guidance to enforceable obligation across multiple jurisdictions simultaneously. This report addresses five interlocking subtopics that together define the current AI liability terrain for attorneys, general counsel, compliance officers, and the executives they advise: liability for autonomous AI agents, mitigation of AI-generated misinformation in enterprise workflows, compliance with the EU AI Act as enforcement begins in August 2026, compliance with state-level AI laws across the fifty states, and compliance with the growing body of federal AI regulation.
The most urgent near-term development is the convergence of state legislation and EU regulation creating enforceable standards before any coherent federal framework has been finalized. California AB 316 removes the autonomous operation defense. Colorado’s original SB 24-205 has been repealed and replaced by SB 26-189, effective January 1, 2027, but the compliance obligation is not suspended during the gap. Texas enacted the Responsible AI Governance Act effective January 1, 2026. Illinois prohibits AI-driven employment discrimination under HB 3773. The EU AI Act’s full high-risk obligations apply from August 2, 2026, with penalties already enforceable for prohibited practices since February 2, 2025.
AI hallucination liability has become a concrete legal risk. The Southern District of New York sanctioned attorneys in Mata v. Avianca, Inc. in June 2023 for submitting fabricated AI-generated citations, and over 700 court cases now involve AI-generated misinformation or fabricated content. The Air Canada Moffatt decision in 2024 established that a company cannot disclaim responsibility for the outputs of its own customer-facing AI systems. Businesses must treat human oversight, audit trails, and contractual risk allocation as mandatory, not optional, features of any AI workflow.
Federal regulation is fragmented but not dormant. The NIST AI Risk Management Framework, published in January 2023 and made mandatory for federal agencies by Executive Order 14110, has become the de facto operational standard referenced by state safe-harbor provisions in Texas and California. The SEC’s 2026 examination priorities expressly target AI-related disclosures, and the FTC continues to apply Section 5 of the FTC Act to AI claims that overstate capability. The Trump administration’s December 2025 Executive Order signals a preference for federal preemption of conflicting state AI laws, but no implementing legislation has passed, and state laws continue to operate and create compliance obligations.
Detailed Findings
Subtopic 1: Liability for Autonomous AI Agents
Autonomous AI agents, which execute multi-step tasks, interact with external services, and take consequential actions without human review of each individual decision, create a liability attribution problem that traditional legal doctrine did not anticipate. The threshold question is not whether an agent caused harm but which entity in the development and deployment chain bears legal responsibility for that harm. Courts and legislatures in 2025 and 2026 have begun to answer that question directly.
California AB 316, effective January 1, 2026, prohibits any defendant who developed, modified, or used an AI system from asserting as a defense that the AI system autonomously caused the harm at issue. This provision covers the complete supply chain: foundation model developers, fine-tuners, system integrators, and enterprise deployers. The law does not impose strict liability; it removes one specific exculpatory argument while leaving intact the traditional negligence and product liability analysis. The practical consequence is that every entity in the AI supply chain must affirmatively demonstrate reasonable care rather than deflecting responsibility onto the machine.
Product liability doctrine offers a parallel route to developer accountability. Under the Restatement (Third) of Torts: Products Liability, a product is defective if it contains a manufacturing defect, a design defect, or an inadequate warning. Legal scholars and plaintiff counsel are actively applying these categories to AI systems: a model that produces systematically harmful outputs may present a design defect claim; a deployment without adequate disclosure of known limitations may present a failure-to-warn claim. The EU’s new Product Liability Directive, which member states must implement by December 9, 2026, explicitly classifies software and AI systems as products subject to strict liability for defective outputs.
The Moffatt v. Air Canada decision in 2024 from the British Columbia Civil Resolution Tribunal established that a company deploying a customer-facing AI system cannot disclaim responsibility for that system’s representations. The tribunal held that Air Canada was responsible for negligent misrepresentations made by its chatbot regarding bereavement fare policies, rejecting the argument that the chatbot constituted a separate legal entity. Singapore’s Model AI Governance Framework for Agentic AI, launched in January 2026 as the first national framework specifically designed for autonomous agents, establishes organizational accountability for agent behavior regardless of the degree of autonomy the agent exercises. NIST’s AI Agent Standards Initiative, launched in February 2026, provides interoperability, security, and evaluation standards that are increasingly referenced by risk management programs.
The Berkeley Technology Law Journal noted in June 2026 that multi-agent architectures, where one AI model orchestrates several subordinate agents, create compounded attribution problems. When a harm results from interactions among multiple autonomous systems operating across organizational boundaries, the traditional question of proximate causation becomes analytically difficult. Contracts among deployers of interconnected agent systems must address indemnification, limitation of liability, and audit rights with this compounding dynamic in mind.
Subtopic 2: Hallucination Mitigation in Workflows
AI hallucination refers to the generation of factually incorrect, fabricated, or contextually inappropriate content by a large language model presented with apparent confidence. In enterprise and legal workflows, hallucinations have produced sanctions, adverse judgments, and significant financial harm. The legal system’s response has shifted from expressing concern to imposing consequences.
Mata v. Avianca, Inc., decided in the Southern District of New York in June 2023, is the foundational case. Attorneys relying on ChatGPT submitted briefs citing fabricated cases, including fictitious holdings attributed to real courts. Judge P. Kevin Castel imposed $5,000 in sanctions under Federal Rule of Civil Procedure Rule 11 and required the attorneys to notify every judge whose name appeared in the fabricated citations. The American Bar Association followed in July 2024 with Formal Ethics Opinion 512, its first guidance on generative AI in legal practice, confirming that lawyers retain an absolute verification obligation for any filing regardless of how the content was generated. At least fifteen documented cases between June 2023 and late 2025 resulted in sanctions for AI-generated fabricated citations, and at least one federal court disqualified counsel as the remedy rather than imposing a monetary fine.
The financial scale of the problem extends beyond the courtroom. Legal analytics estimates place the cost of enterprise AI hallucinations at $67.4 billion in losses in 2024, with projected costs increasing substantially in 2025. These figures include direct losses, operational correction costs, and reputational harm. An estimated $14,200 per employee per year in lost productivity attributable to hallucination verification time compounds the direct loss figures.
Mitigation at the enterprise level rests on three documented practices: retrieval-augmented generation, human-in-the-loop validation workflows, and systematic audit trails. Retrieval-augmented generation constrains model outputs to a curated, organization-controlled corpus, reducing the probability of fabricated citations or outdated legal standards. Human-in-the-loop validation requires a qualified professional to review AI outputs before any consequential action or filing, a requirement that aligns with the ABA’s verification obligation. Audit trails create the documentary record needed to demonstrate good faith in litigation and regulatory proceedings.
Contractual risk allocation is a distinct obligation. Under the reasoning of Moffatt v. Air Canada, an organization that deploys a third-party AI tool and presents its outputs to customers or in legal proceedings cannot disclaim liability by pointing to the tool vendor. Master service agreements, AI-specific terms of service, and indemnification provisions must be negotiated with this principle in mind. Organizations that maintain documentation of testing protocols, model validation procedures, and incident response histories are better positioned to demonstrate reasonable care if a hallucination produces harm.
Subtopic 3: Compliance with the EU AI Act (Enforcement Beginning August 2026)
Regulation (EU) 2024/1689, the EU AI Act, is the world’s first comprehensive binding legal framework for artificial intelligence. Its phased enforcement schedule makes August 2, 2026 the most consequential near-term compliance deadline for any organization that places an AI system on the EU market or whose AI systems affect persons in EU member states.
Article 5 of the Act prohibits a defined set of AI practices without exception. These prohibitions have been enforceable since February 2, 2025. Prohibited practices include: AI systems that deploy subliminal or manipulative techniques to distort behavior causing significant harm; systems that exploit vulnerabilities based on age, disability, or socio-economic circumstances; biometric categorization systems that infer sensitive attributes including race, political opinions, or religious beliefs; and social scoring systems used by public or private entities. Penalties for prohibited practice violations reach 35 million euros or seven percent of global annual turnover, whichever is higher.
Article 6 defines high-risk AI systems by reference to Annex I and Annex III. Eight categories appear in Annex III: biometrics, critical infrastructure, education and vocational training, employment and worker management, essential private and public services (including credit scoring and health insurance pricing), law enforcement, migration and border control, and administration of justice. High-risk AI system providers must satisfy obligations under Articles 9 through 15: a risk management system, data and data administration requirements under Article 10, technical documentation, automatic logging, transparency toward deployers, human oversight measures, and accuracy and rigorousness requirements. These obligations apply from August 2, 2026.
Article 50 transparency obligations also commence August 2, 2026. Deployers of AI systems that interact directly with natural persons, including chatbots and voice assistants, must disclose the AI nature of the interaction. Organizations must also label AI-generated content published in matters of public interest.
Extra-territorial reach is explicit in the Act’s text. A company headquartered outside the EU that places an AI system on the EU market, uses an AI system with outputs in the EU, or provides AI systems affecting persons in EU member states is within scope. American companies serving EU customers through AI-powered products must conduct a classification analysis, assess whether any system falls within the high-risk or prohibited categories, and implement the applicable compliance architecture before August 2, 2026.
Subtopic 4: Compliance with Laws and Regulations of the 50 States
State-level AI regulation in the United States reached critical mass in 2025 and 2026. Lawmakers in forty-five states introduced over 1,500 AI-related bills during 2025 alone, building on the more than 1,200 bills introduced across all fifty states the prior year. A patchwork of enforceable obligations now exists across key commercial jurisdictions, with more statutes effective on the horizon.
California leads the commercial compliance burden. SB 53, the Transparency in Frontier Artificial Intelligence Act, effective January 1, 2026, requires developers of large frontier AI models trained using more than ten to the twenty-sixth power floating-point operations to publish risk frameworks, report safety incidents to the California Attorney General, and implement whistleblower procedures. AB 2013, also effective January 1, 2026, requires AI providers to disclose the datasets used to train publicly available systems, including descriptions of data sources, collection methods, and known limitations. SB 942, the AI Transparency Act, requires AI providers to watermark or otherwise label AI-generated content and to maintain content provenance records. California AB 316, discussed under Subtopic 1, removes the autonomous AI defense.
Colorado SB 24-205, the Colorado AI Act originally effective February 1, 2026, and later delayed to June 30, 2026, imposed a reasonable care standard on developers and deployers of high-risk AI systems to protect consumers from algorithmic discrimination. In May 2026, Colorado repealed and replaced SB 24-205 with SB 26-189, a narrower statute regulating automated decision-making technology that materially influences consequential decisions, effective January 1, 2027. The compliance gap between the old and new statutes does not eliminate legal exposure during the interim period; common law negligence and anti-discrimination obligations remain.
Texas enacted the Responsible AI Governance Act, effective January 1, 2026, which limits its compliance requirements primarily to government use of AI but imposes categorical bans on AI systems designed for behavioral manipulation, unlawful discrimination, violence incitement, and deepfake production of child sexual abuse material. Illinois HB 3773 amends the Illinois Human Rights Act to prohibit employer use of AI that discriminates against protected classes, creating an employment-specific AI liability channel in one of the largest commercial labor markets in the country. Utah SB 149, effective May 1, 2024, requires AI service providers to disclose when a consumer is interacting with an AI system.
The compliance challenge for multi-state operations is the absence of uniformity. Each state statute defines key terms differently, applies to different categories of AI systems, and imposes different remedies. A company operating AI systems that touch California, Colorado, Texas, Illinois, and Utah simultaneously faces at least five distinct compliance frameworks with conflicting definitional structures. Compliance programs must be designed modularly to accommodate jurisdictional variation rather than applying a single national standard that satisfies none.
Subtopic 5: Compliance with Federal Laws and Regulations
Federal AI regulation in 2026 is best understood as sector-specific enforcement by established agencies applying existing statutory authority, overlaid by two Executive Orders with conflicting trajectories and a voluntary technical framework that has achieved de facto mandatory status in some contexts.
The NIST AI Risk Management Framework, published in January 2023 as NIST AI 100-1, provides the most widely referenced operational structure for enterprise AI risk management in the United States. Its four-function architecture, GOVERN, MAP, MEASURE, and MANAGE, establishes the vocabulary and process categories used by sector regulators, state safe-harbor provisions, and procurement specifications. Executive Order 14110, issued in October 2023, made the NIST AI RMF mandatory for federal agencies and contractors. Texas and California state statutes expressly reference NIST AI RMF alignment as the basis for a compliance presumption or safe-harbor defense.
The SEC is the most active federal AI enforcement agency in the securities context. On March 18, 2024, the Commission filed its first AI-washing enforcement actions against two registered investment advisers, Delphia (USA) Inc. and Global Predictions Inc., extracting a combined $400,000 in civil penalties for misrepresenting the use of artificial intelligence in client communications and Form ADV filings. The SEC’s 2026 examination priorities expressly target AI-related disclosures, requiring that descriptions of AI usage in marketing materials, Form ADV filings, and client communications accurately reflect the nature, extent, and limitations of actual AI deployment. Investment advisers using AI for any client-facing function must ensure those representations are specific and verifiable, not boilerplate.
The FTC continues to apply Section 5 of the FTC Act as its primary tool for AI-related enforcement, targeting AI claims that overstate capability, rely on undisclosed processing of consumer data, or constitute unfair or deceptive trade practices. The EEOC has issued guidance on AI in employment decisions, applying existing Title VII and ADA frameworks to algorithmic hiring and performance management tools. The CFPB has issued guidance requiring explainability in credit decisions made with AI models, grounding the obligation in the Equal Credit Opportunity Act and the Fair Credit Reporting Act. The FDA applies its existing medical device and software-as-a-medical-device regulatory framework to AI systems in clinical decision support and diagnostic applications.
The December 2025 Executive Order issued by the Trump administration directs federal agencies to evaluate, restrict funding for, and potentially preempt state AI laws that conflict with a uniform national AI policy. As of June 2026, no implementing legislation has passed, and the preemption effect of the Executive Order on existing state statutes remains contested. Companies should not treat the Executive Order as eliminating current state compliance obligations.
Cross Cutting Themes
Three themes connect multiple subtopics and carry particular weight for compliance program design. First, the supply chain problem cuts across autonomous agent liability, hallucination mitigation, and both state and federal compliance: every entity that touches an AI system, from foundation model developer to enterprise deployer, carries residual legal exposure that cannot be fully transferred by contract. Second, documentation and audit trails appear as a mitigating practice in every subtopic, from demonstrating reasonable care under AB 316 to satisfying the ABA verification obligation to meeting the NIST AI RMF MANAGE function requirements. Organizations that treat documentation as a legal asset rather than an administrative burden will be better positioned across all five areas. Third, the absence of federal preemption creates a compliance multiplier for any organization operating AI systems across state lines or international jurisdictions, requiring modular compliance architecture rather than a single unified standard.
Open Questions
1. Whether the Trump administration’s December 2025 Executive Order on AI will produce implementing legislation sufficient to preempt state AI laws, and which state statutes would survive preemption analysis under the Supremacy Clause.
2. How courts will apportion liability in multi-agent architectures where harm results from interactions among autonomous systems operated by different organizations, particularly in the absence of contractual privity between the harmed party and the upstream model developer.
3. Whether the EU AI Act’s extra-territorial provisions will be enforced against US-based companies in the absence of a bilateral enforcement cooperation agreement between the EU and the United States.
4. How the ABA’s Formal Ethics Opinion 512 verification obligation applies to automated legal workflows, including contract analysis tools and AI-assisted due diligence systems, where no individual attorney reviews each AI output before it influences a legal decision.
5. Whether the NIST AI RMF safe-harbor or compliance presumption provisions in Texas and California will be extended to cover EU AI Act obligations under bilateral or multinational regulatory harmonization efforts.
6. How product liability doctrine will resolve the design defect versus manufacturing defect distinction for large language models that produce harmful outputs inconsistently, particularly when the same prompt produces harmful outputs in some instances and benign outputs in others.
Source List
Primary legal and regulatory sources:
California AB 316 (effective January 1, 2026): eliminates the autonomous AI defense for defendants in the AI supply chain.
California SB 53 (effective January 1, 2026): frontier AI model developer transparency and safety incident reporting requirements.
California SB 942 (effective January 1, 2026): AI-generated content disclosure and watermarking requirements.
California AB 2013 (effective January 1, 2026): AI training dataset disclosure requirements.
Colorado SB 26-189 (effective January 1, 2027): automated decision-making technology regulation replacing SB 24-205.
Texas Responsible AI Governance Act (effective January 1, 2026): government AI use regulation and categorical prohibitions.
Illinois HB 3773 (Public Act 103-0804): AI employment discrimination prohibition under the Illinois Human Rights Act.
Regulation (EU) 2024/1689, the EU AI Act: full text including Article 5 (prohibited practices), Article 6 (high-risk classification), Article 10 (data administration), and Article 50 (transparency).
NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023.
Executive Order 14110, October 2023: mandatory NIST AI RMF compliance for federal agencies and contractors.
Executive Order on AI, December 11, 2025: Trump administration national AI policy framework and state preemption direction.
Mata v. Avianca, Inc., S.D.N.Y., June 2023: Rule 11 sanctions for AI-generated fabricated citations.
Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 2024: deployer liability for AI chatbot misrepresentations.
SEC enforcement actions against Delphia (USA) Inc. and Global Predictions Inc., March 18, 2024: $400,000 combined penalties for AI-washing.
ABA Formal Ethics Opinion 512, July 2024: lawyer obligations when using generative AI in legal practice.
EU Product Liability Directive, implementation deadline December 9, 2026: software and AI classified as products subject to strict liability.
Singapore Model AI Governance Framework for Agentic AI, January 2026: first national framework for autonomous agent accountability.
NIST AI Agent Standards Initiative, February 2026: interoperability, security, and evaluation standards for AI agents.
SEC 2026 Examination Priorities: AI disclosure accuracy requirements for registered investment advisers.
Berkeley Technology Law Journal, June 2026: multi-agent AI liability attribution analysis.
Bibliography
American Bar Association. Formal Ethics Opinion 512: Generative Artificial Intelligence Tools. July 2024.
Baker Botts. ‘California Eliminates the Autonomous AI Defense: What AB 316 Means for AI Deployers.’ January 2026. https://www.bakerbotts.com/thought-leadership/publications/2026/january/california-eliminates-the-autonomous-ai-defense-what-ab-316-means-for-ai-deployers
Berkeley Technology Law Journal. ‘Multi-Agent AI is Outpacing the Liability Frameworks Built for Single-Agent Systems.’ June 2026. https://btlj.org/2026/06/multi-agent-ai-is-outpacing-the-liability-frameworks-built-for-single-agent-systems/
European Commission. Regulation (EU) 2024/1689 of the European Parliament and of the Council (EU AI Act). Official Journal of the European Union. 2024.
European Commission. EU AI Act Enforcement Timeline. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
Illinois General Assembly. HB 3773, Public Act 103-0804. Amending the Illinois Human Rights Act. 2023.
Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. June 22, 2023).
Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 2024.
National Institute of Standards and Technology. AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0). U.S. Department of Commerce. January 2023. https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
SEC. Order Instituting Cease-and-Desist Proceedings Against Delphia (USA) Inc. and Global Predictions Inc. Release No. IA-6573. March 18, 2024.
SEC. 2026 Examination Priorities. Division of Examinations. 2025.
Singapore Infocomm Media Development Authority. Model AI Governance Framework for Agentic AI. January 2026.
State of California. AB 316 (2025). Chapter [x], Statutes of 2025. Effective January 1, 2026.
State of California. SB 53, Transparency in Frontier Artificial Intelligence Act. Effective January 1, 2026.
State of Colorado. SB 26-189. Automated Decision-Making Technology Regulation. Effective January 1, 2027.
State of Texas. Responsible AI Governance Act. Effective January 1, 2026.
Trump, Donald J. Executive Order on Ensuring a National Policy Framework for Artificial Intelligence. December 11, 2025.
Trump, Donald J. Executive Order 14110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (incorporated by reference regarding NIST AI RMF mandate). October 2023.
Interested in analysis about the intersection of corporate law, securities regulation, tech, artificial intelligence, public policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/