There is a question that every organization building or deploying artificial intelligence in the United States should be asking right now — not because a federal statute demands it, but because courts, state legislatures, and regulators are already answering it on your behalf: Have you implemented a recognized AI risk management framework?
If the answer is no, or not yet, or we’re waiting to see what the law requires, that answer is increasingly insufficient. Across the country, the two most prominent voluntary AI governance standards — the National Institute of Standards and Technology’s AI Risk Management Framework (NIST AI RMF) and ISO 42001 — are being woven into the legal fabric of AI regulation through a combination of state legislation, pending bills, and judicial interpretation of negligence and product liability law.
The word “voluntary” is becoming a formality. Understanding how these frameworks are acquiring legal force, what that means in practice for AI compliance teams, product and engineering organizations, and business leadership, and what organizations should do before they find themselves on the wrong side of that question in a courtroom or a regulatory proceeding — that is the subject of this article.
From Best Practice to Legal Benchmark: How It Happened
Voluntary industry standards acquiring quasi-legal status is not new. It happened with cybersecurity frameworks, financial risk management protocols, and environmental safety standards long before AI entered the picture. The pattern is consistent: a voluntary framework emerges to fill a gap that formal law has not yet addressed, industry adoption spreads, and courts and legislators begin treating the framework as the benchmark for reasonable conduct.
The NIST AI RMF, released in January 2023, followed this trajectory at an unusually rapid pace. Developed through an extensive multi-stakeholder process and designed to be technology-agnostic, sector-neutral, and adaptable to organizations of any size, the framework organizes AI risk management around four core functions: Govern, Map, Measure, and Manage. It does not prescribe specific technical requirements. Instead, it establishes a structured approach to identifying, assessing, and addressing the risks that AI systems create across their full lifecycle.
ISO 42001, the international standard for AI management systems published in late 2023, complements the NIST framework by providing a formal management system structure — closer in design to ISO 27001 for information security — that organizations can implement, audit, and certify against.
Neither standard was written as law. Both are now being treated as if they were.
The State Legislative Landscape: Five Different Approaches to the Same Frameworks
What makes the current US regulatory environment for AI governance standards genuinely complex — and genuinely consequential — is not that one federal statute has mandated NIST AI RMF compliance. It is that five states have incorporated these frameworks into law through five meaningfully different mechanisms, creating a patchwork of obligations and protections that organizations operating across state lines must navigate simultaneously.
Colorado: The Mandate-Plus-Defense Model
Colorado’s AI Act (SB 205) was the first US law to require deployers of high-risk AI systems to implement a risk management policy and program that aligns with the NIST AI RMF, ISO 42001, or another nationally or internationally recognized AI risk management framework. That mandate was paired with an affirmative defense: developers, deployers, and other persons that comply with a recognized framework are protected against certain state AG enforcement actions.
Colorado therefore did something legally significant: it simultaneously treated framework compliance as a floor (you must do this) and a shield (if you do, you are protected). The Colorado AI Policy Working Group’s most recent proposed revisions have moved away from explicit NIST and ISO references, but the underlying model — mandated governance aligned with external standards, with legal protection for compliance — remains influential.
Texas: The Incentive Model
The Texas Responsible Artificial Intelligence Governance Act (TRAIGA) took a different approach. Rather than mandating framework compliance, Texas created an affirmative defense to regulatory actions for developers and deployers that comply with the NIST AI RMF or another nationally or internationally recognized AI risk management framework. No compliance required — but comply, and you are protected.
This is a purely incentive-based model. It does not obligate any organization to adopt any specific framework. It simply makes the calculation explicit: if you want protection from state enforcement actions, recognized framework compliance is how you get there. For organizations operating in Texas, the decision not to implement the NIST AI RMF is also a decision to forgo a meaningful legal defense.
California: The Transparency Model
California’s Transparency in Frontier Artificial Intelligence Act (TFAIA) requires frontier AI developers to disclose whether and to what extent they incorporate national standards, international standards, and industry-consensus best practices into their development and risk management processes. It does not mandate adoption of any specific framework. It requires public honesty about whether you have adopted one and, if so, how.
The compliance implication is subtle but real. An organization that cannot point to a documented, coherent approach to AI governance standards — one it is willing to describe publicly — is in a worse position under California law than one that can. The disclosure requirement creates a reputational and legal record. Organizations that disclose robust framework alignment benefit. Those that disclose minimal or no alignment face both reputational risk and a documented admission that may be relevant in subsequent litigation.
New York’s RAISE Act (A 9449) takes a similar approach, requiring developers to describe how they handle the incorporation of external standards into their AI systems.
Montana: The Sector-Specific Model
Montana’s SB 212 is narrower in scope than the other state frameworks, applying specifically to AI deployed in critical infrastructure contexts. It requires deployers to implement a risk management framework that considers external standards when AI is used in critical infrastructure settings. Montana’s approach signals that even states with limited general AI legislation are reaching for the same governance frameworks when the stakes are high enough.
The Pending Wave: What Is Coming
The legislative momentum behind AI governance standards is accelerating, not stabilizing. Pending legislation across multiple states is building directly on existing models, with three categories of bills driving the next phase of development.
Frontier model bills in Illinois (SB 3312, HB 4799, SB 3261) and other states expand California’s disclosure approach by requiring developers and deployers to actually implement — not merely disclose — a governance framework that incorporates or considers external standards. Notably, several of these bills avoid naming NIST and ISO specifically, instead referring to “industry-consensus best practices” and “national or international standards” — language that is broad enough to evolve as the standards landscape develops but narrow enough to exclude ad hoc or internally-defined governance approaches.
Liability bills in Illinois (SB 3502/SB 3590), Maryland (HB 712), and Vermont (H 792) adopt the Texas safe harbor model and extend it to product liability litigation. Under these bills, developers who conduct testing, evaluation, verification, validation, and auditing consistent with industry best practices, and who submit detailed data sheets to the state AG documenting their risk management approach, can access a safe harbor from product liability claims. Deployers face a comparatively lighter burden — implementing a risk management framework that incorporates external standards — to qualify for protection.
Automated decision-making technology (ADMT) bills in Washington (HB 2157) and New York (S 1169) revive Colorado’s hybrid model. Washington’s bill creates a presumption of statutory conformity for developers and deployers that follow NIST’s AI RMF or ISO 42001 — meaning that framework compliance effectively creates a rebuttable presumption of legal compliance. New York’s bill goes further, requiring implementation of a risk management policy conforming to NIST or a standard designated by the state Attorney General, and uniquely granting the AG authority to name qualifying standards — a mechanism that could keep the legal framework current as the standards landscape evolves.
The Courts Are Already There: How Voluntary Standards Are Shaping AI Litigation
State legislators are not the only ones treating AI governance frameworks as legal benchmarks. Courts are already doing it, without waiting for AI-specific statutes, through the application of established negligence and product liability principles.
This development matters enormously for organizations that do not operate in Colorado, Texas, California, Montana, or any other state that has passed AI-specific legislation. In every US jurisdiction, the question of whether an organization exercised reasonable care — the foundation of negligence liability — is answered in part by reference to industry standards and practices. Courts have long recognized that evidence of industry standards, customs, and practices is often highly probative when defining a standard of care, and that advisory guidelines and recommendations, while not conclusive, are admissible as bearing on the standard of care in negligence determinations.
Applied to AI, this means that a court assessing whether an AI developer or deployer acted reasonably will look, among other things, at whether the organization implemented a recognized governance framework. An organization that cannot demonstrate compliance with NIST AI RMF or an equivalent standard is vulnerable to the argument that it fell below the industry standard of care — even in a state that has passed no AI legislation whatsoever.
Negligence: The Duty of Care Question
For negligence claims, the existence and scope of a duty of care is determined by factors that include, prominently, industry standards and accepted practices. An AI developer that implemented the NIST AI RMF’s full governance cycle — identifying risks, documenting mitigations, establishing accountability structures, conducting regular reviews — is in a materially different legal position than one that deployed a system without documented risk management. The former can point to a structured, documented process aligned with recognized industry standards. The latter cannot.
Product Liability: The Defect and Warning Questions
Strict product liability does not require proof of negligence, but it does require establishing that a product was defective or that adequate warnings were not provided. In AI litigation — including the growing wave of chatbot cases — plaintiffs are increasingly pursuing product defect theories. Courts assessing whether an AI system was “defective” or whether warnings were “adequate” are likely to look at whether the developer followed industry standards for risk identification, testing, and user communication. The majority of US states and federal courts consider industry standards in strict liability cases, making framework compliance relevant even in jurisdictions with no AI-specific legislation.
The litigation in Garcia v. Character Technologies (M.D. Fla. 2025) illustrates how this is developing in practice. Chatbot-related product liability claims are testing the boundaries of AI developer responsibility, and the standards question — what did industry practice require? — is embedded in those cases whether or not it has been fully litigated yet.
Punitive Damages: The Good Faith Calculation
Even where liability is established, the question of whether punitive damages are warranted turns significantly on whether the defendant acted in good faith. Courts across the country have looked favorably on framework compliance as evidence of good faith conduct. However, framework compliance is not a blanket insulation against punitive damages — courts have imposed them where defendants followed industry standards but actively resisted safer designs for economic reasons, knew of remaining risks and failed to warn users, or engaged in conduct that knowingly endangered people despite formal compliance.
The lesson is not that framework compliance eliminates punitive damages risk. It is that non-compliance eliminates the good faith argument entirely, and that compliance creates the foundation for a good faith defense — a foundation that still requires the organization to have genuinely acted on the framework’s requirements rather than treated documentation as a box-checking exercise.
What This Means for Your Organization: A Function-by-Function Breakdown
The combined effect of state legislation, pending bills, and judicial interpretation is to create a legal environment in which AI governance framework compliance is, as a practical matter, the baseline for legally defensible AI development and deployment in the United States. The following is what that means for each organizational function.
For AI Compliance Officers and Legal Teams
The immediate priority is a gap analysis against the NIST AI RMF and ISO 42001, mapped against the specific states in which your organization develops, deploys, or markets AI systems. Colorado, Texas, California, Montana, and New York each carry distinct obligations and protections. Organizations that operate nationally need a compliance posture that satisfies the most demanding applicable state requirement while preserving the affirmative defenses available in others.
Beyond state-specific obligations, legal teams should begin building a litigation readiness file: documented evidence of framework implementation that can be produced in discovery to support a duty of care defense, a product liability defense, or a good faith argument against punitive damages. That file should include governance policies, risk assessment records, testing and evaluation documentation, and evidence of ongoing review and update processes. Documentation assembled after litigation begins is substantially less persuasive than documentation that predates the conduct at issue.
For AI Product and Engineering Teams
Framework compliance is not a legal team responsibility that lands on engineering’s desk at the point of deployment. The NIST AI RMF’s core functions — Govern, Map, Measure, and Manage — describe processes that must be integrated into the product development lifecycle from the earliest stages of system design. By the time a system is ready for market, the risk mapping exercise, the measurement approach, and the management protocols should already be documented and operational.
Engineering teams should also understand that the pending liability bills’ safe harbor requirements — testing, evaluation, verification, validation, and auditing consistent with industry best practices — describe technical activities, not legal activities. The data sheets that several bills require developers to submit to state AGs, documenting training data sources, foreseeable risks, risk mitigation steps, and red-teaming results, draw directly on technical documentation that engineering and data science teams are best positioned to produce.
Building that documentation infrastructure as a natural output of the development process, rather than as a retroactive compliance exercise, is both more accurate and more defensible.
For C-Suite and Business Leadership
The question of whether to invest in AI governance framework implementation is no longer a compliance question with a deferred answer. It is a risk management question with an immediate answer, and the answer is yes.
The cost calculation has shifted. Organizations that have not implemented recognized AI governance standards face litigation exposure in product liability and negligence cases across every US jurisdiction, forgo affirmative defenses and safe harbors available under Texas, Colorado, and pending state legislation, and are unable to make the good faith showing that courts consider when calculating punitive damages. They also face the competitive and reputational disadvantage of being unable to demonstrate, publicly or to regulators, that they govern their AI systems responsibly.
Against those costs, the investment in implementing the NIST AI RMF or ISO 42001 is not extravagant. Both frameworks are publicly available. ISO 42001 certification infrastructure is developing rapidly. The NIST AI RMF is designed to be scalable to organizations of any size. The barrier to implementation is organizational will, not resource impossibility.
One additional point that deserves emphasis at the leadership level: framework compliance that is genuine — meaning it is embedded in actual governance processes, reviewed regularly, and reflected in real development and deployment decisions — is worth materially more, both legally and operationally, than framework compliance that exists on paper but not in practice. Courts and regulators are developing the capacity to distinguish between the two. Organizations that treat these frameworks as documentation exercises rather than governance tools are building a compliance posture that will not hold under scrutiny.
The Broader Argument: Good Governance Before the Law Demands It
There is a compelling argument for AI governance framework adoption that has nothing to do with avoiding litigation or satisfying state mandates. The NIST AI RMF and ISO 42001 are well-designed tools for building AI systems that produce reliable, fair, and trustworthy outcomes. Organizations that implement them rigorously are better positioned to catch problems before they cause harm, build customer and partner trust, and engage with regulators from a position of demonstrated competence rather than reactive damage control.
In an environment where AI legislation is proliferating across fifty states at a pace that makes jurisdiction-by-jurisdiction compliance tracking genuinely difficult, a documented commitment to robust governance gives regulators reason to extend good faith when questions arise. An organization that can demonstrate systematic, principled AI governance is in a different conversation with a state AG than one that cannot.
The voluntary frameworks were designed by technical and policy experts to capture what good AI governance actually looks like. The legal system is now ratifying that judgment. Organizations that got ahead of this curve are already benefiting. Those that have not should treat the current moment — before a federal mandate arrives, before a specific liability ruling lands in their jurisdiction, before a state AG inquiry begins — as the optimal window to act.
AI Governance Standards
The AI governance standards landscape in the United States has crossed a threshold. NIST’s AI RMF and ISO 42001 are no longer simply best practices that responsible organizations choose to follow. They are the benchmarks against which courts measure reasonable care, the frameworks that state legislatures are encoding into affirmative defenses and compliance mandates, and the documentation infrastructure that separates organizations with a legally defensible AI governance posture from those without one.
The organizations best positioned in this environment are not those waiting for a federal statute to clarify their obligations. They are the ones that recognized, early, that the voluntary standard was becoming the legal standard — and built their governance infrastructure accordingly.
