Texas TRAIGA in Force: What the Responsible AI Governance Act Requires From January 2026

Texas passed an AI law that was widely feared, then quietly cut most of the feared parts out before it took effect. The Responsible Artificial Intelligence Governance Act is now in force from 1 January 2026, and the version that survived is narrower than the headlines suggested and easier to underestimate as a result.

Most of what was written about TRAIGA during the legislative session described a bill that no longer exists. The original draft carried impact assessments, employer disclosure duties, risk-management policy mandates and disparate-impact liability. The enacted law, House Bill 149, dropped all of them. What is left is a prohibition-and-disclosure statute enforced by the state Attorney General, with an intent standard doing most of the load-bearing work. If your compliance reading is still based on the spring commentary, you are preparing for the wrong law.

Who is actually caught

The scope is deliberately extraterritorial in the way state consumer-protection statutes usually are. TRAIGA reaches anyone who conducts business in Texas, produces a product or service used by Texas residents, or develops or deploys an AI system in the state. You do not need a Texas office to be inside it. If Texas residents use your product, you are in scope, and the definition of an AI system is broad: any machine-based system that infers from inputs how to generate outputs, whether content, decisions, predictions or recommendations.

That breadth of scope is the part that matters for firms operating well beyond Texas. The substance of the obligations is narrow, but the door you have to walk through to check whether they apply to you is wide. The practical consequence is that the first piece of work is not a control build. It is a mapping exercise: which of your AI-touched processes are used by Texas residents at all.

The prohibitions, and the word doing the work

TRAIGA bans a short list of uses. It prohibits developing or deploying an AI system with the sole intent of inciting or encouraging a person to physically harm themselves or others, or to engage in criminal activity. It prohibits systems intended to produce child sexual abuse material and certain non-consensual explicit deepfakes. It prohibits intentional unlawful discrimination against a protected class. And it prohibits government agencies from using AI for social scoring based on personal characteristics.

The load-bearing word throughout is intent. The discrimination prohibition is the clearest example: it requires intent to discriminate, and the statute is explicit that a disparate outcome alone does not establish a violation. This is a very different posture from the impact-assessment regimes emerging elsewhere, where you have to test for and document disparate effect regardless of intent. Texas chose the opposite standard. That makes the prohibitions genuinely hard for the Attorney General to prove, and it means the day-one exposure for a firm using AI for ordinary commercial purposes is low. The risk is not that a well-run deployment trips the ban by accident. The risk is that you never checked, and a use case you inherited or bought was built for something the statute names.

The disclosure triggers are where the routine work sits

Two disclosure duties survived. Government agencies interacting with consumers through an AI system must disclose that fact, clearly and in plain language, before or at the time of the interaction. Healthcare providers must tell patients when AI is used in their care or treatment. Both are narrow, and neither reaches most private-sector chatbots or recommendation engines directly. But if you supply AI to a Texas state agency or to a healthcare provider, their disclosure obligation becomes your contractual and technical problem: they cannot disclose what your system does if you have not told them, in terms precise enough to put in front of a member of the public.

This is the quiet integration cost of the law. The prohibitions are a screening exercise you do once. The disclosure triggers are a standing requirement to be able to say, on demand, where AI sits in a customer-facing or patient-facing flow, and to surface that to the deploying party in a form they can pass on. Firms that cannot produce a current inventory of where their models actually touch an end user will find this harder than the statute makes it sound.

The sandbox: a route worth understanding, not always using

TRAIGA creates a regulatory sandbox administered by the Texas Department of Information Resources, with oversight from a newly established Texas Artificial Intelligence Council. Approved participants can test AI systems for up to 36 months under modified regulatory conditions, filing periodic performance reports in exchange for temporary relief from certain licensing and enforcement pressure. It is aimed squarely at firms deploying AI in constrained or novel contexts who want state-supervised cover while they prove a system out.

For most firms the honest answer is that the sandbox is not for you. The core TRAIGA duties still apply inside it, the reporting obligation is real, and admitting a system to a government programme means handing over detailed descriptions of what it does and how it fails. If your Texas-facing AI use is ordinary commercial deployment that clears the prohibitions comfortably, the sandbox adds obligation without removing risk. It is genuinely useful for a specific class of experimental or high-scrutiny deployment, and a distraction for everything else. Decide which you are before you treat it as an opportunity.

Enforcement, and why the penalties reward acting early

Enforcement rests exclusively with the Texas Attorney General; there is no private right of action, which removes the class-action tail that makes some US privacy statutes so expensive. The Attorney General can issue a civil investigative demand for system descriptions, training data, inputs, outputs and safeguards, and there is a 60-day cure period after a notice of violation. Penalties run from roughly 10,000 to 12,000 US dollars for a curable violation, up to 80,000 to 200,000 dollars for an uncurable one, with a daily charge for continuing breaches. The statute also recognises a safe harbour for firms that discover and cure problems through red-teaming or by following a recognised framework such as the NIST AI Risk Management Framework.

That structure rewards the same discipline good AI governance rewards everywhere: know what you have deployed, be able to describe it, and be able to show you were looking. This is the pattern now repeating across US state AI law, from Colorado’s reset of its own regime to the disclosure duties arriving in California, and it runs in parallel with the deferrals and staged deadlines of the EU AI Act. The firms that will struggle with TRAIGA are not the ones with aggressive use cases. They are the ones carrying governance debt — model deployments nobody has inventoried, provenance nobody can reconstruct, disclosure paths nobody has mapped.

TRAIGA is a light-touch law with a wide catchment. The mistake is not over-compliance. It is assuming a narrow statute means nothing to do, and finding out during a civil investigative demand that you cannot describe your own systems.