TRAIGA and Governance for Physical AI

Which systems and businesses are covered

Section 551.001 defines an AI system by machine-based inference that generates outputs capable of influencing physical or virtual environments. A perception model directing a mobile robot can fall within that definition. The label attached to the product is less useful than its actual functions. A simple mechanism should not be assumed to meet every element merely because its seller calls it a robot. [1, § 551.001]

Section 551.002 reaches persons conducting or promoting business in Texas, producing products or services used by Texas residents, or developing or deploying AI in Texas. Chapter 552 separately defines developers and deployers. A hardware supplier, model developer, integrator, and fleet operator may occupy different roles, sometimes simultaneously. A contract describing one party as the developer does not resolve every statutory question about the work actually performed. [1, §§ 551.002 and 552.001]

What the Act prohibits

The provisions require careful attention to actor and intent. Section 552.052 prohibits development or deployment intentionally aimed at encouraging specified harmful or criminal conduct. Section 552.055 addresses a sole intent to impair federal constitutional rights. Section 552.056 prohibits intentional unlawful discrimination against protected classes and provides that disparate impact alone does not establish that intent. These provisions should not be rewritten as a strict liability rule for every unexpected robotic action. [1, §§ 552.052 and 552.055 through 552.057]

Governmental social scoring and certain biometric identification practices receive distinct treatment. Section 552.054 also makes a violation of the Texas biometric identifier statute a violation of that provision. A municipal robot recognizing visitors, an employer authenticating staff, and a consumer robot navigating around faces therefore require different analyses of purpose, actor, information, and applicable exceptions. CUBI and the TDPSA remain separate sources of obligations. [1, §§ 552.053 and 552.054; 2; 3]

Disclosure depends on the deployment

Section 552.051 requires a governmental agency making an AI system available for consumer interaction to disclose the interaction before or when it occurs. The required notice must be clear, conspicuous, in plain language, and free of a dark pattern. Obvious robotic appearance does not remove that requirement. The provision separately addresses AI used in health care services or treatment, including timing and emergencies. It is not a blanket requirement that every private warehouse label every automated machine under TRAIGA. [1, § 552.051]

For a public reception robot, governance should identify how the disclosure reaches a person who cannot read the screen or hear its speech. That is a deployment design question connecting the statutory notice to accessibility and effective communication requirements that may independently apply. A sign placed elsewhere in the building is not automatically a sufficient answer for every consumer interaction.

Enforcement, cure, and statutory protections

The Attorney General generally has exclusive Chapter 552 enforcement authority; the chapter creates no private cause of action. Section 552.103 permits specified information requests through civil investigative demands, including system purpose, inputs, outputs, limitations, performance measures, and monitoring. Section 552.104 establishes notice and a 60-day cure process with supporting documentation and policy changes. Penalties and injunctions depend on the actual enforcement provisions. [1, §§ 552.101 through 552.106]

Section 552.105 contains a reasonable-care presumption and protections addressing another person's prohibited use and discovery through feedback, testing, applicable agency guidelines, or a qualifying internal review process associated with substantial framework compliance. The exact statutory conditions matter. Merely owning a risk assessment template does not establish that the company satisfies them. These protections operate within the identified TRAIGA enforcement action and should not be described as immunity from an injured person's product liability suit. [1, § 552.105(c) through (f)]

Building a useful robotics governance record

A practical program records the machine's intended task, deployment environment, responsible parties, data flows, operating limits, supervision, and foreseeable misuse. It then maps each legal duty to a person with authority to implement it. A deployment approval should identify the software version, tested configuration, permitted payload, communication assumptions, and conditions requiring suspension. These are recommended governance practices, rather than universal checklist items enacted by TRAIGA.

The distinction matters after a change. A robot approved to move sealed boxes in a restricted warehouse has not necessarily been assessed for public retail service or biometric customer identification. Change review should examine the new legal triggers, not merely whether the same chassis is being used. Testing results, complaints, near misses, and corrective decisions should feed that review. OSHA's application-focused technical guidance and the AI framework's risk-management approach support this operational method without turning either publication into a general robotics regulation. [4; 5]

Hypothetical deployment

A Texas museum proposes a speaking mobile guide that answers questions and recognizes returning visitors. The museum first determines its governmental status, consumer interaction, and any required AI disclosure. It then separates ordinary obstacle detection from biometric identification and examines CUBI, TDPSA coverage, permissions, and retention. Its deployment record assigns someone authority to disable recognition without disabling safe navigation. This illustrates how governance connects distinct duties; it does not assume that consent to enter the museum authorizes every use of a visitor's face. [1, §§ 552.051 and 552.054; 2; 3]