Texas TRAIGA

• TRAIGA is a narrow, intent-based AI law. Early drafts copied Colorado's high-risk model; the enacted version dropped that and mainly lists prohibited uses (self-harm incitement, intentional discrimination, sexual content involving children, rights violations).
• Several duties land only on government entities: AI disclosure to the public, no social scoring, no biometric identification that infringes rights.
• Health care providers must tell patients when AI is used in their care.
• Enforcement is Attorney General only, with a 60-day cure period, a presumption of reasonable care, and defences for companies that find problems through testing or follow the NIST AI RMF Generative AI Profile or a similar framework.
• It also creates a 36-month regulatory sandbox run by the Department of Information Resources (DIR) and a seven-member Texas Artificial Intelligence Council (advisory only, no binding rules).

• Anyone who promotes, advertises or conducts business in Texas, produces a product or service used by Texas residents, or develops or deploys an AI system in Texas (Sec. 551.002).
• Extraterritorial: yes. An Indian SaaS company whose product is used by Texas residents is in scope, even with no Texas office.
• "Developer" and "deployer" are defined broadly (Sec. 552.001). "Consumer" means a Texas resident acting in a personal or household capacity, not in a work context (Sec. 551.001(2)).
• Government-only duties exclude hospital districts and public universities (Sec. 552.001(3)).
• Insurance entities under insurance anti-discrimination law, and federally insured financial institutions that follow banking law, get carve-outs from the discrimination ban (Sec. 552.056(d), (e)).
• The chapter overrides city and county AI ordinances (Sec. 552.003).

• Exclusive AG enforcement, no private lawsuits (Sec. 552.101).
• Complaints come in through an online form on the AG website (Sec. 552.102). After a complaint, the AG can issue a civil investigative demand asking for a description of the system, training data, inputs, outputs, metrics, limits and monitoring (Sec. 552.103).
• Cure: the AG must send written notice naming the provisions breached and cannot sue for 60 days. If within that time you fix the issue and send a written statement with supporting documents and policy changes, no action can be brought (Sec. 552.104).
• Penalties for uncured violations (Sec. 552.105(a)): USD 10,000 to 12,000 per curable violation (or breach of your cure statement); USD 80,000 to 200,000 per uncurable violation; USD 2,000 to 40,000 per day for continuing violations. Injunctions, attorney fees and investigation costs are also recoverable.
• Defences (Sec. 552.105(c) to (f)): rebuttable presumption of reasonable care; no liability if someone else misused your system, or if you found the violation through feedback, adversarial or red-team testing, state agency guidelines, or internal review while substantially complying with the NIST AI RMF Generative AI Profile or another recognised framework. No penalty for systems not yet deployed.
• Licensing agencies may suspend or revoke a licence and fine up to USD 100,000, but only after a court finding under 552.105 and an AG recommendation (Sec. 552.106).
• Sandbox participants are shielded from waived laws during testing, but the prohibitions in Subchapter B can never be waived (Sec. 553.051(c) to (e)).

• Most Indian SaaS and IT firms face low day-to-day burden: the bans require intent, and a disparate impact alone does not prove intent to discriminate (Sec. 552.056(c)).
• Selling to Texas state or local government (GovTech, chatbots, citizen services): your client must disclose AI to every user, even when it is obvious. Build a clear, plain-language disclosure with no dark patterns into the product. Social scoring and rights-infringing biometric ID features will be unsellable.
• Health-tech and clinical AI vendors with Texas provider customers: the provider must disclose AI use by the first day of service. Offer patient-facing disclosure text and workflow hooks.
• GenAI chat, companion or content tools used by Texas residents: put guardrails against content that encourages self-harm, violence or crime, and block child-impersonating sexual chat. These are the cases most likely to draw AG complaints.
• Data processors under the Texas privacy law: HB 149 adds AI system data to what you must help controllers secure (Sec. 541.104(a)(2)).
• Keep evidence ready. A civil investigative demand asks for exactly what an AI system card and NIST AI RMF and GenAI Profile profile contain. Aligning to the GenAI Profile also builds the statutory defence.

References are to the Texas Business and Commerce Code as added or amended by HB 149.

• AI system inventory flagging Texas exposure, government customers and health care use.
• Screenshots of AI disclosures in government and health care deployments, with timing (before or at interaction; by first date of service).
• Content safety policy and test results covering self-harm, violence, crime and child sexual content.
• Fairness testing results and a written statement of each system's intended purpose (useful to rebut any claim of discriminatory intent).
• Biometric data register showing consent source and whether training data was reused commercially.
• System documentation pack matching the eight items in Sec. 552.103(b).
• Red-team and adversarial testing reports, feedback intake logs, internal review records.
• NIST AI RMF GenAI Profile alignment or ISO/IEC 42001 certificate.
• A cure playbook: who receives AG notices, how fixes are documented, template cure statement.
• Sandbox application and quarterly reports, if participating.

• Assuming TRAIGA is a Colorado-style high-risk law with impact assessments. It is not; it is mostly prohibitions plus public-sector and health care disclosure.
• Assuming the consumer disclosure duty applies to every business. In the enacted text it applies to government agencies, plus health care providers.
• Ignoring it because there is no Texas office. Products used by Texas residents are enough.
• Treating the cure period as a reason to wait. Uncurable violations carry up to USD 200,000 each, and a breach of a cure statement is itself a violation.
• Relying on a generic "we follow NIST" line. The defence names the Generative AI Profile and requires substantial compliance plus a discovery route.
• Forgetting that the sandbox never waives the Subchapter B prohibitions.

• Deployer: a person who puts an AI system to use in Texas.
• Developer: a person who builds an AI system that is offered or provided in Texas.
• Civil investigative demand (CID): a formal AG order to hand over documents and answers during an investigation.
• Cure period: the 60 days the AG must give you to fix a violation before suing.
• Social scoring: rating people on behaviour or traits so they are treated worse in unrelated settings.
• Disparate impact: a neutral rule that happens to affect one group more; under TRAIGA it does not by itself prove intent.
• Regulatory sandbox: a supervised test zone where some licensing rules are waived for a limited time.
• NIST AI RMF Generative AI Profile: the US standards body's companion guide applying its AI risk framework to generative AI.
• Dark pattern: a design trick that nudges users into choices they would not otherwise make.

• In force since 1 Jan 2026; signed 22 Jun 2025; reaches anyone whose product is used by Texas residents.
• Core duties are prohibitions based on intent, AI disclosure by government agencies and health care providers, and government limits on social scoring and biometric ID.
• AG-only enforcement with a 60-day cure window; fines run up to USD 200,000 per uncurable violation and USD 40,000 per day for continuing ones.
• Red teaming plus substantial compliance with the NIST GenAI Profile gives a real legal defence.
• No amendments found as of 2026-10-19; the next regular session is in 2027.

• HB 149 bill history, Texas Legislature Online, accessed 2026-10-19
• HB 149 enrolled text (HTML), accessed 2026-10-19
• Texas Attorney General, Texas Responsible Artificial Intelligence Governance Act page and complaint link, accessed 2026-10-19