SB1964

Which AI is covered? “Heightened scrutiny” systems are specifically intended to make consequential decisions autonomously or be a controlling factor in them. A consequential decision materially affects access to a government service; a controlling factor is the principal basis or can alter the outcome. Merely adding a human to the process does not necessarily remove a system from this definition. Excluded are systems intended for narrow procedural tasks, improving a completed human activity, preparatory work for an assessment, or detecting decision patterns or deviations. “Unlawful harm” means harm through a consequential decision to a member of a protected class in violation of law—not every unwanted outcome. Its exclusion for predeployment testing solely to ensure legal compliance is not blanket immunity for testing. Final law, Sections 1 and 5.

Ethics and risk standards: the Department of Information Resources (DIR) must make an ethics code for state and local government AI generally, and separate minimum standards for heightened-scrutiny systems. Covered governments must adopt them. The code addresses oversight, fairness and accuracy, transparency, privacy/security, redress and accountability, and evaluation. The standards require documented security risks, performance and transparency before deployment and after material changes to the system, data or intended use, plus guidance on training and contractual vendor risk management. The rules must align with or be consistent with NIST’s AI Risk Management Framework 1.0. NIST’s framework is voluntary by itself; SB1964 makes the specified Texas rulemaking and adoption duties mandatory. It is not a NIST certification requirement. Final law, Sections 2054.702–.703; NIST framework.

Inventories and scope: state infrastructure inventories add AI; deployment reviews add system purposes, risk mitigation and compliance. Local governments must review heightened-scrutiny deployment and use, but supply that review to DIR on request. Existing exceptions still matter: universities are excluded from the particular infrastructure report in Section 2054.068(g); public junior colleges/districts have a broader chapter exception, subject to specified security/shared-services requirements; lottery operations have planning/procurement exceptions. These duties should not be described as applying identically to every public institution. Final law, Sections 2–3; Infrastructure-report exception; Junior-college exception; Lottery exception.

Data officers and open data: agencies with 150 or fewer full-time employees may designate an employee or, with DIR approval, jointly employ a data management officer. The existing mandate for agencies with more than 150 remains. Agency data officers must now post at least three high-value data sets annually; this does not say three newly created data sets. Confidential or legally protected information remains excluded. Final law, Section 4; Earlier officer and posting requirements.

Advice, training and testing: DIR must provide educational materials, online training and statewide forums. An eight-member advisory board—six agency representatives, including one from an agency with fewer than 150 employees, and two public technology experts—assists agencies, shares resources and recommends improvements to procurement and innovation rules. Governor/designee appointees serve renewable two-year terms without compensation or expense reimbursement; this is an advisory body, not an AI licensing board. Final law, Sections 2054.704–.705.

DIR must also run a “sandbox”: controlled, limited predeployment testing with registered commodity-contract vendors and entities eligible for DIR services. Eligible customers extend beyond executive agencies, subject to the existing best-interest determination. Vendor applications require a system description, public-risk assessment and mitigation plan; participants provide quarterly reports and secure access under DIR oversight. DIR must specify regulatory exemptions, not grant an unrestricted exemption from all law. It may operate the sandbox as a statewide technology center and must share resources with other qualifying testing programs. Although the statute calls DIR’s legislative report “annual,” its explicit deadline is November 30 of each even-numbered year. Final law, Section 2054.706; Eligible DIR customers.

Two notice duties: Section 2054.707 requires state agencies to disclose public-facing AI interaction, unless a reasonable person would already know. Separately, Section 2054.711 requires state and local governments to put a standardized notice on related applications, websites and public computer systems when AI is public-facing or controls a consequential decision. DIR’s form covers the system/data sources and privacy/ethics measures. Listed academic medical centers, state-owned/public hospitals and hospital districts may satisfy this latter section with a generalized patient-consent statement that AI may be used in treatment. This alternative is not a general exemption from all AI duties. Final law, Sections 2054.707 and .711.

Assessments are not public reports: state agencies and their contracting vendors deploying or using heightened-scrutiny systems must assess unlawful-harm risks, system limitations and information governance. DIR can request the assessment. These assessments are confidential; agencies/DIR may withhold protected information without seeking an attorney-general ruling, and DIR must safeguard submissions. This provision does not impose a twice-yearly schedule in the final law or expressly require every local government to do the same assessment. Local reviews and risk standards are separate duties. Final law, Section 2054.708.

Complaints and enforcement: agencies/vendors aware of violations must report them to the attorney general and, if applicable, DIR. The attorney general reviews reports/complaints and decides whether to seek an injunction. A vendor gets written violation notice; failure to respond or cure before the 31st day after receipt leads to agency notice of intent to void the contract. The vendor then has another opportunity to respond and seek a cure before the 31st day after receiving that notice. If still uncured, the agency may void the contract. More than one voided contract triggers DIR referral; the comptroller may bar the vendor under existing procedures. The act does not create a private damages remedy. Final law, Section 2054.709; Vendor-debarment procedures.

The attorney general must create a complaint webpage with DIR, post rights information, send complaints to DIR and report complaints, resolutions and enforcement to the legislature by November 30 of each even-numbered year. Complainants may request an explanation from DIR. DIR must coordinate related AI work and adopt implementing rules. Rules, outreach, the notice form and complaint webpage are required “as soon as practicable” after effectiveness—not by an invented fixed deadline. Final law, Sections 2054.710–.713 and Section 6.

How the bill changed: the Senate committee substitute reorganized definitions, broadened the ethics code beyond heightened-scrutiny AI, clarified the advisory role, removed “semiannual” from assessments, limited interaction disclosure to public-facing AI with an obvious-interaction exception, and clarified the two-stage vendor notice/cure process. The Senate-passed operative text matched that substitute. The House substitute made local submission on request, narrowed assessments to heightened-scrutiny systems, removed the standalone vendor disclosure subsection, and added coordination/resource-sharing provisions. Introduced; Senate committee; Senate-passed; House committee; Version index.

All three indexed floor amendments: Capriglione’s adopted second-reading amendment 1 changed “material, legal” to “material legal,” removed “generated by an artificial intelligence system” from the controlling-factor definition and combined notice-form items; these changes survived. Adopted second-reading amendment 2 initially required standardized hospital disclosure in each patient-consent form. Adopted third-reading amendment 1 replaced that text with the broader listed-provider, permissive generalized-statement option described above. SCR55 then replaced “AI” in that amendment with “an artificial intelligence system” during enrollment. Do not treat amendment 2’s superseded mandatory wording as final law. Second-reading amendment 1; Second-reading amendment 2; Third-reading amendment 1; Enrollment correction; Amendment index; Final law.

Related AI law: HB149 separately amended the same inventory/review sections effective January 1, 2026. That does not by itself erase SB1964’s additions: Government Code 311.025 requires same-session amendments to be harmonized if possible and says reenacting unchanged text does not ordinarily override another amendment’s changes. HB149’s own funding contingency and penalties should not be imported into this act. This is a review of SB1964’s enacted changes, not a complete account of every later rule, overlapping AI law or individual institution’s compliance. HB149, Sections 6–10; Same-session amendment rule.

Compare the published text versions in the official text index. A change is not attributed to a specific amendment here without a verified record.

People dealing with government could gain clearer AI notices, risk controls and a complaint route. Agencies and eligible vendors gain shared guidance and controlled testing opportunities. Benefits depend on implementation and enforcement; the law does not guarantee error-free decisions, cost savings or compensation. Governments and vendors bear review, training, reporting and compliance work. Confidential assessments protect submitted information but limit public access to those assessments; the hospital consent option permits a more general disclosure than system-specific notice. No actual rules, appointments, sandbox operation, complaint webpage or measured outcomes are verified here. Final law.

Estimated public cost: the three indexed fiscal notes estimate a $7,280,205 General Revenue cost for fiscal years 2026–27: $3,255,299 in 2026 and $4,024,906 in 2027, with $4,024,906 annually projected through 2030. They assume ten additional DIR employees, $2,880,205 in biennial personnel costs and $4,400,000 in technology costs. Attorney-general legal work and comptroller costs are assumed absorbable within existing resources; local-government fiscal effects are indeterminate. The bill itself makes no appropriation, and these estimates are not actual expenditure totals or individual fees. April 2 fiscal note; April 16 fiscal note; May 18 fiscal note.

Important fiscal-note limits: their narrative retains proposal descriptions that do not match the final act, including mandatory officers for agencies with “fewer than 150” employees, semiannual assessments and, in later notes, a standalone vendor disclosure duty. The final law instead gives the 150-or-fewer officer option, drops that assessment frequency, narrows assessments and removes that separate vendor subsection. Use the notes for their stated forecasts and assumptions—not as a substitute for the enacted requirements. Latest indexed fiscal note; Final law.