Summary: On January 1, 2026, California SB 53, Texas HB 149 (TRAIGA), and Illinois HB 3773 took effect in the US. California requires covered frontier-model developers to publish safety frameworks, report critical safety incidents within 15 days, and protect whistleblowers; Texas restricts prohibited uses and assigns disclosure, human-review, and impact-assessment duties; and Illinois requires employee notice and bars discriminatory employment AI outcomes. The state obligations remain enforceable while challenges to the December 2025 federal preemption order are unresolved.
January 1, 2026 brought three US state AI laws into force on the same morning: California’s SB 53 (the Transparency in Frontier Artificial Intelligence Act), Texas’s HB 149 (the Responsible AI Governance Act, TRAIGA), and Illinois’s HB 3773. Between them they tile the whole debate — model safety at the frontier labs, deployment limits on high-risk applications, and algorithmic discrimination in employment. With the federal level still nowhere in sight, this is where American AI regulation genuinely touches ground, and it stands as the first compliance bill of 2026 to land on engineering and legal desks.
One piece of context belongs up front: the White House executive order of December 11, 2025 argues for preempting state AI laws that collide with federal policy — but an executive order cannot switch off a state statute. The consensus among legal commentators leaves nothing to interpretation: until the court challenges settle, state laws remain enforceable, and none of the compliance work is optional.
California SB 53: Transparency Duties for Frontier Labs
SB 53 stands as the first US statute to grapple with catastrophic risk from frontier AI, signed by Governor Newsom on September 29, 2025. It captures developers of large models trained with more than 10^26 FLOPS of compute, and its core duties number three: publish a safety framework spelling out how risk assessment and mitigation are embedded in development, report critical safety incidents to the state within 15 days, and stand up whistleblower protections. Penalties climb to $1 million per violation.
For laboratories of the Anthropic, OpenAI, or Google tier, this is not paper compliance. Safety frameworks become public documents; incident reporting runs on a hard clock; internal reporting channels have to function in reality, not on an org chart. California has also directed its Government Operations Agency to deliver a consortium report to the legislature by January 1, 2027 — leaving the door ajar for further mandates and audits.
Texas TRAIGA: Tiered Penalties, AG-Only Enforcement
Texas HB 149 was signed June 22, 2025 and reaches any entity developing or deploying AI inside the state. Its opening move is a list of “restricted purposes” — encouraging self-harm, unlawful discrimination, generating CSAM — all banned outright. Government agencies pick up separate disclosure and human-review duties when they put AI to work, and developers of high-risk systems owe impact assessments.
The enforcement architecture is conservative by design: penalties tiered at $10,000–$12,000 for curable violations and $80,000–$200,000 for incurable ones, enforcement reserved exclusively to the Texas Attorney General, and no private right of action. For companies this amounts to a fines-based regime rather than a class-action magnet — the risk is calculable, and calculable is not the same as negligible.
Illinois and What Comes Next on the Map
Illinois HB 3773 amends the state’s Human Rights Act and governs employers who run AI in employment decisions: employees must be notified, and systems may not deliver discriminatory outcomes against protected classes. Running opposite to Texas, this path preserves a private right of action — meaning vendors and adopters of hiring AI (resume screening, interview analytics, performance scoring) carry genuine litigation exposure.
Two more timers tick in the background. The Colorado AI Act — the first comprehensive US high-risk AI statute — switches on June 30, 2026, demanding impact assessments, consumer disclosures, and reasonable care against algorithmic discrimination. New York’s RAISE Act was signed December 19, 2025. The state map only thickens from here.
Federal Retreat, State Advance
The December executive order directed the Commerce Department to evaluate state AI laws, with a report due March 11, 2026, and singled out the Colorado AI Act as a preemption target. Genuine preemption, however, requires Congress or a courtroom, not a signature. Baker Botts’ January legal watch compresses the situation into four words: federal retreat, state advance — and a federal pullback should not be mistaken for regulatory relief.
That rhymes with the tone of our 2026 opening outlook: rules tighten through 2026, and the only open question is whether the pressure arrives via service terms or via state-law penalties.
What Engineering Teams Actually Need to Do
Rendered into engineering language, the to-do list gets concrete. Audit whether any model you train or fine-tune crosses the 10^26 FLOPS threshold, folding in the training provenance of third-party base models. Stand up a security and safety incident process that can finish a regulatory notification inside 15 days. Add disclosure flows for US users touched by AI in employment decisions. And write the impact-assessment documentation for high-risk deployments in healthcare, credit, employment, or government services. Where several states’ duties overlap, implementing once to the strictest standard costs far less than state-by-state customization.