Jan 1, 2026
date TRAIGA (Texas Business & Commerce Code Chapters 551-553) took effect, enforced solely by the Texas Attorney General with no private right of action
Jan 1, 2027
new effective date for Colorado's AI law after the original SB 24-205 was delayed twice, blocked in federal court, and repealed and replaced by SB 26-189
$200,000
maximum per-violation penalty under TRAIGA for uncurable violations, on top of $2,000 to $40,000 per day for violations that continue after the cure period

SponsoredHorizon3.ai

Proactive Security for the AI Era

NodeZero continuously and autonomously pentests infrastructure, identity, cloud, and now web applications, chaining weaknesses across every domain the way real attackers do. Every finding ships with replayable proof showing exploitable business impact, not theoretical risk.

See NodeZero WebApp in action

A lot of AI governance advice published through 2026 repeats the same clean split: Texas requires proof of intent before an AI system creates legal exposure, while Colorado imposes a duty of care regardless of intent. That split was real when it was written, and it is still a useful way to understand the two basic theories of AI liability now spreading across state legislatures. But if a compliance program is being built today around that description of Colorado's law specifically, it is being built around a law that no longer exists in that form. Colorado's original AI Act, SB 24-205, was delayed twice, blocked by a federal court, and then repealed and replaced by the legislature before its own effective date ever arrived. What is actually scheduled to take effect in Colorado on January 1, 2027 is a narrower law with different obligations than the one most 2025 and early-2026 compliance guides still describe. This piece works through what Texas TRAIGA and Colorado's current AI law actually require, why the underlying legal theories still genuinely conflict even after Colorado's rewrite, and what a security or compliance team should build so one program covers both without betting on either state's law staying exactly as it reads right now.

What Texas TRAIGA actually requires

The Texas Responsible Artificial Intelligence Governance Act took effect January 1, 2026, codified at Texas Business & Commerce Code Chapters 551 through 553. It applies broadly to any entity that promotes or advertises in Texas, produces a product or service used by Texas residents, or develops or deploys an AI system in the state, and its definition of AI covers recommendation algorithms and autonomous systems, not just generative AI. Government agencies must disclose AI use to a person before or during an interaction, and healthcare providers must disclose AI use connected to treatment. On discrimination specifically, TRAIGA Section 552.056(b) prohibits intentionally developing or deploying an AI system that unlawfully discriminates or infringes a constitutional or statutory right, and Section 552.056(c) is explicit that disparate impact alone does not establish that intent. TRAIGA also bans AI systems designed to manipulate someone into self-harm, harm to others, or criminal activity, and separately restricts government use of social scoring and biometric capture, with private companies exempted from that specific restriction. Enforcement runs exclusively through the Texas Attorney General, there is no private right of action, and the AG must issue a notice with a 60-day cure period before penalties attach. Penalties under Section 552.105(a) run $10,000 to $12,000 per curable violation not cured, $80,000 to $200,000 per uncurable violation, and $2,000 to $40,000 per day for violations that continue. A meaningful safe harbor exists too: Section 552.105(e)(2) protects a company that discovers a violation through its own testing, red-teaming, or internal review, and that substantially complies with a recognized framework such as the NIST AI Risk Management Framework's Generative AI Profile. Texas also runs a regulatory sandbox allowing up to 36 months of testing outside standard licensing requirements, subject to quarterly reporting.

What Colorado's AI law actually requires now, and why that answer changed in 2026

Colorado passed the first comprehensive US state AI law, SB 24-205, in 2024, and most compliance content written since then, including much of what is still circulating online, describes that version: a duty of reasonable care to protect consumers from algorithmic discrimination, developer and deployer obligations to document intended use and known risks, mandatory periodic impact assessments for high-risk systems, and a rebuttable presumption of compliance for companies aligned with a recognized framework like NIST AI RMF or ISO 42001. That law never took effect. Its start date was first pushed from February 1, 2026 to June 30, 2026 through SB 25B-004, signed during an August 2025 special session after lawmakers could not agree on amendments. Before the later date arrived, a federal magistrate judge issued an order in April 2026 blocking Colorado from enforcing SB 24-205 after AI company xAI filed a constitutional challenge and the US Department of Justice intervened in support, and the parties agreed the AG would not enforce or investigate the law until well after a ruling on xAI's request for a preliminary injunction. Rather than fight that fully out, Colorado's legislature repealed SB 24-205 outright and replaced it with SB 26-189, signed May 14, 2026, now scheduled to take effect January 1, 2027. SB 26-189 is a materially narrower law: it drops the duty-of-care standard and mandatory impact assessments and replaces them with a notice-and-disclosure regime. Developers must give deployers documentation of a system's intended uses, training data categories, and limitations, and notify them of material updates. Deployers must notify consumers when AI is used in a consequential decision covering education, employment, lending, housing, healthcare, or insurance, and must provide a plain-language explanation within 30 days of an adverse outcome, plus a human review and correction process. Liability is allocated between developer and deployer by fault rather than joint and several, and the Attorney General enforces with a 60-day cure period that sunsets after three years. Even this replacement law is expected to face its own legal challenges from the same parties that blocked its predecessor, so treat January 1, 2027 as the current best information rather than a certainty.

Free daily briefing

Briefings like this, every morning before 9am.

Threat intel, active CVEs, and campaign alerts, distilled for practitioners. 50,000+ subscribers. No noise.

Intent versus process: the distinction that survives the rewrite

Even after Colorado's overhaul, the two states still rest on genuinely different legal logic, just not quite the intent-versus-impact split that circulated when SB 24-205 was still expected to take effect as written. TRAIGA conditions liability on subjective intent: a company that can show it did not mean to build or deploy a discriminatory system has a real defense, and disparate impact statistics alone cannot substitute for proof of that intent. Colorado's current law, SB 26-189, does not ask about intent at all, in either direction. Its notice and disclosure duties are triggered by objective facts, whether a system falls into a listed consequential-decision category, whether an adverse outcome occurred, not by anyone's state of mind. A company that has built a complete TRAIGA defense file, red-team logs, a NIST AI RMF alignment memo, documented internal review, can still owe a Colorado consumer a notice or a 30-day explanation it has never built a workflow to produce, because that duty does not care whether the company intended harm. The reverse is also true. A company that has built out Colorado's consumer notices and human-review workflow has not thereby proven anything about intent, and disclosure records without genuine underlying bias testing can even work against a TRAIGA defense by showing the company knew a system produced adverse outcomes and did not meaningfully investigate why. It is also worth remembering that the older, stricter duty-of-care model Colorado originally enacted has not disappeared as a legislative template just because Colorado itself walked it back under court pressure. Other states drafting AI legislation may still land closer to that original design, so understanding it remains relevant even though it is not currently in force in Colorado.

Where a program built for only one law leaves real gaps

A program built narrowly around TRAIGA's intent defense, red-team documentation, a framework-alignment memo, records of internal review, does a genuinely good job of proving the company did not mean to discriminate. It produces none of what Colorado's SB 26-189 actually requires: no consumer-facing notice when AI is used in a covered consequential decision, no 30-day plain-language explanation after an adverse outcome, no developer-to-deployer technical documentation hand-off, and no human review or correction channel a Colorado consumer can actually use. Those are freestanding disclosure obligations under Colorado's law, and having a strong no-intent case under TRAIGA does not satisfy any of them. Run the comparison the other direction and a program built purely around Colorado-style disclosure, consumer notices, plain-language adverse-outcome explanations, a human-review workflow, produces exactly the kind of paper trail that documents a system's outcomes without necessarily documenting why they happened or what was done about it. Without genuine bias testing and a record of remediation behind those disclosures, that paper trail can cut the wrong way under TRAIGA, since it can show a company was aware of adverse outcomes for a protected group without showing it investigated whether those outcomes were the product of intentional design choices. Neither program builds the evidentiary record the other regime's theory of liability actually needs, absence of intent on one side, completeness of disclosure and process on the other, so a security or compliance team has to build both categories of control deliberately, not inherit one by accident from whichever regulator's guidance it read first.

A reconciliation approach: controls that hold up under both

The practical path is a small set of controls that were never actually specific to one state's legal theory in the first place.

Documented pre-deployment bias and impact testing

Aligning this testing to the NIST AI Risk Management Framework, or another recognized framework, builds direct evidence for TRAIGA's Section 552.105(e)(2) safe harbor and also supplies the technical basis Colorado's disclosure obligations under SB 26-189 assume already exists somewhere in the pipeline.

An AI use-case inventory that flags consequential decisions

Tagging every system that touches employment, lending, housing, healthcare, education, or insurance identifies exactly which Colorado notice duties apply, and the same inventory doubles as the classification record TRAIGA implicitly expects a company to have before deployment.

A standing internal-review or red-team cadence with logged remediation

This is the specific mechanism TRAIGA's safe harbor rewards, and it is also the proof of genuine responsiveness that would matter if Colorado's adverse-outcome explanation duty is ever triggered for a real consumer complaint.

Developer-to-deployer technical documentation packages

SB 26-189 makes this an explicit developer duty in Colorado, and the same package gives a deployer the material it needs to run its own bias check before deployment, which closes the loop TRAIGA's intent standard cares about on the deployer side.

A consumer notice and human-review workflow for consequential decisions

This satisfies Colorado's core deployer obligation directly, and it also functions as documented human oversight, the kind of evidence that helps establish a company was not simply letting an automated system run unchecked if intent is ever disputed under TRAIGA.

Version-pinned recordkeeping tied to the guidance in force at assessment time

Given how quickly both laws have already changed once, each assessment record should note which version of which law and which framework guidance it was made under, so a past risk classification can be defended on its own terms rather than judged against rules that did not exist yet.

This will keep happening: build for the strictest common denominator

Two states rewrote, delayed, or had their flagship AI law blocked by a federal court within about eighteen months of signing it. That is a strong signal that state AI regulation is still being actively contested, not settled, and more states are working through their own AI bills right now with their own liability theories, some closer to TRAIGA's intent standard, some closer to the duty-of-care model Colorado itself walked away from under legal pressure, and some combination not yet drafted. The durable response is not to chase whichever state's law reads as final this month. It is to build the underlying controls, an inventory, documented impact testing, disclosure and notice capability, human review, that would satisfy the strictest plausible version of a state AI law, so that whichever theory a new state adopts, or whichever way an existing state's law gets rewritten next, the program already produces most of what is needed. Teams that are also managing EU AI Act obligations alongside this US state-law patchwork will recognize the same overlap: the EU AI Act compliance platform comparison covers tools built around Annex III high-risk classification and conformity documentation, and an AI-system inventory, a documented impact assessment, and a disclosure trail built once for that regulation is largely reusable for Texas and Colorado's obligations too. Building the control once and mapping it to every applicable regime is cheaper than rebuilding it every time a state legislature changes its mind.

The bottom line

Texas TRAIGA and Colorado's AI law are still built on genuinely different legal theories, TRAIGA requires proof of intentional discrimination while Colorado's current law imposes notice and disclosure duties regardless of intent, but Colorado's actual current law is not the duty-of-care, impact-based statute most 2025 and early-2026 compliance guidance still describes. That original version, SB 24-205, never took effect: it was delayed twice, blocked by a federal court, and repealed and replaced by SB 26-189, now scheduled for January 1, 2027 and still subject to further legal challenge. A compliance program built for only one of these laws, or built around a description of Colorado's law that predates its rewrite, will have real gaps against the other. The more resilient approach is a small set of controls, an AI use-case inventory, documented bias and impact testing, disclosure and notice capability, and human review, that were never actually specific to one state's legal theory, because more states are going to pass their own AI laws with their own theories, and any one of them can change again before it even takes effect. This is informational analysis of publicly available legislative and legal-industry sources, not legal advice, and state AI law is moving fast enough that a company should confirm current requirements with its own counsel before relying on any specific provision described here.

Frequently asked questions

What is the actual difference between Texas TRAIGA and Colorado's AI Act?

TRAIGA (Texas Business & Commerce Code Chapters 551-553, effective January 1, 2026) only creates liability when a company can be shown to have intentionally developed or deployed an AI system to discriminate; disparate impact alone is explicitly insufficient under Section 552.056(c). Colorado's AI regulation went the opposite direction in its original 2024 design, then changed dramatically before it ever took effect. The version currently scheduled to take effect January 1, 2027, SB 26-189, does not ask about intent either, but it also dropped the duty-of-care and mandatory impact-assessment framework its predecessor had. It instead imposes notice, disclosure, and human-review obligations that trigger automatically whenever AI is used in a listed category of consequential decision, regardless of whether anyone meant to cause harm.

Is Colorado's AI Act still an impact-based, duty-of-care law?

Not as currently scheduled to take effect. The original Colorado AI Act, SB 24-205, signed in 2024, did impose an affirmative duty of reasonable care to protect against algorithmic discrimination and required periodic impact assessments, which is the version most compliance guides published through mid-2026 still describe. That law never actually took effect. Its start date was pushed from February 2026 to June 2026, a federal court then blocked its enforcement in April 2026 after a constitutional challenge, and Colorado's legislature repealed and replaced it entirely with SB 26-189, which drops the duty-of-care standard and mandatory impact assessments in favor of a narrower notice-and-disclosure regime effective January 1, 2027.

What happened to Colorado's original AI law, SB 24-205?

SB 24-205 was signed into law in 2024 but never took effect. Colorado's legislature first pushed its start date from February 1, 2026 to June 30, 2026 through SB 25B-004, signed during an August 2025 special session. Before that later date arrived, a federal magistrate judge blocked Colorado from enforcing the law in April 2026 after AI company xAI filed a constitutional challenge and the U.S. Department of Justice intervened in support. Colorado's legislature then repealed SB 24-205 outright and replaced it with SB 26-189, a substantially narrower law focused on disclosure rather than a duty of care, now scheduled to take effect January 1, 2027.

Does TRAIGA's intent requirement mean disparate impact is legally irrelevant in Texas?

Not entirely, but it changes what evidence matters. TRAIGA's Section 552.056(c) states that disparate impact alone is not sufficient to prove intent to discriminate, so a company cannot be held liable under TRAIGA purely because an AI system produces statistically unequal outcomes across groups. But disparate impact data can still be part of the evidence in a broader case, and TRAIGA's own safe harbor under Section 552.105(e)(2) rewards companies that actually look for and fix disparate outcomes through red-teaming, internal review, or alignment with a recognized framework like the NIST AI Risk Management Framework. Ignoring impact data entirely removes the very evidence that would support a genuine no-intent defense.

What controls satisfy both TRAIGA and Colorado's current AI law?

Four categories of control do double duty. Documented pre-deployment bias testing aligned to the NIST AI Risk Management Framework builds the record TRAIGA's safe harbor rewards, and also supplies the technical basis Colorado's disclosure duties assume exists. An AI use-case inventory that flags which systems touch a consequential decision, such as employment, lending, housing, healthcare, or insurance, identifies exactly which Colorado notice obligations apply and doubles as the classification record TRAIGA implicitly expects. A standing internal-review or red-team cadence with logged findings and remediation is the specific TRAIGA safe-harbor mechanism, and it is the proof of genuine responsiveness Colorado's adverse-outcome explanation duty effectively demands. A consumer notice and human-review workflow satisfies Colorado's core deployer obligation directly and gives a company a documented oversight process to point to if intent is ever disputed under TRAIGA.

Will more states pass conflicting AI laws like this, and how should a security team prepare?

Almost certainly. Two states rewrote, delayed, or had their flagship AI law blocked by a court within about eighteen months of signing it, which is a strong signal that state AI regulation is still being actively contested rather than settled. Other states will keep introducing bills built on different liability theories, some closer to an intent standard, some closer to a duty-of-care or disclosure standard, and any given state's current law can change again before a compliance program built around it even goes live. The more durable approach is building controls, an AI use-case inventory, documented impact testing, disclosure and notice capability, and human review, that would satisfy the strictest plausible version of a state law, rather than optimizing narrowly for one state's law as it reads today.

Sources & references

  1. Norton Rose Fulbright, The Texas Responsible AI Governance Act: What your company needs to know
  2. ComplianceHub.Wiki, Texas's AI Law Took Effect January 1: TRAIGA
  3. Sonomos, The Colorado AI Act (SB 24-205): A Compliance Guide for 2026
  4. McDermott Will & Emery, Colorado AI Law in Flux: Comprehensive Replacement Bill Signed After Federal Court Blocks Predecessor's Enforcement
  5. Akin Gump, AI Law and Regulation Tracker: Colorado Postpones Implementation of Colorado AI Act, SB 24-205

Free resources

25
Free download

Critical CVE Reference Card 2025–2026

25 actively exploited vulnerabilities with CVSS scores, exploit status, and patch availability. Print it, pin it, share it with your SOC team.

No spam. Unsubscribe anytime.

Free download

Ransomware Incident Response Playbook

Step-by-step 24-hour IR checklist covering detection, containment, eradication, and recovery. Built for SOC teams, IR leads, and CISOs.

No spam. Unsubscribe anytime.

Free newsletter

Get threat intel before your inbox does.

50,000+ security professionals read Decryption Digest for early warnings on zero-days, ransomware, and nation-state campaigns. Free, daily, no spam.

Unsubscribe anytime. We never sell your data.

Eric Bang
Author

Founder & Cybersecurity Evangelist, Decryption Digest

Cybersecurity professional with expertise in threat intelligence, vulnerability research, and enterprise security. Covers zero-days, ransomware, and nation-state operations for 50,000+ security professionals every morning.

Giveaway: InfoSec World 2026 All Access Pass ($3,895 value)

Details →
Daily Briefing

Subscribe to enter the giveaway

Every subscriber is automatically entered. You also get daily threat intel every morning: zero-days, ransomware, and nation-state campaigns. Free. No spam.

Already subscribed? You're already entered.

Giveaway

Win a $3,895 InfoSec World 2026 pass.