The US mosaic: federal signals, state laws, and the agencies that already reach workplace AI
The short answer
"No federal AI law" is true and nearly irrelevant
The United States has no comprehensive federal AI statute as of 2026, but Title VII, the ADA, the FTC Act, the FCRA, and the Fair Housing Act all reach AI without naming it. The absence of a horizontal law redistributes exposure to older statutes; it does not remove it.
What you will be able to do
- Explain why the United States has no comprehensive federal AI statute as of 2026, and why "no AI law" does not mean "no law reaches AI."
- Distinguish the three layers of the US mosaic: federal signals (executive orders, agency guidance, sector statutes), state laws (biometric, employment, consumer, frontier-safety), and enforcement agencies that already claim jurisdiction.
- Map a specific AI system your organization runs onto the exact federal signals, state laws, and agencies that already reach it, with a citation for each.
- Analyze how existing general-purpose laws (biometric privacy, employment discrimination, consumer protection) are applied to AI without any AI-specific text, using the Illinois Biometric Information Privacy Act and the Lensa avatar case as the worked example.
- Identify which state you are actually exposed to, recognizing that state AI laws follow where your users and workers are, not where your headquarters is.
- Recognize the volatility of the federal layer: how an executive order rescinded, guidance quietly removed from a website, or a state law repealed and replaced changes your exposure without changing your product.
- Defend a US exposure map against the challenge that the underlying laws are "old" or "not about AI," and against the challenge that removed federal guidance means removed legal risk.
The ideas, one by one
The mosaic has three layers
Federal signals (volatile executive orders and agency guidance plus durable sector statutes), state laws (biometric, employment, consumer, frontier-safety, no two alike), and enforcement agencies (FTC, EEOC, state AGs, city agencies, and private plaintiffs). Map all three separately.
State law follows your people, not your headquarters
Illinois BIPA reaches you if you touch an Illinois resident's face. New York City Local Law 144 reaches you if you screen an NYC applicant. Draw your map by the geography of the users, applicants, and workers your AI touches.
Biometric privacy is the sharpest weapon
Illinois BIPA, with its private right of action and statutory damages of USD 1,000 to USD 5,000 per violation, is the most consequential AI-adjacent US law, and it never mentions AI. Facial geometry, voiceprints, and fingerprints are the highest-risk data an AI feature can touch domestically.
Old general laws are more dangerous than new AI-specific ones
The broad, well-established statutes (biometric privacy, employment discrimination, consumer protection) carry decades of case law and eager enforcers. The novelty of your AI does not shrink their reach. Never let anyone dismiss an exposure because the statute predates the technology.
Removed guidance is not removed law
When the EEOC removed its AI technical-assistance documents in January 2025, the underlying Title VII and ADA obligations stayed exactly the same. The disappearance of a roadmap is not the disappearance of the road; map to statutes, not to guidance.
The federal and state layers are both volatile
Executive orders flip with administrations (EO 14110 rescinded by EO 14179). State laws are repealed and replaced (Colorado's AI Act became a narrower ADMT law). Your exposure map is a living document that must be re-verified, especially the state layer, at least quarterly.
Match the system to its enforcer
A biometric feature means private plaintiffs under BIPA. A hiring tool means the EEOC, state civil-rights agencies, and NYC DCWP. A consumer AI claim means the FTC and state AGs. Knowing the enforcer tells you the shape of the threat and the evidence you need to survive it.
Read a state law for its doors, not only its prohibitions
TRAIGA bans certain AI uses, but the provisions you would actually operate are the 60-day cure period that follows written notice from the Texas Attorney General (fix, document, and confirm inside the window, because there is no second one) and the Department of Information Resources sandbox, a supervised test of up to 36 months with quarterly reporting. Every state law on your map deserves the same second read: what does it let me do, and on what clock.
The map is the artifact
The output of this topic is not knowledge that the US is complicated; it is a specific, sourced, one-page exposure map for your highest-stakes system, with the single highest exposure flagged. That map is the domestic half of your cross-border decision and an entry in your accountability dossier.
The worst federal remedy deletes the model, not the quarter
The FTC can order algorithmic disgorgement, the destruction of a model and the data it was trained on, as it did in the Everalbum matter. That reframes training-data provenance from a documentation chore into an existential question: if you had no right to the data, an enforcer can erase the asset you built on it. Weigh model-deletion risk, not only fines, and keep the provenance file that proves you owned what you built. (see Topic 2.6)
Preemption is the volatility hanging over the whole state layer
As of 2026 no federal law preempts state AI laws; a ten-year moratorium on state AI regulation was stripped from the federal budget bill by a 99-to-1 Senate vote in July 2025, so the states still govern. But the federal-state boundary is contested, and a single act of Congress could redraw it without your product changing. Map to the states as they stand, and flag the preemption debate as your map's top volatility note.
Listen to it
The same lesson, talked through by two hosts as episode 41 of the podcast.
Real cases
These examples show the mosaic reaching real AI systems, with the specific law and enforcer named in each case.
Example 1: Lensa and the Illinois biometric statute (Flora v. Prisma Labs, 2023). In late 2022 the Lensa app's "Magic Avatars" feature went viral: upload eight to twenty selfies, receive stylized AI-generated portraits. In early 2023, Illinois residents filed a class action, Flora v. Prisma Labs, No. 5:23-cv-00680 (N.D. Cal.), alleging that Prisma Labs collected scans of their facial geometry to generate the avatars and to train its neural networks, without the informed written consent and disclosures that Illinois BIPA requires (Loevy & Loevy; Artnet, 2023). No federal AI law was involved. An eight-year-old (at the time) state biometric statute, enforced by private plaintiffs, reached a cutting-edge generative-AI feature the moment it touched Illinois faces. This is the mosaic's sharpest tool in action, and it is your anchor for this topic.
Example 2: A drugstore chain banned from facial recognition (FTC). The FTC used Section 5 to ban a national retailer from using facial-recognition technology for five years after finding it deployed the technology without reasonable safeguards, falsely flagging shoppers. (see Topic 4.4) The point for the mosaic: a consumer-protection statute from 1914 produced a five-year ban on an AI system, with no AI law required.
Example 3: A hiring-AI vendor treated as the decision-maker (Mobley v. Workday). A federal court allowed an age-discrimination collective action to proceed against a hiring-software vendor, treating the vendor's screening AI as an agent making employment decisions covered by federal anti-discrimination law. (see Topic 5.4) The mosaic lesson: Title VII and the ADEA reached the AI vendor itself, not only the employers using it.
Example 4: New York City's bias-audit mandate in practice. Employers using automated tools to screen New York City applicants must commission an annual independent bias audit and post the results publicly. A company headquartered in another state that screens New York City candidates is covered. The DCWP enforces it. This is a concrete, standing obligation on any hiring AI that touches New York City, independent of any federal law.
Example 5: The Texas intent-focused approach (TRAIGA). Texas's Responsible AI Governance Act, effective January 2026, prohibits using AI to manipulate behavior, to discriminate unlawfully, or to produce unlawful deepfakes, and imposes disclosure duties on government agencies and in healthcare. A Texas user of your AI feature brings Texas's rulebook, which looks nothing like Illinois's biometric statute or New York City's audit mandate. One country, incompatible obligations.
The two TRAIGA provisions a Texas-exposed operator actually uses. Prohibitions tell you what not to build; these two provisions tell you what to do. First, enforcement: TRAIGA is enforced exclusively by the Texas Attorney General, and the statute states that it does not provide a basis for a private right of action (Texas HB 149, 89th Legislature, 2025, enrolled text). Before the Attorney General may bring an action, the office must send written notice identifying the specific provisions of the chapter it alleges have been or are being violated, and the recipient then has 60 days to cure. If, before those 60 days run out, the recipient both resolves the violation and sends back a written statement confirming the cure, with supporting documentation and confirmation that internal policies were updated to reasonably prevent a further violation, the Attorney General may not bring the action. That changes what you do the morning a notice arrives. You preserve the evidence of what the system actually did, you fix the specific provision the notice names, and you document the fix and the policy update inside the window. The cure period is a compliance sprint, not a briefing period; an organization that spends the 60 days deciding whether it agrees with the Attorney General has spent the only remedy the statute handed it.
The Texas sandbox as a live deployment option. Second, TRAIGA creates a regulatory sandbox program administered by the Texas Department of Information Resources. An approved participant may test an AI system for up to 36 months, extendable if the department finds good cause, without obtaining state licenses, registrations, or authorizations that would otherwise apply, and files quarterly reports to the department covering performance, risk mitigation, and consumer or stakeholder feedback. The statute's core prohibitions still apply inside the sandbox, so it is supervision, not permission to manipulate or discriminate. Treat it as a third option on the decision you would otherwise force into two: ship into uncertainty, or shelve the system. A supervised test with a defined term and a reporting duty lets you build the evidence that the system behaves before you carry full exposure, and the reports themselves become dated artifacts for your dossier.
Example 6: The Colorado reversal as a currency lesson. A company that built its 2025 compliance program around the Colorado AI Act's high-risk architecture found, by mid-2026, that Colorado had repealed and replaced it (SB 26-189) with a narrower disclosure law effective 2027. The lesson is not about Colorado specifically; it is that the state layer is volatile enough that an exposure map is a living document, not a one-time deliverable. Verify each state law's current status before you rely on it.
Example 7: The federal guidance that vanished but changed nothing legally. In January 2025 the EEOC removed its AI technical-assistance documents from its website. A company that had been treating those documents as the extent of its AI hiring obligations now had no federal roadmap, but Title VII, the ADA, and the UGESP validation requirements still applied to its hiring AI exactly as before. The disappearance of the guidance changed the company's information, not its liability.
Example 8: The FTC orders a company to delete its model (In re Everalbum). Everalbum, Inc. ran a photo-storage app called "Ever." It used the photos that users uploaded to develop facial-recognition technology, which it then marketed to enterprise customers, and it misled users about when facial recognition was on and about deleting photos after account deactivation. In January 2021 the FTC settled the matter, and the settlement, finalized in May 2021, required Everalbum to delete not only the improperly used photos but also the models and algorithms it had built from them (FTC, "California Company Settles FTC Allegations It Deceived Consumers about use of Facial Recognition in Photo Storage App," 11 January 2021; FTC final order, May 2021). It was the FTC's first enforcement action focused on facial recognition, and the order required the destruction of AI models trained on data collected improperly, the remedy now called algorithmic disgorgement. The company later renamed itself Paravision and confirmed it had deleted the models as ordered. The mosaic lesson is stark: a consumer-protection statute from 1914 (Section 5 of the FTC Act), with no mention of AI, produced an order that erased a company's AI models. This is the "delete the model" remedy from Section 3H in a real, documented case, and it is why the provenance of your training data is an existential question, not a paperwork one. (see Topic 2.6)
Where people go wrong
- "There is no federal AI law, so we are fine." The single most expensive belief in American AI governance. There is no comprehensive federal AI statute, but Title VII (1964), the ADA (1990), the FTC Act (1914), the FCRA (1970), and the Fair Housing Act all reach AI without mentioning it, and the agencies that enforce them do not need a new law. The absence of a horizontal AI act is a redistribution of exposure to older laws, not an absence of exposure.
- "Our headquarters is in a state with no AI laws, so state laws do not reach us." State laws like Illinois BIPA and New York City's Local Law 144 protect the people in those places, not the companies located there. If your AI touches an Illinois resident or a New York City applicant, that jurisdiction's law reaches you regardless of where your office sits. Exposure follows the affected people.
- "BIPA is an old privacy law, not an AI law, so it does not apply to our generative-AI feature." BIPA applies to any private entity that collects biometric identifiers such as facial geometry without consent. A generative-AI avatar feature that scans facial geometry from selfies is squarely within BIPA's terms, as the Lensa litigation showed. The age of the statute and the novelty of the technology are both irrelevant; the conduct is what matters.
- "The EEOC removed its AI guidance, so AI hiring risk is gone." The removed technical-assistance documents were non-binding explanations of how existing law applies. Removing them deleted the roadmap, not the road. Title VII's disparate-impact prohibition and the UGESP validation requirements still apply to AI-driven hiring. Removed guidance arguably increases uncertainty and therefore risk.
- "An executive order is like a law, so we can rely on the current one." Executive orders are not statutes. EO 14110 was rescinded by EO 14179 with a signature. Building a governance foundation on executive policy means rebuilding it every time the administration changes. Anchor to statutes and durable agency authority, not to the executive order of the month.
- "We commissioned a bias audit for New York City, so we are covered everywhere." New York City's Local Law 144 audit satisfies New York City. It does not satisfy Illinois HB 3773's outcome-based prohibition, Title VII's federal requirements, or any other jurisdiction. Each layer of the mosaic must be mapped separately; compliance with one is not compliance with all.
- "The Colorado AI Act is the US model, so we should build to it." As of 2026 the Colorado AI Act (SB 24-205) has been repealed and replaced by a narrower disclosure law (SB 26-189, effective 2027). Building to a repealed statute is worse than useless. The state layer is volatile; verify the current status of every state law before relying on it, and treat your exposure map as a living document.
- "We only need to worry about the laws that specifically say AI." The most dangerous exposures come from general-purpose laws that do not mention AI at all: biometric privacy, employment discrimination, consumer protection. The AI-specific state laws are newer, narrower, and less tested. Mapping only the AI-labeled laws misses the biggest threats.
- "The worst case is a fine we can absorb." The worst federal case is not a fine; it is algorithmic disgorgement, an FTC order to delete the model itself along with the data it was trained on, as happened in the Everalbum matter. Deleting a model can erase the product, not dent the quarter. The fix is to weigh the deletion risk, not just the monetary penalty, whenever the training data's provenance is questionable, and to keep the data provenance file that shows you had the right to build what you built. (see Topic 2.6)
- "We settled with the agency, so the matter is closed." An FTC settlement typically comes with a consent order that binds the company for years, often two decades, with deletion, audit, reporting, and practice-ban obligations that are enforceable on their own. Violating the order triggers penalties without the agency having to re-prove the original conduct. The fix is to map the multi-year consent order as a standing obligation, not to treat the settlement as an ending.
- "If the states are a mess, Congress will just preempt them and simplify everything." As of 2026 there is no federal preemption of state AI laws. A ten-year moratorium on state AI regulation was proposed in 2025 and stripped from the federal budget bill by a 99-to-1 Senate vote in July 2025, so the state layer stands in full. The fix is to map to the law as it actually stands (states apply), while flagging the preemption debate as the top volatility risk in your map's notes; do not pre-comply with a moratorium that does not exist, and do not assume the current federal-state split is permanent.
Questions people ask
- What is US mosaic?
- The patchwork of federal signals, state laws, and enforcement agencies that governs AI in the United States in the absence of a single comprehensive federal AI statute. Its three layers must each be mapped separately because none of them alone captures a system's full exposure.
- What is comprehensive (horizontal) AI statute?
- A single law that governs AI across sectors with defined obligations, such as the European Union's AI Act. The United States has no such statute as of 2026; its AI governance runs through general-purpose laws and agency authority instead.
- What is federal signal?
- A federal-level indication of AI policy that is not a comprehensive statute. Includes volatile executive orders and agency guidance and durable sector statutes. The volatile parts change with administrations; the durable parts (Title VII, the FTC Act, the FCRA) do not.
- What is Illinois Biometric Information Privacy Act (BIPA)?
- An Illinois statute (740 ILCS 14, enacted 2008) requiring informed written consent before a private entity collects biometric identifiers such as facial geometry, a published retention-and-destruction policy, and disclosure limits. Its private right of action and statutory damages of USD 1,000 to USD 5,000 per violation make it the sharpest AI-adjacent state law in the United States, though it never mentions AI.
- What is biometric identifier?
- Under statutes like BIPA, a physiological measurement used to identify a person, including a scan of facial geometry, a voiceprint, a fingerprint, and a retina or iris scan. AI features that derive these from images, audio, or video collect biometric identifiers and trigger biometric privacy law. More on Biometric identifier
Keep going
This lesson builds US federal and state AI regulation, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →