The US mosaic: federal signals, state laws, and the agencies that already reach workplace AI

You know, there is this one specific sentence that is currently echoing through practically every American corporate boardroom right now. Oh, yeah. I hear it constantly.

Right. And it's practically hardened into an absolute fact at this point. The sentence is this, there is no federal AI law, so we are fine.

Right. And it is arguably, well, it's arguably the most expensive sentence a company can believe right now. It is a profound liability.

I mean, a massive liability. And what makes it so dangerous is that if you read it very narrowly, it's technically true. Right.

Technically. Like as of 2026, the United States has no single comprehensive horizontal statute that governs artificial intelligence. We do not have an equivalent to the European Union's AI Act.

Which is what everyone is looking for, right? The big binder. Exactly. Congress has passed nothing like it.

But believing that the absence of a federal AI law equals the absence of legal exposure is a complete fallacy. If you are building or deploying AI, you are operating in a highly regulated, high-stakes environment. Yeah.

It just, you know, it doesn't look the way most executives expect it to look. So if that sentence is a lie and there is no massive binder labeled the U.S. AI Act sitting on a compliance officer's desk, what invisible tripwires are these executives actually stepping on? I mean, if there isn't one big law, what are we dealing with? We're dealing with what professionals in the field call the U.S. mosaic. A mosaic.

Right, the mosaic. Yeah. Instead of one clean rule book, you have three distinct overlapping layers of authority.

Okay. Lay those out for us. Sure.

So first, you have federal signals. And this is a mix of highly volatile executive orders and durable general purpose sector statutes. Which we'll get into.

Right. Second, you have state laws. This is a chaotic, highly varied geographical web covering everything from biometrics to employment to consumer protection.

No two states are alike. Not even close. And third, you have enforcement agencies.

This isn't just one federal AI watchdog. It's a sprawling network of federal bodies, state attorneys general, city agencies, and perhaps most dangerously private plaintiffs. So our mission for this deep dive is to map this mosaic.

Because relying on the absence of an AI-specific law to protect you is like, well, it's like driving at 100 miles an hour and thinking you're completely safe from a ticket. Because there's no law on the books titled, Rules for Driving Blue Cars. That's a great way to put it.

Right. I mean, the speed limit still applies, regardless of what you call the car. The laws that reach these AI systems were written before anyone even used the term artificial intelligence.

That captures the dynamic perfectly. To map this mosaic, we really have to look closely at that first layer, the federal layer. And we need to understand why these old laws are effectively the apex predators of the AI landscape.

It helps to divide the federal layer into two distinct parts, the volatile signals and the durable statutes. Okay. Let's start with the volatile part.

So the volatile part consists of executive action and agency guidance. And the defining characteristic here is that it flips depending on who is sitting in the Oval Office. Right.

Because we saw a massive, I mean, a massive political whiplash on this recently. We really have. In October 2023, there was a sweeping directive on AI safety.

It was Executive Order 14110. And it mandated safety testing, red teaming. It heavily scrutinized frontier models.

It was incredibly comprehensive. But then, jump to January 2025, Executive Order 14179 was signed, which basically rescinded the previous order outright. Outright.

Gone. It replaced that safety-first approach with a highly deregulatory agenda. It focused almost entirely on compute infrastructure and workforce readiness instead of binding safety obligations.

So the entire federal executive posture on AI just flipped overnight. Overnight. And this teaches us a load-bearing lesson for corporate governance, which is that an executive order is not a statute.

Right. Governing your company's AI on the assumption that federal executive policy is stable is an exercise in futility. It's just a weather forecast.

Exactly. You treat the volatile federal layer like a weather forecast. You watch it.

You note the direction the wind is blowing. But you do not pour the concrete foundation of your compliance program based on it. Okay.

So that foundation has to be built on the durable statutes. These are the laws reaching AI without actually naming it. Yes.

These are the laws that executives tend to overlook because, well, they sound like ancient history. We're talking about Title VII of the Civil Rights Act of 1964. Wow.

Okay. The Americans with Disabilities Act, the ADA from 1990, the Federal Trade Commission Act from 1914. Wait.

1914. 1914. And the Fair Credit Reporting Act, the FCRA, from 1970.

Let me stop you there. Because 1914, we are regulating neural networks with a law passed before the invention of the zipper. I know.

It sounds crazy. It sounds absurd on its face. How does a law written during the Taft administration even begin to parse the latent space of a billion-parameter generative model? Well, it parses it by completely ignoring the technology.

Yet, these laws are ruthlessly outcome-focused. They do not care about vector mathematics or attention mechanisms. They ask fundamental questions.

Like what? Like, was a protected group harmed? Did a deceptive practice occur? Was a required legal process followed? The law fundamentally does not care whether the decision was made by a human hiring manager, a mechanical sorting machine, an Excel spreadsheet, or a state-of-the-art neural network. It's all the same to the law. Exactly.

If the outcome violates the statute, the tool used to achieve that outcome is liable. Which brings us to a massive trap that a lot of corporate compliance teams walked right into earlier this year. The guidance trap.

The guidance trap. Let's look at the Equal Employment Opportunity Commission, the EEOC. In January 2025, they quietly removed their AI-related technical assistance documents from their website.

Yeah, they just took them down. Right. These were the documents explaining how Title VII and the ADA apply to AI tools used in hiring and management.

And when they vanished, a lot of companies breathed a massive sigh of relief. They did. The naive red from many legal departments was, oh, the government deleted the AI hiring rules? The risk is gone.

We're safe. Right. We could deploy our AI resume screeners without worrying about regulatory blowback.

It's a catastrophic category error. Removed guidance is not removed law. I mean, if removing the guidance doesn't remove the law, doesn't that actually increase our risk? Absolutely.

It's silly, like removing the trail map at a national park. The cliff is still there. You're just navigating it blind now.

That's spot on. It increases your risk because the statutory obligations remain entirely intact. But you've lost the agency's interpretation of how to navigate them safely.

So what are we actually navigating blind? Specifically, we're talking about Title VII's prohibition on disparate impact. The EEOC doesn't need to publish a special AI webpage for disparate impact to be illegal. It has been illegal since the 1960s.

Let's define disparate impact for everyone because it is a crucial concept here and it operates very differently than what people traditionally think of as discrimination. Sure. Disparate impact is a form of unlawful discrimination where a facially neutral practice disproportionately disadvantages or harms a protected group without an adequate job-related justification.

Okay, so you don't have to program the AI to be racist or ageist? Not at all. There doesn't need to be discriminatory intent anywhere in the code or in the minds of the engineering team. Intent is completely irrelevant under disparate impact.

Wow. Let's say you build an automated AI screener to filter resumes for a software engineering role, right? You train the model on your past 10 years of successful hires. The model identifies that candidates who played lacrosse in college or who live within a certain zip code correlate highly with long-term retention.

So it starts heavily favoring those resumes. Right. But it turns out those seemingly neutral data points lacrosse and zip codes inadvertently serve as proxies for race or gender.

Yeah, resulting in the AI consistently filtering out female or minority candidates. If you cannot prove mathematically that playing lacrosse is a strict, necessary business requirement for coding, which you obviously can't, you have a disparate impact violation. Even if the AI didn't know it was discriminating.

Exactly. And to prove that this isn't just theoretical risk, we have to look at a real-world case regarding this. The case is MobileEV Workday.

Right. This case is a wake-up call for the entire tech vendor ecosystem. Unpack what happened here.

So Workday is a massive provider of enterprise software, and they offer automated applicant screening tools that employers use to sort through massive volumes of job applications. Standard applicant tracking system stuff. Right.

A plaintiff named Derek Mobley filed a collective action in federal court. He alleged that he applied for over 100 jobs at companies using Workday software and was rejected every single time. 100 jobs.

Over 100. He claimed that these screening tools systematically discriminated against applicants based on age, race, and disability. But here's the critical pivot and why we're talking about this.

He didn't just sue the dozens of companies he applied to. He sued Workday directly. And that is what makes this case a landmark moment in the U.S. mosaic.

Historically, software vendors rely on the shield of their commercial contracts. The sauce liability model. Exactly.

They say, hey, we just sell the software. We don't make the hiring decisions. Talk to the employer who bought the license and configured the tool.

But the court didn't buy them. No. The federal court looked at the actual mechanics of the AI screening tool and allowed the Title VII and Age Discrimination in Employment Act, the ADA, claims to proceed against Workday.

Wait, against the vendor. Against the vendor. Yeah.

The court essentially treated the software vendors screening AI as an effective agent making employment decisions on behalf of the employers. Hold on. You're saying the old laws reached right through the commercial contract, bypassed the employer entirely and grabbed the tech vendor directly.

That fundamentally breaks the B2B liability model. If I'm building B2B AI software, I assume my enterprise clients hold the liability for how they use it? Usually yes. But the court recognized that when an AI system is aggressively filtering candidates before a human being at the employer ever even sees the resume, the vendor is no longer just providing a digital filing cabinet.

They're acting as a gatekeeper. They're actively participating in the screening process. Title VII and the ADEA reached the AI vendor itself because the algorithm was doing the substantive work of an employment agency.

Wow. And that is exactly why these old general laws are far more dangerous than any new AI specific law could ever be. They're broad.

They're established. They carry decades of settled case law, and they are backed by plaintiffs and enforcers who know exactly how to leverage them. Okay.

So if Title VII creates massive civil liability in the hiring space, that is a huge financial and legal headache. You might have to pay a settlement, tweak your algorithm, apologize. But if we look at the Federal Trade Commission, the FTC, we're looking at something else entirely.

We are moving from a headache to an existential threat to the AI product itself. Absolutely. The Federal Trade Commission uses Section V of the FTC Act, which broadly prohibits unfair or deceptive acts or practices, as a blunt and incredibly powerful instrument against AI.

Unfair or deceptive. Yes. They pursue deceptive marketing, like a company slapping an AI-powered label on a product that's really just a basic rules interest.

We see that all the time. All the time. But they also aggressively pursue unfair practices, like deploying a tool that harms consumers without reasonable safeguards, or, crucially, training a model on data you didn't have the legal right to use.

Okay. Let me put on my cynical executive hat for a second. Go for it.

When a company hears that a federal agency might come after them for a data sourcing violation, they picture a line item fine. A cost of doing business. Exactly.

The agency issues a civil investigative demand, the company's lawyers spend two years negotiating a settlement, they write a five million dollar check, put out a PR statement about improving their processes, and the product just keeps running. It's just a parking ticket. And that mental model is going to get a lot of companies erased.

Graced. Erased. The FTC has developed a remedy that a monetary fine does not even approach in severity.

It's a concept called algorithmic disgorgement. Some in the industry refer to it as model destruction. Algorithmic disgorgement.

What does that actually mean mechanically? It means that if an enforcer determines you built your AI using ill-gotten, deceptively obtained or improperly sourced data, they don't just order you to delete the bad data from your servers. Okay. They order you to delete the actual AI models and algorithms that were trained on that data.

Wait, practically speaking, how does that even work? If a startup spends 20 million dollars and three years of engineering runway training a massive foundation model, and it turns out a fraction of their training data was scraped without proper consent, the FTC doesn't just issue a parking ticket. They repossess and crush the car. They absolutely repossess and crush the car.

Oh my God. Because of how neural networks function, you know, you can't easily unlearn specific data points. Yeah.

The weights and parameters of the entire model have been influenced by the illicit data. The toothpaste is out of the tube. Right.

The model itself is the fruit of the poisonous tree. Deleting a trained model changes the corporate math entirely. It is not a dent in your quarterly earnings.

If your model is the product, an algorithmic disgorgement order fundamentally destroys the core asset of the company. So how does this change the behavior of the engineering and legal teams on a daily basis? What should they be doing differently today? It completely reframes the concept of data provenance. Data provenance.

Yes. Historically, documenting exactly where your training data came from, what terms of service applied to it, what consent was gathered, that was seen as a tedious compliance chore. Just paperwork that slows the engineers down.

Exactly. But under the threat of algorithmic disgorgement, data provenance becomes an existential question. If your foundation is tainted by scraped data or data gathered under a privacy promise that the company later broke to its users, the FTC can pull the entire house down.

So it's not just paperwork. Your data provenance file is literally the deed to your AI model. If you cannot prove you have the right to the data, you do not own the asset you built with it.

We have to look at the Everal album case to prove the government actually does this because this sounds almost like science fiction. It's very real. This is from 2021 and it's the perfect illustration of this happening in the real world.

Walk us through Inree Everalbum. Okay, so Everalbum ran a consumer photo storage app called Ever. Simple enough.

Yeah, it was marketed as a safe place for users to upload their personal photos to the cloud. But behind the scenes, the company was using those millions of uploaded photos to develop and train facial recognition technology. And then they marketed this facial recognition tech to enterprise customers.

The core issue was that they misled users about when the facial recognition feature was turned on, and they explicitly broke promises about deleting photos after users deactivated their accounts. They kept the photos anyway. They kept the data to keep training the models.

Classic deceptive practice under Section 5. Exactly. So the FTC comes knocking. What did the settlement actually look like? The final order, which was issued in May 2021, required Everalbum to delete the improperly used photos, which is a standard regulatory remedy.

But crucially, it also ordered them to destroy the models and algorithms they had built from those photos. So they actually pulled the trigger. They did.

This was the FTC's first enforcement action specifically focused on facial recognition, and it firmly established algorithmic disgorgement as a real-world weapon. The company didn't just pay a fine. They had to certify to the federal government that they had vaporized the proprietary models they spent years building.

A 1914 consumer protection law produced an order that forced the deletion of a cutting-edge AI model. That is the mosaic at work. But it doesn't just end with the deletion, right? I mean, these FTC settlements aren't just a one-time execution.

They haunt the company for decades. The pain lingers for a generation, literally. When the FTC settles a case like this, the company signs a consent order.

That order binds the company for up to 20 years. 20 years? 20 years. It comes with mandated, highly intrusive independent audits, intense reporting requirements, and explicit legally binding bans on certain business practices.

And that consent order becomes enforceable in its own right. Meaning, if you step out of line five years later, the FTC doesn't have to take you to court to prove you did something deceptive all over again. They just prove you violated the consent order.

Precisely. Violating a consent order triggers massive civil penalties instantly, without the FTC having to re-litigate the original bad conduct. So a single settlement converts your past exposure into a long-running, highly restrictive compliance obligation with actual teeth.

It's a 20-year ankle monitor. Yes. When you're mapping your exposure to the federal layer, the remedy isn't just the risk of model deletion, it's the 20-year ankle monitor that follows.

So the FTC can force you to delete your model if your data provenance is bad. But what if your data is perfectly sourced? You bought all the licenses, everyone consented, everything is clean, but the math itself is a black box. Does the federal government care if the AI can't explain its own decisions? Because that brings us to consumer credit, and this is where the federal layer actually reaches inside the math and dictates how the AI is built.

Yes. This is a corner of the mosaic that reaches a massive class of AI systems. Anything that scores a consumer for credit, a loan, an apartment, insurance, or employment.

High stakes decisions. The highest. And the two apex predators in this space are the Fair Credit Reporting Act, the FCRA from 1970, and the Equal Credit Opportunity Act, ECOA, from 1974.

Specifically, it's implementing RULE, Regulation B. Let's ground these acronyms in real-world stakes. What do these laws actually demand from a company? Okay, so the FCRA governs consumer reports and the agencies that assemble them. If your AI scores people using third-party data and sells that score for credit or housing decisions, you might accidentally classify your tech vendor as a consumer reporting agency.

Like an Equifax. Exactly. And that brings a mountain of accuracy, disclosure, and dispute duties.

You're suddenly regulated like Equifax or Experian. That sounds like a compliance nightmare. It is.

But ECOA and Regulation B are where the specific, highly technical AI engineering constraints live. ECOA requires that when a lender takes an adverse action, meaning they deny an applicant for credit or offer them significantly worse terms, the lender must provide the applicant with a statement of the specific principal reasons for that denial. In writing.

And this creates a massive collision with modern AI, because modern neural networks, transformer models, deep learning arrays, they are famously opaque. They're black boxes. Right.

A bank might use a highly accurate AI to deny a loan, but the bankers themselves, and even the data scientists who built it, might not know exactly why the AI made that specific choice for that specific user. The latent space is just too complex. It's a product of a billion interacting parameters.

And the Consumer Financial Protection Bureau, the CFPB, which enforces these laws, saw this collision happening in real time. They issued two critical circulars to clarify exactly how a 1974 regulation applies to a 2026 neural network. Let's get into those circulars.

Let's look at CFPB Circular 2022-03. This established a hard, unyielding line. The black box nature of an AI model is absolutely no defense against the adverse action notice requirements of Regulation B. So you can't just tell the regulator, sorry, the algorithm is highly accurate, it reduces default rates by 20%, but it's just too complex to explain.

You cannot. The CFPB stated plainly that if your model is too complex or opaque to accurately produce the specific principal reasons why someone was denied credit, that it is illegal to use that model for that decision, period. Period.

Uninterpretability is not a valid defense. It's a fundamental compliance failure. The AI must be explainable.

The black box defense is dead in finance. It adds like a chef refusing to list the allergens on a menu because the recipe is a secret. I love it.

The law says, if you can't tell me what's in it, you can't serve it to the public. But you know, I know how compliance teams work. Oh, I'm sure you do.

If you tell them they have to provide a reason, they will just build a system that spits out a plausible sounding reason to make the regulator happy. And that is exactly the work around the industry attempted, which forced the CFPB to tighten the screw even further the very next year with CFPB Circular 20-2303. They saw the loophole.

They did. Since companies had to provide a reason for the denial, they were looking at the sample checklist of reasons provided in the back of the Regulation B forms. Things like insufficient income or limited credit experience.

Exactly. They would build a secondary AI layer that would just spit out the closest plausible reason from that generic list, even if it wasn't the exact mathematical trigger inside the black box. So the model denies the loan for some complex, obscure correlation, but the letter sent to the consumer just checks a box for insufficient income because it sounds reasonable.

Exactly. And Circular 20-2303 banned that practice entirely. It established the strict requirement of per-decision faithfulness.

Per-decision faithfulness? Yes. You cannot rely on a checklist of sample reasons. If those reasons do not perfectly and accurately reflect the actual individual mathematical decision the AI made on that specific applicant.

This forces incredible engineering constraints. We're talking about a rare instance where a law from 1974 reaches its hand inside a 2026 neural network and actively dictates model selection to the data science team. It absolutely dictates model selection.

And the tension between the legal requirement and the vector mathematics is profound. Think about the engineering reality here. Walk us through it.

Circular 2022-03 says the model must be able to produce reason codes. Circular 20-2303 says the reason codes must be perfectly faithful to the individual file. The kind of failure this catches is when a deep learning model actually denies an applicant because of an obscure behavioral signal, say, a complex vector correlation between their late-night food delivery habits, their social media usage, and historical default risk.

Wait, can AI actually track late-night food delivery for credit risk? Alternative data is a huge field. But let's say it does, but the system prints out an adverse action notice that checks a tidy box saying insufficient income. Because insufficient income is a legally safe, easy-to-understand reason, whereas your late-night Uber Eats habits correlate with bankruptcy opens up a massive can of worms regarding proxy variables and fairness.

Precisely. But if the reason code provided to the consumer is not faithful to the actual computation on a per-decision basis, the entire system fails compliance. And here is the kicker that keeps general counsels awake at night.

Tell me. A reason code layer that is highly accurate on average across a massive portfolio of tens of thousands of loans can still be completely illegal if it is wrong on the specific individual file the applicant is holding in their hand. So if your data scientists come to you with a highly performant, incredibly accurate black that boosts revenue by 10%, but they can't mathematically guarantee per-decision faithfulness.

You have a choice to make. You have to leave that model on the shelf. The 1974 law overrides the 2026 math.

You leave it on the shelf, or you use it for something entirely divorced from adverse consumer decisions like marketing or internal analytics. But for credit, housing, and insurance, explainability and faithfulness are hard, non-negotiable legal requirements. Wow.

And look, the federal agencies are powerful, and their reach is at least somewhat uniform across the country. If the FTC or the CFPB issues a rule, it applies broadly. Right.

But that brings us to the next layer of the mosaic, which is where things get truly chaotic. We are moving into the state layer. Up until now, we've talked about federal laws.

But the vacuum left by Congress, the lack of that single U.S. AI Act, has been filled by the states. And the states are building a geographical trap that catches companies completely off guard. To navigate the state layer without stepping on a landmine, you have to establish the golden rule immediately, right? Yes.

The golden rule is, state law follows your people, not your headquarters. This is the headquarters fallacy. An executive sitting in a corporate park in a state with zero AI laws thinks they are immune.

My state hasn't passed anything, so we don't have to worry about this. It is the most common and fatal error in state-level compliance. If your company is headquartered in a state with no AI laws, but you use an automated tool to screen a job candidate who physically resides in New York City, New York City Local Law 144 applies to you.

It reaches across state lines. Completely. If your AI system interacts with a user residing in Texas, the Texas Responsible AI Governance Act, TRAGA, applies to you.

Your exposure map is drawn by the physical geography of the people your AI touches, not by the mailing address of your corporate office. Let's examine a few of these specific state-level regulatory structures to show just how wildly different they are from one another. Let's start with New York City Local Law 144.

What does a compliance flow actually look like under this law? Local Law 144 focuses heavily on employment and hiring. It requires an annual independent bias audit for any automated employment decision tool, or AEDT, used to screen candidates in the city. And we should be precise here.

What legally constitutes an AEDT? An AEDT is essentially any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that substantially assists or replaces discretionary decision-making for employment. That is a very broad definition. Very broad.

And if you use one of these tools, you cannot just have your internal data science team run a quick fairness check. You must pay an independent, third-party auditor annually. You must publicly post the results of that audit, specifically the selection rates and impact ratios across race, ethnicity, and sex on your website.

Publicly. Yes. And you must provide explicit notice to candidates at least 10 business days before using the tool on them.

10 days' notice, plus an annual public audit enforced by the Department of Consumer and Worker Protection, the DCWP. That is a highly prescriptive, process-oriented law. It mandates very specific operational hoops.

Yes. But then you look at a state like Texas, and Texas took a completely different path. Texas enacted TRAGA, which goes into effect in January 2026.

Rather than focusing on mandatory audits and public reporting, TRAGA is intent-focused and narrower in its scope. How so? It strictly prohibits specific harmful uses of AI, things like behavioral manipulation, unlawful algorithmic discrimination, or generating unlawful deepfakes. But what makes TRAGA fascinating for an operator, and why corporate counsel needs to understand it deeply, are its unique mechanical provisions.

It gives companies very specific doors to walk through to mitigate risk. Tell us about the 60-day cure period, because this completely changes how a legal team reacts to an enforcement notice. Okay, so TRAGA is enforced exclusively by the Texas Attorney General.

There is no private right of action for individuals to sue. That's a relief for some. Right.

But before the Attorney General can bring a lawsuit against your company, the statute requires them to send a written notice identifying the specific provisions you are allegedly violating. From the moment you receive that notice, you have exactly 60 days to cure the violation. Okay, if I'm a general counsel, and I receive a notice from the Texas AG that my AI is violating TRAGA, my instinct is to fight it.

Naturally. My instinct is to draft a massive legal brief explaining why the AG misunderstands my technology and spend the next two months arguing. And if you do that, you are wasting the only remedy the statute gave you.

It is a 60-day compliance sprint, not a briefing period. Oh, I see. What you're supposed to do is preserve all the evidence, fix the specific algorithmic or operational provision named in the notice, update your internal policies to prevent it and send a written, sworn statement back to the AG with supporting documentation confirming the cure.

And if you do all that? If you execute all of that inside the 60-day window, the Attorney General is legally barred from bringing the action. That is incredibly practical advice. Don't argue.

Fix the code in 59 days. Texas also has something called a regulatory sandbox, right? Yes. The Department of Information Resources, the DIR, administers a 36-month regulatory sandbox.

This is a live deployment option for uncertain or highly novel technology. Like a safe harbor. Exactly.

An approved participant can test an AI system in the real world for up to 36 months without having to obtain the state licenses or authorizations that might otherwise apply. You have to file quarterly reports covering performance and risk mitigation, and the core prohibitions of the law still apply. You can't use the sandbox to unlawfully discriminate, but it provides a supervised path to market.

It's a pragmatic alternative to the binary choice of either shelving a risky product or shipping it into total legal uncertainty. Exactly. So, New York City wants annual public audits.

Texas wants to stop behavioral manipulation but gives you a 60-day cure period and a testing sandbox. It's a massive, confusing patchwork. It's a mess.

But the real nightmare of the state layer isn't just that it's a patchwork. It's that the patches are constantly being ripped up and restitched. Let's talk about state volatility and specifically the Colorado reversal.

The Colorado reversal is the ultimate cautionary tale for compliance teams trying to build durable infrastructure. Set the stage for us. In 2024, Colorado signed SB24205, the Colorado AI Act.

For two years, this was the absolute reference point for a comprehensive U.S. state AI law. It was modeled loosely on the European Union's risk-tier approach, focusing heavily on imposing rigorous duties of care on deployers of high-risk AI and algorithmic discrimination. And companies spent real money on this.

Oh, massive amounts of time, money, and engineering bandwidth building their compliance architecture around this specific law. And then what happened? The effective date was pushed back. And then, in a stunning legislative move in May 2026, the Colorado legislature passed SB26-189.

This new bill abruptly repealed the original statute entirely and reenacted a fundamentally different law. They just threw it out. They tore down the whole high-risk architecture, the heavy duties of care, and replaced it with a much narrower automated decision-making technology ADMT disclosure framework that doesn't even take effect until 2027.

So if you spent all of 2025 building your back-end compliance systems to satisfy the 2024 Colorado law, you built a house for a ghost. The law vanished before it ever governed a single algorithm. You cannot build compliance to repealed ghosts.

This proves that the state layer is incredibly volatile. A U.S. exposure map is not a one-time deliverable you draft, present to the board, and put in a drawer. It is a living document that must be re-verified quarterly because the laws themselves mutate even when your product doesn't change a single line of code.

Okay, let's step back. Because if Colorado changes its mind entirely, and Texas has a cure period, and New York City wants independent audits, and California has its own frontier safety transparency laws, wouldn't a rational company just say, this is impossible, we cannot build software this way, let's just pause, wait for Congress to assert federal preemption, wipe all this out and give us one single federal law to follow? It is a highly illogical question. But operating on that assumption is a remarkably dangerous strategy.

Let's define federal preemption first. Under the Supremacy Clause of the U.S. Constitution, a valid federal law can displace or override conflicting state laws. If Congress passed a comprehensive federal AI statute that expressly preempted state laws, the entire chaotic state mosaic could collapse into one legible rulebook overnight.

It sounds like a gene for the tech industry. But is that actually going to happen? Is Congress riding to the rescue? The historical evidence says no, not anytime soon. In July 2025, there was a very real, very constrained attempt at this.

A provision was inserted into a massive federal budget reconciliation bill that would have imposed a 10-year moratorium on state and local AI regulation. So a freeze. Yeah, it would have frozen thousands of pending and enacted state measures in their tracks, effectively clearing the board.

A 10-year freeze on the states. What did the Senate actually do with that provision? The United States Senate held a roll call vote on July 1st, 2025, specifically on whether to keep that preemption provision in the bill. And they voted 99 to 1 to strip that moratorium out.

Wow. The preemption attempt died instantly on the floor. 99 to 1. That is about as bipartisan as the U.S. Senate gets on literally anything.

It signals a massive, entrenched bipartisan reluctance at the federal level to strip the states of their authority to govern artificial intelligence. So waiting for Congress to save you from state laws means running massive, unmanaged legal exposure today based on a prayer that 99 senators will suddenly change their minds. It's not a strategy.

Not at all. Preemption remains a volatility risk hanging over the state layer. A future Congress could pass it.

But as of right now, in the real world, the states govern. You have to map the states. And if you think the compliance headaches of a New York City audit or a Texas cure period are bad, wait until you meet the apex predator of the entire U.S. mosaic.

Yes. We are going to Illinois, and we are going to talk about a law that terrifies every competent AI engineer and lawyer in the country. Let's dive into biometric privacy.

We are going to Illinois to look at a law passed in 2008, the Illinois Biometric Information Privacy Act, universally known in the industry as BiPA. Define BiPA for us. Why is an 18-year-old privacy law passed before Siri or Alexa even existed? The single sharpest weapon wielded against modern AI.

BiPA makes facial geometry, voice prints, and fingerprints the absolute highest risk data you can touch in the United States. It requires a private entity to obtain informed written consent before collecting a person's biometric identifiers. Written consent, yes.

And it requires a public retention and destruction policy. But what makes it the sharpest weapon isn't just the requirement to get consent. It's the enforcement mechanism.

BiPA has two teeth that make it uniquely devastating to AI companies. Let's look at tooth number one, the private right of action. Yes.

Most consumer protection laws are enforced by a government regulator, like the FTC or the state attorney general. You generally get a warning letter or a civil investigative demand first. You have time to negotiate, to cure, to prepare a defense.

Like we saw in Texas. Exactly. BiPA contains a private right of action, which means individual citizens can sue a company directly.

No regulator is needed. If the regulator is effectively an incredibly aggressive, highly motivated plaintiff's class action bar in Illinois. So no warning letters.

They do not send warning letters. They just file a lawsuit in federal or state court. And tooth number two is where the math gets genuinely terrifying.

The damages. Crushing statutory damages. BiPA allows a plaintiff to recover $1,000 per negligent violation and $5,000 per intentional or reckless violation.

Okay, that sounds bad, but... And the Illinois Supreme Court ruled in a landmark case called Rosenbach that a plaintiff does not even need to prove they were actually harmed or suffered any concrete real world damage. So you don't even have to be hurt by the data collection. No.

The bare procedural violation of not getting the proper written consent is enough to trigger the penalty. Let's do the math on that because I think people miss the scale of this. Please.

If you launch a fun new generative AI feature and 100,000 users in Illinois try it out over a weekend and you didn't get BiPA compliant consent, 100,000 users times $5,000 per violation, that is a $500 million existential threat. And it arrives overnight as a filed class action with zero warning from a government agency. A $500 million liability created over a single weekend.

That's insane. It is a corporate death sentence. And companies constantly walk right into this trap because they fundamentally misunderstand BiPA's fine print, specifically the exclusions.

Ah, the photograph exclusion. Right. The engineering teams read the statute, they see a specific line, and they think they are safe.

Let's deconstruct the photograph exclusion trap. Right. Because an engineer reads the text of IEK, sees a line that expressly says biometric identifiers do not include photographs, and they celebrate.

They say, great. Our AI app just analyzes the selfies that users upload. It's just photographs.

We are outside BiPA's jurisdiction. And that reading is fatally wrong and hit has cost companies hundreds of millions of dollars in settlements. Why is it wrong? Because BiPA's exclusion covers the photograph purely as an image.

It covers the storage of a JPEG. It covers a grid of RGB pixels sitting on a server. It does not cover a scan of facial geometry that a software system derives or extracts from that photograph.

Ah, the extraction. Yes. Illinois courts have consistently held that when an AI system processes an uploaded image to map the distances between the pupils, the contour of the jawbone, the shape of the nose extracting a geometric template to train a model or apply a filter, that resulting template is a biometric identifier.

You know, thinking the photograph exclusion protects your face scanning AI is exactly like thinking that because it's perfectly legal to own a camera, it's therefore legal to hide in the bushes and secretly record your neighbors. The picture itself isn't the crime. What you are extracting and doing with it is what breaks the law.

That is a phenomenal analogy. The exclusion protects the passive storage of pictures. It provides zero protection for the active extraction of face prints from them.

And almost every interesting generative AI feature, every filter, every avatar generator, every deepfake tool does the second thing. Let's look at the real world case that anchors this entire concept and proves how fast this can take down a company. Flora v. Prisma Labs from 2023.

What exactly happened here? Prisma Labs created the Lensa app, which went incredibly viral in late 2022. They launched a feature called magic avatars. Users would upload eight to 20 selfies and the app's AI would generate these beautiful stylized portraits in different artistic styles.

I remember this vividly. My entire social media feed was just these AI avatars for a solid month. It was an absolute phenomenon.

It was everywhere. But to generate those avatars, the neural network had to do more than just look at a JPEG. It had to detect and encode facial landmarks.

It had to map the faces to understand the user's features so it could reconstruct them in a stylized way. And Prisma Labs didn't get by PA consent. No.

In early 2023, Illinois residents filed a massive class action against Prisma Labs. The plaintiffs alleged that the company collected scans of their facial geometry to generate the avatars and to train its underlying neural networks, and that they did so without the explicit, informed, written consent and disclosures that Illinois PAA requires. And again, no federal AI law was needed to trigger a massive legal threat to a cutting edge generative AI tool.

There was no U.S. AI Act involved, an old privacy law, and an aggressive plaintiff's bar were more than enough to threaten the core of the business. The mosaic strikes again. And it strikes instantly the moment the feature touches the biometric data of an Illinois resident.

Which brings us to the ultimate question for our listeners, right? If BiPA is the biggest, sharpest threat in the mosaic, how do we actually disarm it? And how do we map this chaotic landscape for our own companies so we don't become the next case study? This brings us to the remediation phase. We spent a lot of time outlining the terrifying reality of the U.S. mosaic. Now it's time to provide the concrete steps to survive it.

Let's fix the BiPA threat first. If I'm an engineering leader or a product manager, how do I actually build a legally compliant flow? The fix for the BiPA gap is a highly specific, buildable engineering and legal task. It requires two distinct documents or flows to be generated and implemented before you ever touch the data.

Okay, what's step one? Step one addresses BiPA section 15a. You must publish a public retention and destruction policy. This must establish firm guidelines for permanently destroying biometric identifiers with a hard statutory rule that destruction happens when the initial purpose for collecting that data is satisfied or within three years of the person's last interaction with the company, whichever comes first.

Okay, so that's the public policy. It sits on the website, easy enough. But what about the actual user experience in the app? Step two? Step two addresses section 15b.

You must implement an informed, written consent flow before any collection happens. And let me be perfectly clear, because this is where UX designers fight with compliance teams constantly. Oh, I'm sure.

UX hates friction. They do. But a generic terms of service checkbox that says, I agree to the privacy policy is legally worthless here.

It doesn't count. You can't bury it in paragraph 47 of the ELA. It does not count.

And courts have rejected it repeatedly. The consent flow must explicitly tell the person that a biometric identifier or biometric information is being collected or stored. It must state the specific purpose for the collection, and it must state the exact length of term for which it will be stored and used.

So it's a very specific granular pop-up. Very specific. You have to build this gate into the product architecture so that the neural network literally cannot ingest the selfie or the audio clip until that specific granular consent is captured.

And thanks to a 2024 amendment to BiPA, an electronic signature or digital consent can satisfy this written requirement, but it absolutely has to hit those specific points. Correct. Okay.

So that fixes BiPA. But what is the overarching action item for the executive listening right now? If I'm a listener walking into work on Monday, and my CEO wants to launch a new AI tool by Friday, what are the literal steps I need to take to stop us from getting sued into oblivion? Your mandate for Monday morning is to walk into your office and build a one-page U.S. AI exposure map for the single highest stakes AI system your organization currently runs or plans to deploy. Just one page? Keep it to one page so the board actually reads it.

You need to pull your AI system's inventory, pick the one touching the most sensitive data like faces, voices, employment data, or credit decisions, and build a map that answers three concrete questions. Let's role-play this. Question one.

What does it touch? Look closely at the actual engineering data pipeline. Does it extract facial geometry? Does it make or assist an employment decision? Does it score consumer credit? Because the data determines the law. Exactly.

The exact nature of the data and the decision determines which federal and state laws apply. A chatbot giving IT support is fundamentally different from a resume screener. Question two.

Where are the people? Map the physical geography of the users, the applicants, and the workers, not the location of your corporate headquarters. If your system touches an Illinois resident, BPA is on the map. If you screen a New York City applicant, local law 144 is on the map.

If you touch a user in Texas, TRE is on the map. And question three. Who enforces and how? For every law you identify in steps one and two, name the specific enforcer.

Is it the FTC looking for deceptive practices? The EEOC looking for disparate impact? The Texas Attorney General demanding a cure? Or is it private plaintiffs under BAPA? And the remedy. Yes, crucially, name the remedy. Is it a monetary fine? Is it algorithmic disgorgement? Is it statutory damages per user? This converts a vague, generalized sense of legal liability into a concrete, quantifiable threat model that a board of directors can actually understand and act on.

We started this deep dive with a single, highly dangerous sentence. There is no federal AI law, so we are fine. We've systematically dismantled that fallacy.

We've seen how a mosaic of durable federal statutes, volatile executive signals, and a chaotic web of state laws reaches directly into the vector mathematics and code of modern AI. We have. And we've seen that the enforcers of these laws do not care how novel, complex, or opaque your technology is.

They care about what it does, who it touches, and whether you can mathematically and legally prove you have the right to do it. So as you head into the office to build your exposure map on Monday, I want to leave you with one chilling question to mull over. We learned today that the FTC's remedy of algorithmic disgorgement can literally order a company to delete its model and all the data it was trained on.

And we learned that state laws can change your liability overnight, bypassing your commercial contracts entirely. So if data provenance is the literal deed to your AI model, how much of your company's billion-dollar valuation is currently built on a legal hallucination that could be deleted tomorrow morning by a consumer protection law passed in 1914? It is the question every AI operator has to answer. Get to mapping.

We'll see you on the next Deep Dive.