In a recent post from “the biggest donor in the midterm elections,” the patchwork metaphor - a go-to attack line on state AI laws - finally frayed into threads.1
To appreciate how quickly the ground is shifting, consider the volume of state activity in 2026 alone. States have introduced more than 1,500 AI-related bills. In the first six months of 2026, 84 measures became law across 27 states, with several others still pending or expected to be reintroduced in 2027.
To see why this is either meaningless or indefensible, let me put it in different terms:
Imagine the Golden State Warriors went 78-4 on the season and then fans couldn’t shut up about how the NBA is biased against them and constantly preventing them from winning. (Maybe a better example is Patriots fans ca. 2001-2019)
If this feels like the “ground is shifting” you might want to get your inner ears checked for equilibrium issues.
Less than six percent is the passage rate within this “volume of state activity.”
84 is still a lot of new laws though my dude!2
Indeed. So let’s dispense with the drive-by about 1,500+ bills as if it says anything.
The 84 actual passed laws isn’t even the meaningful policy change. About one-third of those have no effect on AI companies or any other tangible outcome (ex: Alabama HJR 51—establish a commission to study AI and children; report due in November).3
So we’re left with 56. Please don’t dig into the specifics of the remaining 56 bills. I’m begging you.
Okay, fine (don’t say I didn’t warn you).4 Here's the first 27:
CO SB 26-189: In 2024, Colorado passed a first-in-the-nation comprehensive AI law. As I understand, it required impact assessments, notices, disclosures, all the things.5 It never took effect. This 2026 law is a repeal-and-replace eliminating most of the requirements. For those keeping score, this new bill LOWERED the compliance obligation from the status quo ante.
9 healthcare laws: Seven are prior auth and insurer-facing, such as requiring a physician to make the final decision re: a denial of coverage. The liability nexus is the health insurer and not the vendor selling AI-assisted claims review software. One law is specifically on downcoding. Two are patient consent and scope of practice.
13 extensions of existing criminal/tort laws: they take existing crimes, and update or expand accordingly, like modifying the definition of sexual extortion, CSAM, NCI, libel, and identity theft (via penalty increase). In other words, “we have felonies in this state - don’t use AI to commit one.”
4 pertaining to elections: disclosure in campaign calls and materials.
Before closing out with the rest of the bills, I’d like to survey the class. Raise your hand if there are any of the above bills that:
(a) you’d like to oppose on the record?
(b) you characterize as threatening innovation and suffocating startups?
(c) violate the “regulate uses and not development” maxim?
(d) are unnecessary because deepfake child porn doesn’t merit criminal penalties???
Go ahead get your hand up, don’t be shy.
Okay then. To wit:
None of this is an argument against protecting children, promoting transparency, or addressing the genuine risks that AI presents. States are responding to real concerns….But perpetually shifting, state-by-state development mandates will impose costs on Little Tech that will have corresponding costs to competition in AI markets.
…States would retain their essential and robust authority to police harmful uses of AI within their borders through their consumer protection, criminal, and civil rights laws
So we’re in agreement. Of the 1,500+ bill patchwork threat:
> 94% aren’t “perpetually shifting” anything because they aren’t law
< 6% ARE law
≈ 4% are the ones we should be concerned about
Heard.
That’s 29 remaining, conceding these bills do impose some compliance and market-limiting burdens on AI companies—
17 chatbot safety: real disclosure and safety requirements for developers. Non-trivial. A real place for debate and calibration.
12 misc.: everything from training data requirements, provenance, algorithmic pricing, employment restrictions, and the Illinois law which is a real frontier safety auditing regime. Again, real things, that real developers need to accommodate or work around.
For sake of argument let’s assume all 29 apply to all AI companies—that these are all broad-based and generally applicable (keeping in mind the CO law significantly eliminated a bunch of regulatory burdens).
Is this really “one of the biggest challenges facing AI startups?”
If so they could use some kind of state AI legislation tracker to follow it all.
Fortunately I made one and available for free (link below). Happy to share the JSON file, but I assume AI startup people are way better at this than me.
Mine’s pretty simple and incomplete, but it does the basic job. I’m not sure it solves what a16z says is the problem, because apparently AI startups need to:
"reconcile conflicting obligations….[for] competing and potentially contradictory requirements”
and simultaneously “always build to the most restrictive law.”
Nothing can help with these because they are categorically mutually exclusive. Having to build a ladder to get over the tallest wall does not also require building a series of shorter ones. That’s not how superlatives work.
Like, if my job required me to respond to the most obtuse policy argument made by advocates, I don’t need to spend time responding to less vacuous ones, much less “reconcile” them.*
*Contradictions by definition don’t reconcile. Indeed an incredibly burdensome task for any startup.
I considered a more troll-y title like “a16z Opposes Laws Against CSAM, Deepfakes, and Exploitation of Children,” but thought better of it. But…it’s not not true. Other titles included: “What’s the Compliance Cost for a Dead Bill?” and “Name One” (a challenge to identify the actual startup, law, or product feature you’re talking about).
Comparison: 213 abortion-related laws passed (2023), 176 state telehealth laws (2024), 109 trans rights related (2025), 67 gun laws (2018). I don’t know what the implication of any of this is, but apparently we’re just able to throw out a seemingly large number and some obvious conclusion falls out. These numbers include ALL laws enacted on these topics, not ones that go one direction or the other, which again the plaintiffs in this case don’t bother delineating either.
Probably they should’ve leaned more heavily on AI to analyze all those bills.
I’m including examples beyond the January-June 2026 window because they tell us to worry about “several others still pending or expected to be reintroduced in 2027.”
What Chamber of Progress argued would “stifle innovation in the entire AI ecosystem - both at the model level and the application level” and “[t]hreaten to…torpedo the state’s tech economy.”