Twenty-nine states have AI laws on the books. The United States federal government has none.
Sit with that for a second, because it explains almost everything about why building and buying AI tools in America right now feels like assembling furniture with instructions from twenty-nine different manufacturers, none of whom agree on what a screw is.
I test AI agents for a living. I read the docs, I break the products, I write down what actually happens versus what the launch post promised. And the single most common thing I hear from teams shipping these tools is not “we can’t get the model to work.” It’s “we don’t know what we’re allowed to do with it.” Those are very different problems, and only one of them gets solved by a better base model.
Speed is the whole story
The framing I keep seeing across the policy commentary — Spencer Fane’s piece on models moving faster than the rules, ChinaTalk’s argument that AI is becoming a political crisis rather than merely a technical one — points at the same mismatch. Model capability moves on a release cadence measured in weeks. Legislation moves on a cadence measured in sessions. When those two clocks run at that different a speed, you don’t get regulation. You get archaeology.
State legislatures noticed the vacuum and filled it, which is what legislatures do. That’s not a criticism of the states. It’s a description of what happens when the level of government best positioned to set a single national standard doesn’t set one. The result is a patchwork, and patchworks have a specific cost profile: they’re cheap for the biggest players and expensive for everyone else.
Think about who can afford to run compliance across twenty-nine different rulebooks. It’s not the four-person agent startup you found on Product Hunt last week. It’s the company with a legal department bigger than that startup’s entire headcount. Fragmented rules don’t slow down big AI labs. They slow down competitors to big AI labs. If you care about a market where small tools can actually challenge incumbents, this is the part that should bother you.
What the geopolitics adds
Carnegie’s work on a “compute coalition” — the idea of organizing AI development across free-world countries — adds a layer that domestic debates usually skip. AI policy isn’t just consumer protection. Compute is infrastructure, and infrastructure decisions are strategic decisions. A country that can’t articulate a coherent domestic position on AI is going to have a hard time coordinating an international one.
I’m not going to pretend I know how that plays out. I review tools. But I’ll note the practical version of it: the companies I evaluate increasingly make architectural choices based on guesses about future rules. Where they host. What they log. Whether they build in an audit trail nobody has asked for yet. That’s engineering effort spent on speculation, and speculation is not a great foundation for a product roadmap.
The labor question nobody wants to answer
Carnegie also lays out three competing views on AI and the future of work, which is a polite way of saying the experts don’t agree. I find that clarifying rather than frustrating. When serious researchers hold three different positions, anyone selling you certainty about job displacement — in either direction — is selling something else.
From where I sit, testing agents that claim to replace roles, the honest answer is messier than any of the three views. Most agents I try are genuinely useful for a narrow slice of a job and genuinely useless for the rest of it. That produces neither mass unemployment nor business as usual. It produces a slow, uneven reshuffling that’s hard to legislate because it doesn’t announce itself.
What I’d tell a buyer
If you’re purchasing AI tools into this mess, a few things follow:
- Ask vendors which state rules they’re built against. Vague answers are answers.
- Favor tools with real logging and data controls, even if you don’t need them today. Rules tend to arrive retroactively.
- Treat “we’re compliant” as a claim to verify, not a feature to check off.
- Assume the rules will change mid-contract, and read the exit terms accordingly.
None of that is exciting. All of it is cheaper than finding out later.
The uncomfortable part is that the current arrangement isn’t really a policy choice. It’s the absence of one, and absence has consequences that compound. Twenty-nine state answers to a national question is not a system. It’s what a system looks like before someone builds it.
Meanwhile the models keep shipping. They don’t wait for anybody.
🕒 Published: