Pass the Popcorn: America Discovers It Can't Regulate AI

Pillar 05 · Algorithmic

I’ve been watching Washington try to figure out how to regulate AI. It’s better than anything on Netflix.

The latest episode comes from Jack Thorlin, a law professor at the University of Arkansas, who published a piece in Just Security called “The Supreme Court’s Open-Weight AI Regulation Gauntlet.” It’s careful, footnoted and deeply serious. It’s also, from up here, a comedy. The world’s most powerful government has worked out that it has no legal way to regulate the most important technology of the decade. It got there one Supreme Court ruling at a time, starting in 1995.

Here’s the setup.

Today the US regulates frontier AI by phone call. The president rings a few closed-model companies, and they listen, because they have federal contracts and shareholders and a strong preference for not being the next example. We all watched it in June. Anthropic suspended access to Fable and Mythos to comply with Commerce export controls, then switched them back on three weeks later when the controls lifted. Was it legal? Thorlin shrugs, and so did everyone else. It worked.

Now picture the day an open-weight model catches the frontier. Experts say that’s months away. The weights get published, and suddenly the list of people running a frontier model includes a few thousand startups, every university, and a guy in Moncton with a gaming rig and opinions. None of them have a DoD contract to lose. You can’t jawbone a torrent.

So Thorlin goes looking for Plan B. Every door he tries is locked.

Build an independent AI regulator, like the Fed? Trump v. Slaughter lets the president fire the commissioners whenever he likes. Congratulations, you’ve built a department with extra paperwork.

Have agencies regulate under existing law? The major questions doctrine says something this big needs Congress to say so clearly. Loper Bright killed Chevron deference, so no creative reading of a 1970s statute will save you.

Call it national security? Learning Resources v. Trump says the Court isn’t reading emergency powers generously these days. IEEPA even exempts informational materials. Lawyers will bill thousands of hours arguing whether weights count.

Get Congress to pass a law? This is a Congress that needs a hostage negotiator to keep the lights on. But fine, say it passes. Banning people from running a model at home hits Lopez and Morrison, because running software in your basement isn’t interstate commerce. Then nondelegation shows up and asks Congress to write precise, durable rules for a technology that reinvents itself every quarter.

Thorlin’s conclusion is that what survives is narrow. Regulate commercial deployment and hosting. Harden critical infrastructure. Even that does nothing about weights published in Beijing. And to his credit, he says plainly he doesn’t have a hopeful answer.

I do. It just isn’t a legal one.

The model was never the thing to protect

Model weights are a file. Files have one hobby, and it’s being copied. Every plan to govern AI by controlling the model is a plan to stop a file from being copied, and we’ve been losing that fight since Napster.

The thing you can actually control is what the model touches. Your data. Where it gets decrypted. Who can see it in the clear. And whether you can prove any of that to an auditor afterwards.

That’s an engineering problem, and engineering problems have the nice property of not needing five votes.

Where SkyeConnex fits

This is exactly the layer we built.

SkyeConnex encrypts data on the client, then erasure-codes it into shards placed by geopolicy across independent hosts. No single operator holds anything usable — not the cloud provider, not the host. On the sovereign deployment, where the customer is the operator and we are outside the trust boundary entirely, that includes us. A subpoena served on any one host gets confetti.

SkyeGXU runs the AI where the data lives. Shards only reassemble inside the customer’s boundary, on their own machines or inside a hardware enclave that proves what it is before it gets a single shard. The model loads in there, does its work, and writes the results back into custody. Plaintext never leaves. Neither do the weights.

Workloads are packaged and signed, so you know exactly which model, which container and which version ran against which data. That’s the provenance record every regulator is going to ask for the moment they find a lever that still works.

And here’s the punchline. The one lever Thorlin says survives is deployment and hosting. That’s where we sit. We don’t replace your cloud. You keep AWS, Azure, Google and Canadian hosts like ThinkOn, and we put custody on top. If a rule ever lands on hosting, the hosts running SkyeConnex already have the answer.

The fair objection is that a custody host looks like a blind spot. If the host can’t see what’s running, how does anyone police it? Easy. You don’t need to see the data to prove the rules were followed. You need signed workloads and an attestation trail. Trust the math, then check the receipts.

The Canadian bit

None of the cases Thorlin lists bind Canada. We could build durable technical controls tomorrow without waiting for a court or a Congress to clear the hallway. Instead our default AI strategy is “use the American model and hope the phone call goes well.” June was the fire drill. The next one might not come with a three-week restore.

Washington can’t regulate the weights. Ottawa doesn’t need to. Lock down the data and the weights can go wherever they like.

I’ll be over here with the popcorn. And the custody layer.

Sovereignty by architecture, not by promise.

Bias declaration. I’m the founder and CEO of SkyeConnex. I build the thing I just told you to buy. Read accordingly.

Source: Jack Thorlin, “The Supreme Court’s Open-Weight AI Regulation Gauntlet,” Just Security, September 24, 2026. justsecurity.org