Giving Congress subpoena power over AI sounds comforting until you remember what Congress is actually like: slow, political, and usually years behind the thing it’s trying to control. Still, I don’t think this is automatically a bad idea. I think it’s a late idea. And late ideas can either become guardrails—or become theater.
Based on what’s been shared publicly, Rep. Don Beyer is pushing to set up a House Select Committee on Artificial Intelligence. Not just a discussion club. The pitch is real authority: subpoena power, plus legislative authority. On top of that, he’s working on an AI safety draft bill with Rep. Jacobs, aiming to introduce it later this year, with a possible path to action in early 2026. He’s also talking up a bipartisan approach, even after Democrats did well in the midterms.
The part that grabs me isn’t “select committee.” Congress makes committees the way people buy gym memberships: a hopeful gesture, not always followed by real change. The part that matters is subpoenas. Subpoenas mean they’re thinking about leverage. It’s a signal that lawmakers want access to what companies know, what they’re building, and what they’re hiding behind glossy demos and vague promises.
And yes, “hiding” is the right word sometimes. If you’re a company building powerful systems, you’re incentivized to share the wins and bury the mess. That’s not evil; it’s predictable. The question is whether a committee can force the truth into daylight without turning the process into a partisan circus.
Because that’s the risk here: subpoenas can be used to learn, or they can be used to perform. If this turns into televised gotcha moments where the goal is to clip a CEO saying something dumb, then we’ll get noise, not safety. The public will learn to tune it out. The serious people inside companies will learn to lawyer up and say less. And the committee will “do something” while the actual problems keep moving.
But I can also picture the opposite, and it’s the first time in a while that Congress doing oversight doesn’t sound completely imaginary.
Imagine you’re running a school district. A vendor is selling an AI tool that flags students for “risk,” or helps decide who needs extra attention. It sounds helpful. It might even save time. But if the model is wrong, the kid pays the price, not the vendor. A serious committee with subpoena power could actually dig into basic questions: What data did you train on? How often is it wrong? Who audits it? Who is responsible when it harms people?
Or imagine you work at a hospital. A tool summarizes patient notes or suggests next steps. If it makes things up, that’s not a funny glitch. That’s someone’s life. Oversight can’t just be “trust us.” It has to be boring, specific, and enforced.
Now flip it. Imagine you’re a startup trying to build something useful with AI. You’re not trying to replace democracy or melt the planet. You just want to ship. If Congress writes sloppy rules, you don’t have a giant legal team to help you comply. The big companies will survive and even benefit, because complex rules can act like a moat. That’s the uncomfortable truth: regulation can “protect the public” and also freeze out smaller players at the same time.
That’s why a bipartisan approach sounds nice but isn’t automatically virtuous. “Bipartisan” can mean “durable and serious.” It can also mean “watered down until it doesn’t bite.” If the end product is a bill that mostly protects incumbents and creates a handful of committees and reports, we’ll call it progress and then act surprised when nothing changes.
Early 2026 as a target for legislative action is also a tell. That’s not soon. In AI time, that’s a different era. Companies will release new models, new products, and new ways to shove AI into everything people do at work—hiring, firing, grading, lending, policing, health decisions. The legal system will still be catching up to harms that already happened. A select committee could speed up learning, but it could also just document the disaster politely.
Still, I’d rather see an attempt at serious oversight than the current vibe, which is a mix of panic, awe, and wishful thinking. Subpoena power, used well, could cut through the fog. It could force clear answers about testing, failures, internal warnings, and how decisions get made before a product touches millions of people. It could also help draw a line between “AI safety” as a real public goal and “AI safety” as a branding strategy.
I’m not convinced Congress can do this cleanly, because the incentives are messy. Lawmakers want headlines. Companies want freedom. Voters want protection but also want cheaper, faster, better tools. And everyone wants someone else to be responsible when things go wrong.
So I’m left with one uncomfortable, practical question: if this select committee gets real power, will it use that power to set tough rules that slow down risky AI releases, even when that means taking heat from donors, companies, and voters who just want the benefits now?