This is the part where AI companies tell you they’re terrified of “bad regulation,” while quietly fighting the only kind of regulation that would actually make them slow down.
OpenAI and Google are opposing proposed AI safety rules in Massachusetts. Anthropic, meanwhile, is backing them. The rules, from what’s been shared publicly, involve the Massachusetts House pushing for strict independent reviews focused on catastrophic risk for advanced AI models. So we’re not talking about a slap on the wrist. We’re talking about someone outside the company getting real authority to ask, “Could this system cause serious harm, and can you prove it won’t?”
And the split here matters. Not because one company is “good” and the other is “bad,” but because it exposes how incentives work when the product is powerful and the downside is vague until it isn’t.
On paper, independent safety reviews sound like basic adulthood. If you build airplanes, bridges, or medicines, you don’t get to wave away oversight by saying you’re moving fast. You prove things. You document things. You accept that other people get to say “not yet.”
AI companies want the cultural status of building infrastructure, but they often argue like social apps: trust us, we’ll self-police, regulation will kill innovation, and anyway our intentions are pure.
I don’t buy it.
When OpenAI and Google push back on rules like this, I read it as a bet that speed is worth more than caution. Maybe they’ll say the rules are too broad, or would force them to reveal sensitive details, or would create delays that make them less competitive. Those could be real concerns. But it’s also true that “independent review” is exactly what you would resist if your edge comes from moving fast, shipping first, and dealing with consequences later.
Anthropic supporting the rules is interesting, and not automatically noble. It can be principle. It can also be strategy. If your brand is “we take safety seriously,” then backing strict reviews makes you look responsible and raises the bar for everyone else. If you’ve already built internal processes that look like what regulators want, then regulation hurts your competitors more than it hurts you. That’s not a conspiracy. That’s just how companies behave.
But even if Anthropic’s support is partly self-serving, the policy itself can still be the right move.
Because the real issue isn’t whether AI can write emails or code faster. It’s whether the people building the most powerful systems should be the ones deciding what “safe enough” means. History says no. Not because they’re evil. Because they’re busy, ambitious, under pressure, and paid to win.
Imagine you run a hospital system. Your team wants to use a new AI model to help with triage, insurance checks, patient messages, and maybe even suggestions for care. It’s tempting. Budgets are tight. Staff are burned out. If the AI makes operations smoother, it feels like relief. Now imagine a catastrophic-risk review says, “Not until you can show it won’t do X under stress.” That’s annoying in the short term. It might even cost money. But it’s also the difference between “we tested this carefully” and “we hoped the edge cases wouldn’t happen to our patients.”
Or imagine you’re a small startup trying to build on top of these models. You want stable rules. You want to know what will be allowed next year and what will get banned after a scandal. If Massachusetts forces serious reviews, that could create clearer norms. It could also create friction that only the richest companies can handle. That’s the tradeoff people should argue about honestly.
And yes, regulation can be clumsy. A state-level rule could become a patchwork. Companies might avoid launching certain tools in Massachusetts, or they might build a watered-down version just for compliance. Independent reviewers could become performative box-checkers if the process is designed badly. There’s also the risk of freezing a fast-moving field with requirements written for last year’s models, not next year’s.
But here’s the thing: the “do nothing” option isn’t neutral. It’s a decision to let a handful of companies set the pace for everyone else, and to find out what “catastrophic risk” looks like only after it shows up.
And catastrophic doesn’t have to mean sci-fi. It can mean systems that help real criminals at scale. It can mean brittle tools being trusted in high-stakes settings because the demo looked good. It can mean a race where every lab feels they have to ship first because they assume everyone else will. When you build a world where slowing down is punished, you don’t get careful behavior. You get rational shortcuts.
There’s also a bigger political consequence. If companies fight even modest-sounding safety rules now, they’re basically training the public to assume the worst about them later. That’s how you end up with regulation that’s harsher, dumber, and fueled by panic. If the industry wants smart rules, it has to stop treating any external check as an attack.
I’m not pretending Massachusetts will solve AI safety on its own. But I like the direction: independent review, catastrophic-risk focus, and real teeth. The burden should be on the builders to show restraint, not on the public to prove harm after the fact.
If OpenAI and Google think these rules are bad, they should say clearly what they would accept instead that still slows down risky deployments in a real way, not just in a press release.
So here’s the debate I actually want people to have: should we accept slower AI progress if it means independent reviewers can block releases that companies claim are safe, or is that too much power to hand to regulators?