Anthropic Wants States to Regulate AI Harder, Not Softer

Here is a sentence you do not read often: an AI company is asking the government to regulate it more, not less. That is roughly what Anthropic has spent the last few weeks doing, and it is worth paying attention to, because the usual script for a company this size is to lobby for the softest possible rules and call it a day.

Anthropic's pitch, laid out around the middle of July, is that the patchwork of state AI laws now taking shape should get tougher over time, not get flattened into one gentle national standard. The company wants frontier developers, the outfits building the most capable models, to face independent safety audits, real security requirements, and actual government authority to step in when something looks dangerous. "Transparency and self-reporting, we don't believe are sufficient anymore," Cesar Fernandez, Anthropic's head of U.S. state and local government relations, told reporters. Coming from a company that would have to live under those rules, that is a striking thing to say out loud.

What Anthropic actually wants

The specifics come from a policy framework Anthropic published in June. It asks for public summaries of how models are tested, regular independent reviews by outside evaluators, security programs that protect model weights and training infrastructure (the crown jewels, basically), and government power to block or deter a deployment that poses catastrophic risk. In plain terms: show your work, let strangers check it, lock the doors, and give regulators a real lever to pull.

The company also has a message for Congress. If lawmakers want to write one national rulebook, fine, but they should not use it to wipe out stronger state laws in the meantime. Anthropic's position is that federal law should "preserve state authority until it passes rules at least as strong" as what the states have already done. That is the opposite of the preemption most big tech companies quietly root for.

The states are already moving

This is not hypothetical. Illinois got there first: Governor JB Pritzker signed a law in early July requiring annual independent safety audits for frontier developers pulling in more than $500 million a year. California went ahead in 2025 with SB 53, which makes large developers publish safety plans and report incidents. New York's RAISE Act layered on independent audits, and Massachusetts is weighing rules that would bring in outside evaluators to judge whether a system could help someone build a bioweapon. Anthropic has backed language in several of these fights.

Put together, they form what critics and supporters alike call a ratchet: each new state can stack fresh obligations on top of the last one, so the floor keeps rising. Anthropic likes that design. It means the rules get stricter as the models get more capable, instead of freezing in place at whatever felt reasonable a year or two ago.

Where OpenAI parts ways

Not everyone in the industry sees it this way, and the split with OpenAI is the interesting part. OpenAI also supported the Illinois law, but it wants states to converge on a single common baseline that can harden into a national standard. "Step by step, they can create a de facto national standard," Chris Lehane, OpenAI's chief global affairs officer, put it. OpenAI's version leans on written safety standards, mandatory reporting of serious incidents, and independent audits, while pushing the thorniest technical and national-security questions up to federal agencies.

The difference sounds procedural. It is not. A common baseline is a ceiling dressed up as a floor: once everyone agrees on one set of rules, it gets much harder for any single state to ask for more. Anthropic's ratchet is built to keep climbing. OpenAI's baseline is built to hold steady and cut down on the headache of complying with fifty different rulebooks, a concern former Facebook security chief Alex Stamos has raised on developers' behalf. Both things can be true at once: divergent state laws are genuinely annoying to comply with, and a tidy national standard is exactly what you would want if your real goal were softer oversight.

Why any of this matters to you

You are not going to file a compliance report with Illinois. But the models these rules govern are the ones increasingly answering your emails, writing code at your job, and sitting behind the chatbots your kids talk to. Whether an outside auditor ever gets to look under the hood, and whether a regulator can say "no, not yet" to a risky release, is the difference between an industry that polices itself and one that answers to someone. We have watched the self-policing version play out in social media. It did not go great.

The skeptical read, and it is a fair one, is that Anthropic is playing a longer game. It is already built for this kind of scrutiny, so rules that reward safety-first companies also happen to reward Anthropic. Cynicism is cheap here. But even if the motive is partly self-interested, a frontier lab arguing on the record that "trust us" is no longer good enough is a rare and useful thing. The test is not what Anthropic says this month. It is whether the audits have teeth, and whether anyone with the power to stop a bad launch is actually willing to use it.