Congress came back the week of November 9 with a stack of unfinished AI bills. Two of them get talked about as if they were the same idea. The FRONTIER Act (H.R. 9925) would make the largest AI developers hire federally licensed auditors and would let the Secretary of Commerce suspend or restrict a model found to pose an “imminent catastrophic risk.” The AI Kill Switch Act (H.R. 9917) would make large AI providers keep the technical ability to throttle or shut down their systems, and would let the Department of Homeland Security order them to do it.
Both are pitched as “oversight.” Both are pitched as “safety.” And both supporters and critics use the word “accountability” as if everyone agreed on what it means.
They don’t. I think most of this debate is people arguing past each other because they’re using the same words for different things. So instead of picking a side, I want to take the problem apart, define the words, find where the sides actually agree, and then pinpoint the exact step where they split. I’ve numbered the steps. If you disagree with me, I’d like to know which step you get off the train at.
What the bills actually do
It’s worth being precise here, because the headlines aren’t.
The FRONTIER Act (Reps. Obernolte and Trahan, introduced July 23, 2026) requires large frontier developers to publish a safety framework and get an annual independent compliance audit. The very largest must also retain an “independent verification organization” (IVO), licensed by Commerce, for ongoing assessment. If an IVO concludes a model poses an imminent catastrophic risk, it must refer the matter to the Secretary within 72 hours. The Secretary can then issue an emergency order suspending or restricting the model’s development, deployment, or internal use. Provisional orders expire after 45 days and final orders after 90, with judicial review in federal court. It’s also in flux: Rep. Trahan is reportedly reworking it without her Republican co-sponsor ahead of the new Congress.
The AI Kill Switch Act (Reps. Lieu and Moran, introduced July 23, 2026) requires covered entities to maintain the ability to stop inference, cut off users, and shut down a covered system, and to report covered incidents within 15 days. If the DHS Secretary, consulting Commerce and the Director of National Intelligence, determines a covered incident has occurred, DHS can order “proportionate” action up to a full shutdown. The company can petition for reconsideration within 48 hours, but the petition doesn’t pause the order, and if DHS doesn’t answer within five days the answer is no. Penalties run up to $20 million per day for ignoring an order. It has been referred to a Homeland Security subcommittee and hasn’t moved since, though Rep. Lieu has said he wants it passed in the lame-duck session.
Neither is law. Neither has passed a chamber. That makes this a good time to get the concepts right.
Defining the terms
Here are the words this debate runs on, and how I’m going to use them.
- Accountability. When harm happens, you can trace it to a decision and to a responsible person or entity, such as a company. Any party that gets involved in something automatically picks up accountability for its part. An oversight committee is accountable for observing and reporting truthfully and thoroughly. Whoever holds a kill switch is accountable for using it correctly, and failing to use it when it should have been used makes them accountable for what happens next.
- Safety. Fewer bad outcomes. That includes harm to people, but also damage to property and other real costs. A shutdown that strands a hospital system is a safety outcome too.
- Control. Who or what holds the power to stop or change a system.
- Oversight. This one is a trap, because it has two meanings:
- Supervisory oversight is the ability to direct something, and with it, responsibility for it. A CEO has oversight of his employees.
- Observational oversight is watching and reporting. An oversight committee may have no power at all other than to publish what it finds.
- Licensing. In regulation, something is forbidden by default and a license makes it allowed. You can’t launch a rocket from the United States without one.
With those definitions, look at the bills again. The FRONTIER Act’s auditors are observational oversight. Its suspension power, and the entire Kill Switch Act, are supervisory oversight. Both get sold under the same word. And the FRONTIER Act’s licensed IVOs edge toward licensing, since a very large developer can’t operate without retaining one.
The common ground
Before I get to where the sides split, here’s what almost everyone already agrees on:
- Harm to people and property should be prevented where it reasonably can be.
- When harm happens, someone should answer for it.
- Very capable systems can cause harm at a scale that matters.
Safety advocates and the AI industry both sign on to all three. The disagreement is not about whether harm matters. It’s about who holds control and who holds accountability, and whether those are the same party.
Loop 1: Control and accountability have to travel together
Step 1. Accountability follows involvement. If you have the power to act and you act, or you fail to act, you own the result.
Step 2. Look at where the bills put the pieces. The government gets the switch. The duty to detect a problem and raise the alarm sits with someone else: the company’s 15-day incident report under the Kill Switch Act, or the IVO’s 72-hour referral under the FRONTIER Act. The party holding control is set up to act on other people’s findings.
Step 3. That’s the shape of control without accountability. The government wants the power to pull the switch, but makes it other parties’ job to tell it when. If the switch is pulled wrongly, the blame flows back to whoever filed the report. If it isn’t pulled, the blame flows to whoever didn’t file one.
Step 4. Now the two sides’ divergence is easy to name. Safety advocates treat control as if it were accountability: “If we can shut it off, we’re covered.” Industry treats observational oversight as if it were accountability: “We report, therefore we’re accountable,” even when the reporting carries no consequence for getting it wrong. Each side has taken one piece and called it the whole.
Step 5. My position: whoever holds control must also hold the accountability that comes with it. You don’t get one without the other.
The next question: if the government does hold a switch, what would make it accountable for using it?
Loop 2: Whoever flips the switch answers for it
Step 6. The person who uses the switch is accountable for using it. “They told me to” is not a defense, and neither is “the data was bad.”
Step 7. I can easily picture the failure mode. A politician flips the switch. Later it turns out it shouldn’t have been flipped, and the politician blames the agency that requested it or the company whose data drove it. If that’s possible, the system is broken. The official who holds the switch should be responsible for having a process where requests are independently reviewed and the decision is actually adjudicated before anything is flipped. Not arbitrarily, and not on a memo.
Step 8. This is really the old argument about good laws versus discretion. A well-written law tells you in advance what the standard is. Discretion lets appointed people write memos and enforce subjective judgments after the fact. The SEC is the clearest example of the second model. The IRS is more rooted in statute, but it still has room to decide how things should be accounted for. I’ve seen it many times: the SEC and the IRS want the same thing accounted for two different ways, neither one defines how it should be done, and they just kick it back or change it when it fails a test they never wrote down.
Step 9. A kill switch run on that model is control without accountability, with a very large blast radius.
To be fair to the drafters, the bills aren’t silent here. The FRONTIER Act requires a written finding, sets expiration dates on orders, and allows judicial review. The Kill Switch Act limits orders to defined “covered incidents,” such as unintended conduct that kills at least ten people or causes at least $100 million in damage, or a system subverting its own shutdown mechanism, and it requires a report to Congress. Those are real constraints. But the Kill Switch Act’s review comes after the order, the petition doesn’t pause it, and silence counts as denial. As I read the text, neither bill makes the official who orders a wrongful shutdown answerable for its costs, or answerable for failing to act. (Note that the Kill Switch Act itself counts economic damage as harm. So does my definition of safety. A wrongful shutdown is economic damage too.)
The next question: before we argue about when to flip the switch, what do we even mean by “safe”?
Loop 3: Absolute safety versus relative safety
Here’s another pair of terms that gets swapped mid-argument.
- Absolute safety: zero harm.
- Relative safety: less harm than the alternative, weighed against the benefit.
Step 10. The ban-hammer always wins on absolute safety. We could eliminate 100% of motor-vehicle deaths by banning vehicles and tearing up the roads. Nobody proposes that, because the benefit of cars and roads is worth the cost of some number of deaths. Everyone wants that number as low as possible, but the trade-off is clear. U.S. roads killed an estimated 36,640 people in 2025, about 1.10 deaths per 100 million vehicle miles, which NHTSA describes as the second-lowest rate on record. We accept that cost every year, deliberately.
Step 11. Banning all AI would reduce AI-related injury, too. That doesn’t make it the right answer, for the same reason.
Step 12. Here’s what people do without noticing: they hold machines to a far higher standard than they hold themselves. Every self-driving injury is treated as an absolute failure that should be near zero, while human drivers get the relative standard.
The data on driverless cars is already good enough to make the comparison. Through June 2026, Waymo has driven 271.3 million rider-only miles with no human driver. Compared with human drivers over the same distance on the same kinds of city streets, its published figures show 82% fewer injury-causing crashes and about 95% fewer crashes involving a serious injury or worse. An independent check comes from insurance data. A study with the reinsurer Swiss Re looked at 25.3 million driverless miles and found 88% fewer property-damage claims and 92% fewer bodily-injury claims than the overall driving population. Even against new 2018 to 2021 model-year cars with modern driver-assist features, the reductions were 86% and 90%.
The caveats matter, and I’d rather state them than have someone in the comments do it for me. This is one company’s fleet, in cities without much snow, mostly on surface streets rather than freeways. The crash-rate comparisons come from Waymo’s own researchers, though the methods are published in peer-reviewed journals and the raw data is downloadable. Human crashes are underreported, which the benchmarks only partly correct for. And there still aren’t enough miles to make a statistically meaningful claim about fatalities on their own. None of that flips the conclusion. It does mean “safer than human drivers, where it operates today” is the honest claim, not “safe everywhere.”
Step 13. The right question for any automated system is never “did it cause harm?” It’s “did it cause more harm than a human doing the same job?” An automated forklift should be judged against the human-driven forklift it replaces, not against a warehouse where nobody ever gets hurt.
The next question: if beating the human is the test, is it a one-time test, or a moving one?
Loop 4: The switching threshold versus a rising standard of care
Again, two things that get treated as one:
- The switching threshold: is the automated system clearly safer than a human doing the same task? If so, that justifies switching.
- The standard of care: the ongoing bar every product has to meet, which rises over time.
Step 14. The first baseline is the switching threshold. Clearly safer than a human is what justifies adopting the automated system at all.
Step 15. After that, safety keeps improving, and the standard has to keep pushing with it. A 1950s car and a car built today are far apart on safety. That isn’t because people in 1950 didn’t want safe cars. It’s because they couldn’t build them. Making a 1950s car as safe as today’s, with the same performance, would have been an Apollo-sized project.
Step 16. So the usual calculus is cost. Does it make sense to require a given improvement? The shoulder belt cost very little and added a lot of safety. Airbags are now cheap enough, and safe enough themselves, to clear the bar, but in the 1970s they wouldn’t have. NHTSA’s own numbers show the shape of that trade. It estimates seat belts add about $206 to a 2019 passenger car and dual frontal airbags about $365. In 2017 alone it estimated seat belts saved 14,955 lives and frontal airbags 2,790. If adding $100 to the price of a car makes it 1% safer, that’s probably worth requiring.
Step 17. Where I’m honestly unsure is where the requirement should stop and the free market should take over. If I can sell a robot for $1,000 that’s pretty safe and one for $100,000 that’s very safe, some people will buy each. Both choices can be reasonable.
The next question: where’s the line between what the law requires and what the buyer gets to choose?
Loop 5: A legal floor, a free market above it, and declared scope
Step 18. The difference between the two robots is who carries the risk. A buyer can choose the cheaper, riskier robot for their own property and their own life. But the neighbor’s kid, the pedestrian, the other driver didn’t make that choice.
Step 19. So the mandatory floor should be set by harm to third parties, the people who didn’t choose the risk. Above that floor, the buyer decides how much risk to accept for themselves and their own property.
Step 20. For the market to do that job, buyers need transparency into how safe something actually is. The $1,000 robot may be perfectly safe for cutting my grass, but it should be clear that I shouldn’t leave my infant in its care.
Step 21. That points to a requirement I’d support: manufacturers should have to state clearly what they’re willing to take responsibility for. Call it the declared scope. If the lawn bot runs over your dog while mowing your lawn, that’s inside the scope, and the manufacturer answers for it. If you tell it to mow inside your house, maybe not. Accountability moves to whoever moved the machine outside its scope, which is the same principle as Step 1: accountability follows involvement.
Step 22. There’s also a case that a lesser bot should be responsible for saying no. If it can recognize that it’s being used outside its scope, it should refuse.
Step 23. But there’s an obvious loophole. A manufacturer could have the robot refuse everything by default and require the user to override it for ordinary use, just to shed accountability. That would be bad. One way to close it, which I’d offer as a proposal rather than a settled answer: the declared scope must cover the product’s ordinary, advertised use. A lawn bot that has to be overridden to mow a lawn isn’t a case of the owner assuming the risk. It’s a defective product.
Notice that this loop doesn’t require anyone to hold a kill switch over the lawn bot. It works through clear rules written in advance, disclosed scope, and liability that lands on whoever was involved. That’s the same structure I’d want for much larger systems.
Where I land
Here’s the whole chain in one place:
- Accountability follows involvement.
- The bills separate control from accountability: the government holds the switch while others must tell it when to use it.
- That’s control without accountability.
- Safety advocates mistake control for accountability; industry mistakes observational oversight for accountability.
- Whoever holds control must hold the matching accountability.
- Whoever flips the switch answers for it.
- Requests to flip it must be independently reviewed and adjudicated first, and the official can’t hide behind the requester.
- That requires written law, not discretion and memos.
- A switch run on discretion is control without accountability.
- A ban always achieves absolute safety, at the cost of all the benefit.
- Banning AI would reduce AI harm the same way banning cars would reduce road deaths.
- We hold machines to an absolute standard and ourselves to a relative one, even where the data already favors the machine.
- The test is harm compared with a human doing the same job.
- Clearly safer than a human justifies switching.
- After that, the standard of care has to rise as capability improves.
- Requirements should follow cost versus benefit, the way seat belts and later airbags did.
- Above some line, the market should decide.
- The difference is who carries the risk.
- The legal floor is set by harm to third parties; the buyer chooses above it.
- The market needs transparency into real safety.
- Manufacturers must publish a declared scope and answer for harm inside it.
- Products should refuse uses they can recognize as out of scope.
- Declared scope must cover ordinary, advertised use, or the product is defective.
The open question, and where I want your help
The step I’m least sure of is where exactly the mandatory floor sits. I’ve said it should be set by harm to third parties and by cost versus benefit. But how much third-party risk is acceptable? Who sets the number, and how does it get written down so it isn’t one more subjective test applied after the fact? And for systems as general as frontier AI models, who counts as a third party when nearly everyone is?
If you disagree with any of this, tell me the step number where you get off. “I disagree at Step 4 because…” is far more useful to me, and to everyone else reading, than a general objection. I’d rather find the one place where we actually diverge than argue about the twenty-two where we don’t.