My bottom line
The next century will be decided by who leads on artificial intelligence, and leading means writing the rules, not just building the tools. America is not behind in this race. The question is whether we set the standards while we're ahead, or let someone else set them for us.
My rules are simple. The most capable systems get tested before they ship, the way we already test aircraft and weapons systems. A human stays in the loop in every life-or-death decision. Your face, your voice, and your name belong to you, not to anyone with a laptop. And where health, safety, and freedom aren't at stake, Washington stays out of the way. Rules where the stakes demand them. Restraint everywhere else.
Safe before it ships
Nothing this consequential deploys untested, and nobody gets to bury an incident report. That is not a radical idea; it's how this country already treats every other technology that can hurt people at scale. Binding federal safety requirements for the most capable AI systems: independent testing before deployment, and mandatory reporting of serious safety and security incidents.
When a system shows a critical failure mode, you ground the fleet until you understand it. You don't keep flying while lawyers argue about whose job it is to prove the crash. So regulators need real authority to pause development or deployment when a system crosses defined high-risk thresholds. And when a system could create catastrophic public risk, the burden of proof sits with the developer to show it's safe to proceed, not with the public to prove it isn't.
That includes the far end of this road. The United States should not permit deployment of a superintelligent system that cannot be demonstrated to remain reliably under human control. "We think we can control it" has never been an acceptable answer for any system that can kill people when control is lost. It isn't acceptable here either.
To be clear about scope: this is about the handful of frontier systems built by the largest companies on earth. It is not about the software a Bozeman startup ships, and I'll keep it that way.
A human decides
A human stays in the loop in every life-or-death decision. All of them. Targeting and strikes in war. Your VA claim. Your Medicaid coverage. Anywhere a machine's output could end or upend a human life, a person with authority and accountability makes the call. Machines can help us see faster and sort more information; they do not decide who lives. Speed never comes by deleting the human in the loop. That's already my standard for insurance companies, and it applies tenfold to your government and a hundredfold in uniform.
Nuclear decisions are categorically human, at every step of the chain. No exceptions, no automation, ever.
The same rule holds in cyberspace. An AI-enabled cyber operation against another country's infrastructure is a use of force, and a human with authority decides it. A system that picks its own targets at machine speed is how an incident becomes a war before a human being has seen the screen.
One honest boundary: the narrow place where automation already outruns human reflexes is intercepting incoming munitions, and those defensive systems stay bounded, supervised, and pointed at missiles, never at people.
And when your own government uses AI to decide something about you — your benefits, your claim, your rights — you get three things: the right to know AI was used, the right to an understandable explanation, and the right to appeal to a named human being with the power to change the outcome. That includes the VA. No algorithm denies a veteran's claim or flags a veteran's risk without a human who answers for it. We have already seen what the other path looks like: engineers handed the keys to Treasury and Social Security systems, and 300 million Americans' records copied into a database without the required security. Legal walls go around those systems, with real accountability.
Your face, your voice, your name
They belong to you. Not to a company, not to an ex, not to a scammer, not to anyone with an app. Nobody gets to manufacture "you" without your consent.
Start with the ugliest corner of this. When researchers first measured the deepfake internet, roughly 96 percent of it was fake intimate imagery of real people who never consented, and 99 percent of those people were women. The apps built to do this are one download away. I've sat with victims of image crimes most people can't imagine, and I can tell you the harm is not theoretical and it does not fade. Congress finally acted: the TAKE IT DOWN Act now requires platforms to remove this material within 48 hours, with real fines behind it. A law on paper is not enough. Fund the enforcement, pass the DEFIANCE Act so victims can sue the people who do this, and end the immunity of apps whose entire business is stripping and humiliating real people. No blanket immunity for what they build. That has been my position on AI and children from the start, and it holds here.
Then the scams. Last year, AI-enabled scams took more than $893 million from Americans, and $352 million of it from older Americans: the cloned voice of a grandchild begging for bail money, a fake emergency from a familiar number. The average victim over 60 loses nearly twice what a younger victim does, and most of this is never even reported. Voice-clone fraud is fraud. Enforce it like fraud, and hold the tools marketed for impersonation to account.
Then our elections. The consultant behind the fake presidential robocall in New Hampshire admitted what he did, was fined six million dollars by the FCC, has paid nothing, and was acquitted at trial. Right now, a lie in a candidate's voice two days before an election effectively costs less than a parking ticket. That is an argument for a real federal law, and Montana already drew the map: our state law bans unlabeled AI deepfakes of candidates in the 60 days before an election, with penalties that escalate to a felony. Montana led the country on digital privacy, and Montana led here. I'll take Montana's standard to Washington, and I'll write it carefully, because courts have already struck down versions that swept in satire. Deception is the crime, not comedy. Parody, satire, and commentary stay protected, full stop.
And pass the NO FAKES Act, which cleared committee unanimously, so that your likeness and your voice are legally yours, with the same carve-outs for news, criticism, and parody.
Proven safe in the job, before it takes the job
Where health, safety, or personal freedom is on the line, AI doesn't take over a job until it's proven safe in that job. The nurse reading your chart. The pilot in your cockpit. The caseworker who can cut off your benefits. Nobody hands those roles to a system that hasn't been independently tested in that role, and the burden of proof sits with whoever wants to deploy it. Prove it first. A human stays accountable for the call.
And that is exactly where the federal role ends. Where those stakes aren't involved, Washington has no business dictating what tools a Montana shop, farm, or business uses. I've watched one-size-fits-all rules from back east punish the people they were supposed to protect. Not on my watch.
The economics get the same honesty. AI's gains should not be captured by a handful of firms. Four meatpackers cornered Montana's ranchers; we are not letting five tech companies do that to the whole economy. That means real antitrust enforcement, real transition support tied to documented displacement, retraining and the trades pathways that let people land on their feet, and enforceable community benefit agreements when these projects profit off our communities, the same standard I hold data centers to.
Prove what's real
There's a quieter danger under all of this. When anything can be faked, everything real can be dismissed as fake, and the liars get the benefit of the doubt. Courts feel it. Journalism feels it. Military command feels it.
The answer is not to police every pixel. It's to make the real thing provable. I'll push content-provenance standards so official government communications can be cryptographically verified as genuine, and back the same standard as an industry floor. In a world where anything can be faked, the high ground belongs to whoever can prove what's real.
Where this meets my other fights
The infrastructure question, who builds the data centers and who pays for them, has its own full position: Montanans decide what gets built, and no one gets to spy on them. The data question, what happens to the information about you that these systems run on, lives in my privacy plan: one national law, built on Montana's, and a warrant before the government tracks you. The competition question, keeping America ahead and keeping our research and chips out of our adversaries' hands, lives in Technology & National Security. And the fight I've carried longest, protecting kids from what these systems can do to them, runs through all of it. One thread ties them together: these tools answer to people, never the other way around.
Sources
FBI data via the U.S. Senate Special Committee on Aging, "Age of Fraud" annual fraud report (2025 edition) · Sensity (formerly Deeptrace), deepfake landscape research · FCC, $6M forfeiture for the New Hampshire AI robocalls; verdict reported by NHPR, June 13, 2025 · Montana SB 25 (2025), enrolled text · Stanford HAI, 2025 AI Index · Also cited: the TAKE IT DOWN Act (enforceable since May 19, 2026); the NO FAKES Act of 2026 (S.4591); the DEFIANCE Act (pending, House Judiciary).