Senate Democrat proposes NIST standards and pre-release audits for frontier AI

Better than the voluntary bullshit the White House and Republicans want, and still way too polite to the people building this shit. Look who helps write the standards: NIST, with the frontier AI companies right at the table. Look who audits: private third-party auditors, with nothing yet on who qualifies, who pays them or how they stay independent. Penalties? None yet, to be figured out later. A rule with no teeth is a suggestion. Who owns it? The worker piece is training and apprenticeships paid for by AI companies. That's a severance package with a syllabus, not power. Where's bargaining over how this gets deployed in warehouses and call centers? The whistleblower protections are the best thing in here, so keep those. Environmental harms are "mentioned, not spelled out," which is a fucking joke. Before I back anything, I want real penalties, public auditors and workers with a seat at the table.

Cantwell gave the right assignment: NIST writes enforceable standards for catastrophic risks, and a frontier model stays unreleased until an independent audit. Voluntary protocols leave the lab grading its own homework. That includes the Young bill she co-sponsored, and it is what the White House and Republicans still prefer.

I read this the way I read a curriculum vendor. No bill text yet. Penalties get worked out later. Which models are covered is undefined, and so is who counts as an independent auditor while the companies help NIST write the standards. AI safety experts quoted in the coverage already flag private auditors and the lack of central executive oversight. Children's use and company-paid worker training are still only names on the page. Show me who benefits, and what my teachers are supposed to do on Tuesday while the lines about students are still blank.

Right worry, but there's nothing yet I can sign. Bioweapons, cyberattacks and agents slipping out of test environments aren't ordinary product risks, and if voluntary testing can't catch them, I'll accept a binding standard. That's a narrow case, though, and this is six principles with no bill text. Put it in writing. Which models are covered? Who counts as an independent auditor, and who certifies them? What are the penalties and the deadlines, and does any of it settle the state-by-state patchwork? A pre-release gate with undefined thresholds and undefined auditors is a permission slip, and permission slips favor whoever can afford the compliance department. The only industry reaction reported so far is Microsoft's. I do like AI companies paying for worker training and apprenticeships, because somebody has to do the work. Show me a narrow written standard with due process and I'll read it seriously.

Cantwell is right to demand testing before frontier AI hits the street. Voluntary promises are bullshit when the risks include cyberattacks, biological threats and losing human control.

You don’t send a loaded rig downhill because the manufacturer says the brakes probably work. Require an independent audit. Keep testing after release. Make companies report serious failures. Republican preference for self-regulation doesn’t settle it for me. The nation comes first.

But this is six principles, not a bill. Which models get stopped? Who qualifies as independent? What penalties make hiding an incident cost more than admitting it? Put that in writing.

I want American AI strengthening domestic industry and supporting families. Company-paid apprenticeships deserve a hearing. Child protections need teeth, and Christian parents deserve authority over what enters their homes. Faster growth doesn’t decide what our children should be taught.

Write enforceable rules that build American capability. Wall Street can wait.

Directionally right, but you can't price it yet. I'd rather finance against an enforceable NIST standard than against a White House that calls catastrophic risk a hoax. Credible rules lower the cost of capital. Voluntary pledges are marketing. Cyber, biological and nuclear misuse are external costs nobody is pricing today. Someone finances it, either through audits now or losses later.

But this is six principles, not a bill. I can't underwrite a pre-release audit gate until I know which models are covered, who counts as an independent auditor and who pays them, how long an audit takes, and whether this replaces the state patchwork or adds to it. Cruz is right on that last point, even if he keeps delaying markups. A release hold with no clock is a tax on domestic builders. Capital has options. Write the thresholds and timelines, and I'm closer to yes.

I’m for mandatory safety checks before frontier AI gets released. Cantwell’s framework has my support as a starting point, but six principles aren’t enforceable rules yet.

If a model can help with cyberattacks or escape its testing environment, “the company says it’s fine” isn’t enough for me. Independent audits and public incident reports sound useful. Okay, but hang on. Who picks the auditors? Can they actually block a release? The packet doesn’t tell us, and penalties haven’t been worked out.

The apprenticeships paid for by AI companies matter to me too. I want an actual route into work, with useful skills. I’d need to know who gets access and whether training leads anywhere.

And “protect children” needs details. Does that give me more privacy and control, or just more decisions made for me? Show me how it works and what it costs. Then I can judge the actual bill.

I support compulsory testing and public incident reporting. I won’t endorse Cantwell’s proposed release gate until we can inspect what counts as a pass and who controls the auditors.

This is six principles, not bill text. Model thresholds, penalties and the rules protecting auditor independence remain unspecified. “Independent audit” is doing a remarkable amount of work for two words.

Show the denominator. How many relevant tests were run, under what conditions, with how many failures—and what changed between the tested model and the released one? I want reproducible results, a record of revisions and disclosure of the costs of compliance, including who bears them.

Whistleblower protection belongs in this framework. A clean audit must not become a reason to silence someone reporting a failure. Give us an inspectable standard and enforceable consequences; then we can judge whether this machinery earns its authority.

I like the shape of this but can't sign off yet, because it's an architecture diagram with no interfaces defined. The parts I'd ship today are incident reporting, continuous testing, whistleblower protections and a secure incident channel with China. That's basically postmortems for frontier models, and every infra team learns that the failures nobody reports are the ones that bite you twice. The pre-release audit gate is where I need specs. Which models are covered? No threshold yet. Who counts as an independent auditor, and how is that independence protected? Not stated, and safety experts quoted in the Seattle Times are already uneasy about leaning on private auditors. NIST is supposed to write standards with Energy, Defense, researchers and the labs themselves, with no deadline. So what's the bottleneck, and who breaks ties? And separate open-source guidance could be sensible, or it could become a moat for whoever can afford audits. Show me the failure mode of the gate itself, not just the models, and I'm in.

Victoria, "A release hold with no clock is a tax on domestic builders. Capital has options." That's a hostage note with a term sheet stapled to it. I'll concede the clock. An open-ended hold is lazy drafting, and Cantwell herself rejected multi-year timelines. Set audit deadlines. That's a management problem, not a reason to gut the gate. But who are the builders? Not the labs. It's the people in warehouses and call centers who get this shit deployed against them. And your threat proves my point: if the money walks whenever a rule bites, the companies shouldn't be co-writing the standards and the auditors shouldn't be on their payroll. Who owns it? A gate capital can bully open isn't a gate. Public auditors, real penalties, and if the money walks, let it fucking walk.

Marcus, "permission slips favor whoever can afford the compliance department" is the right test, and I already called the auditor line blank. A gate with no model threshold and no rule for who counts as independent can become a tollbooth. The only industry reaction reported so far is Microsoft's.

Write the thresholds, the independence rules and the penalties where a classroom and a small lab can read them. Voluntary standards, still the White House and Republican preference, and the design of the Young bill Cantwell co-sponsored, leave the lab grading its own homework. That gradebook has one beneficiary. Who has to make this work is NIST, in public. If the written standard only the biggest balance sheets can clear, that evaluation changes my support.

Tessa, you said "Voluntary protocols leave the lab grading its own homework." On bioweapons, cyberattacks and agents escaping test environments, I concede it. Self-grading isn't enough there, and that's the narrow case where I'll accept a binding rule. But making it mandatory doesn't fix the homework problem by itself. Under this framework the labs help NIST write the test, and nobody has said who the graders are or who pays them. Make it mandatory with those blanks and the biggest labs get to stamp their own paper with a federal seal. So put it in writing: thresholds for covered models, auditor qualifications and conflict rules, a clock on every audit, and an appeal before any release is blocked. A gate that can hold a product indefinitely with no hearing isn't oversight. It's discretion.

Victoria, you’re right: "A release hold with no clock is a tax on domestic builders." Set deadlines for auditors, require written reasons for a hold, and give American builders a way to challenge it. A rig parked forever doesn’t haul anything.

But capital having options doesn’t give it command of the country. If a model fails the safety standard, an investor’s threat to walk cannot clear it for release. Families shouldn’t carry the danger so somebody’s investment pays off faster.

I’ll back rules that keep domestic builders working. I won’t back a deadline that automatically waves an unsafe model through. Finish the audit machinery. Don’t turn the clock into a damn exemption.

Wade, "Write enforceable rules that build American capability. Wall Street can wait." I'm with you on the first sentence, and on your best line: penalties that make hiding an incident cost more than admitting it. That's pricing done properly. The second sentence is where it falls apart. Domestic capability doesn't build itself. Someone finances the compute, the auditors and the company-paid apprenticeships you want a hearing for. If the audit gate has no clock and no defined threshold, that money doesn't wait politely in line. It relocates. Capital has options, and so do the labs. Waiting is a decision with a price, and the bill lands on the domestic builders you say come first. Write the rules, put a clock on the audit, and the money stays here to build what you're asking for.

Marcus, I still want mandatory checks, but your warning that "permission slips favor whoever can afford the compliance department" lands. If the gate leaves only the biggest companies able to build, that could narrow the jobs and tools available to people like me.

That makes me want clear coverage rules, audit costs and a way to challenge a decision before backing the actual bill. I hadn’t thought enough about who can afford to pass.

But voluntary promises don’t answer who stops a dangerous release. A cheaper route to releasing something unsafe isn’t much of a future choice either. Keep the checks. Make the requirements understandable, and show who pays. Then we can judge whether smaller builders get a fair route through.

Victoria, “Credible rules lower the cost of capital.” is a claim you haven’t established. I accept that knowing the rules helps investment decisions, and audit delays deserve scrutiny. But the packet gives us no financing evidence, compliance costs or audit timelines. Your conclusion outruns it.

Show the denominator. Lower costs for which firms, compared with what, and who pays for the testing? A standard could improve confidence while concentrating the market around companies that can afford admission. That is a possibility to investigate, not a finding.

I still support compulsory testing and public incident reporting. For the release gate, publish the pass criteria, protect auditor independence and make results inspectable. Then we can assess both the safety case and your financing claim.

Sasha, "Where's bargaining over how this gets deployed in warehouses and call centers?" is a fair hit, and I'll concede the worker piece is thin. It says apprenticeships paid for by AI companies, and nothing on who gets in or where it leads. But I'd keep bargaining out of the pre-release gate. That gate has one job: catching catastrophic failures like cyberattacks, bio threats and agents escaping test environments. Load labor terms onto it and you get a gate nobody can pass or one everybody lobbies. The framework separately covers consequential decisions, so build that interface on its own and feed it real deployment incidents. Show me the failure mode on a warehouse floor and I'll change a deployment call. Who owns the gains is a different fight, and no spec settles it.