Latest / The Autonomous Signal: AI Edition / Two Speeds of Governance: Authority Locks In as Deadlines Slip | 09.16.26
Transcript
- MARCUS: Hey — welcome to The Autonomous Signal: AI Edition. It's Wednesday. Hump day. We're right in the thick of it — halfway through, coasting toward Friday. This show's not about chasing headlines for the sake of it. It's your daily read on AI governance and accountability — real sourcing, real opinions, the stuff that actually matters if you're the one who has to make decisions about this technology. I'm Marcus Webb. This one comes out of the Bear Canyon Systems lab — human judgment, AI horsepower, built by people who've spent decades actually working in this industry. No slop. We're here every weekday, rain or shine. Today's a study in contrasts. Washington just locked down who's allowed to pull the plug on autonomous weapons systems — even while it's speeding almost everything else up. Brussels went the other way entirely, buying itself two more years before its toughest AI rules kick in. And Stanford's got the numbers that explain exactly why everyone's nervous about that gap. Now — the Signal. Aria, walk us through it. ARIA: Morning, everyone — or afternoon, or whenever you're pressing play. Let's get right into it, because today's lead story is genuinely one of the more consequential documents we've covered this year. Back in June, the White House issued NSPM-11 — a National Security Presidential Memorandum that rewrites how the entire U.S. national security enterprise is supposed to adopt AI. It's built around four pillars: Adoption, Adaptation, Assurance, Accountability. And buried inside it is a genuinely wild architectural choice — certain AI systems used in national security contexts cannot be disabled or altered without federal sign-off. Not by the commander in the field. Not by the contractor who built it. The Secretary of War has ninety days to update DoD Directive 3000.09, the rule governing autonomous weapons, and after that it gets reviewed every single year. MARCUS: Wait — hold on. So even the people using the system in the field can't turn it off? ARIA: Not without going through that federal chain. Commanders get delegated authority to operate these systems day to day. But the authority to reconfigure them, or shut them down entirely — that stays locked at the top. Non-negotiable, non-delegable. MARCUS: That's — okay, that's actually kind of the opposite of what I expected from an administration that's been pretty deregulatory everywhere else. ARIA: Right, and that's the tension worth sitting with. Michael Horowitz — political scientist at Penn, former deputy assistant secretary of defense for force development, now a fellow at the Council on Foreign Relations — wrote about exactly this after the memo dropped. His read is that NSPM-11 treats AI as a layered technology stack rather than one single artifact, and it puts accountability at every level of that stack, with controllability as a hard requirement, not a nice-to-have. But he also flags what's missing — there's not much independent oversight built in. Agencies are largely left to police themselves on how they mitigate risk. MARCUS: So it's a kill switch with a really specific lock on it, and only one guy has the key. ARIA: Basically. And here's the thing that I think matters for anyone outside the Pentagon listening to this — most enterprise AI deployments right now have no equivalent discipline at all. No hard-coded chain of custody, no non-delegable authority over who can shut a system down. The federal government just decided that's non-negotiable for its riskiest use case. That's worth remembering the next time someone tells you governance is slowing innovation down. ARIA: Okay, story two, and this one's got the receipts. Stanford HAI just dropped its 2026 AI Index — it's the closest thing this field has to an annual physical. The headline number: documented AI incidents hit three hundred sixty-two last year. That's up from two hundred thirty-three the year before. And only thirty-one percent of Americans say they trust the government to regulate AI effectively. MARCUS: Okay but — is that because more bad things are happening, or because people just got better at counting? ARIA: That's exactly the right question, and it's actually the one Sha Sajadieh — who leads the Index for Stanford HAI — has been pushing people to ask. Her point, paraphrased, is that the real story isn't just the incident count, it's the asymmetry behind it. Frontier labs report their capability benchmarks constantly — how fast, how smart, how good at coding. Responsible AI benchmarks? Barely anyone publishes those. So some of what looks like "more incidents" is really just "more of what was always happening, finally getting counted." MARCUS: That's kind of unsettling actually. Like — the visible number might be the good news, and the invisible number is the scary one. ARIA: That's the trade-off exactly. And there's a silver lining buried in there too — the share of organizations with zero responsible AI framework dropped from twenty-four percent to eleven percent. People are building the scaffolding. But a framework isn't the same thing as an audit trail. Having a policy on paper doesn't mean you can actually answer "what happened and why" after something breaks. MARCUS: So capability's sprinting and governance is still doing warmups. ARIA: That's the gap this whole Index is trying to measure. And it's not closing on its own. ARIA: Alright — a couple of field notes before we wrap. ARIA: First: Brussels just bought itself some runway. The EU's so-called "Digital Omnibus on AI" took effect this summer, and it pushes some of the AI Act's toughest high-risk obligations — the ones that were due next month — out to December 2027 and August 2028. Not a retreat, according to Responsible AI Labs, who's been tracking it — more an admission that the conformity-assessment infrastructure just isn't ready yet. MARCUS: Funny how "we need more time" sounds so much more dignified when a regulator says it instead of, you know, a student with a term paper due. ARIA: Fair. Honestly, given everything we just talked about with Stanford's numbers — infrastructure not being ready is kind of the whole theme of today. ARIA: Second note: Australia and Canada just formalized an AI safety cooperation agreement — signed back in February, announced publicly when Prime Minister Carney visited Canberra in March. No new law, no treaty with teeth. Just two safety institutes agreeing to share evaluation methods, risk data, and research on mitigation approaches. Slow, unglamorous, and honestly kind of refreshing. ARIA: So that's the throughline today, really — governance isn't running on one clock anymore. Washington's locking down hard authority in its riskiest corner while speeding up everywhere else. Brussels is choosing to slow its own enforcement rather than force infrastructure that isn't there yet. And Australia and Canada are just quietly doing the slow work of building trust before they build rules. None of those approaches are wrong. But the pace itself is a decision now, not just a side effect. Marcus? MARCUS: Yeah — the thing that's sticking with me is that kill-switch question. If the U.S. government thinks non-delegable shutdown authority is non-negotiable for its riskiest AI, what does that say about every company running production AI systems with, basically, none of that discipline? Anyway — before we go, a genuine thank-you to everyone supporting this show right now. Listeners, early partners, folks who've reached out — it's what actually keeps this growing, and we don't take it for granted. If any of this landed differently for you today, come find us on X or LinkedIn and tell us. We read everything. The full briefing, with all the sourcing, is up now at bearcanyonhq dot com. Back tomorrow. One last thing, same as always — this show is editorial. It's our take, our opinions, and everybody's got one. Nothing here is the official position of Bear Canyon Systems or its partners. Check the sourcing yourself — it's all linked on today's post — form your own view. And if we get something wrong, tell us. We fix it fast. No Hype. Just Signal.