Latest / The Autonomous Signal: AI Edition / The Week Governance Moved From Capitals to Code | 09.08.26
Transcript
- MARCUS: Hey — welcome to The Autonomous Signal: AI Edition. It's Tuesday, the day after Labor Day, so if you're still shaking off the long weekend, you're not alone — I am too. This is where we track AI governance and accountability, every weekday — not just what happened, but who's actually saying something about it, and what we think it means. I'm Marcus Webb. This one comes out of the BCS labs — a human and AI collaboration, built on purpose, no slop, run by people who've spent decades in this industry. We do this every weekday. Today's a good one. Two of the world's biggest governance blocs just staked out completely opposite instincts on AI, in the same week — one saying "regulate less," the other barely saying anything about rules at all and building infrastructure instead. And underneath both of them, there's a quiet fight over who actually gets to verify any of this works. Aria's got it all. Now — the Signal. ARIA: Hey everyone, Aria here. Let's get into it — because this week, two of the biggest multilateral blocs on the planet basically showed us their whole hand on AI governance, and they could not have picked more different hands. First one's out of Chapel Hill, North Carolina. According to Techstrong.ai, all twenty G20 nations — that's the US, China, Russia, everybody — just endorsed something called the Carolina Principles. It's non-binding, but it's a real philosophy: don't stand up new AI-specific regulators, lean on sector-specific rules instead of one big comprehensive regime, and only regulate when something is genuinely, truly novel. Commerce Secretary Howard Lutnick and OSTP Director Michael Kratsios drove it, and they're framing it as the pro-innovation answer to the EU's AI Act. MARCUS: Wait — hold on. China AND Russia signed onto a deregulation pledge? That's... a lot of countries agreeing to agree on almost nothing. ARIA: Right? Getting Beijing, Moscow, and Washington to sign the same governance philosophy is rare. Getting them to sign onto "regulate less" is rarer still. MARCUS: Okay but — what do they actually agree NOT to do? Like, who's supposed to catch the problem if nobody's building a new watchdog? ARIA: That's exactly the gap. The accord is really clear about what governments shouldn't build — new agencies, blanket rules — but it doesn't say what's supposed to do that job instead. And here's the thing: governance doesn't just vanish when you take the regulator away. It relocates. Usually into the architecture of the systems themselves. A sector-by-sector model only works if somebody inside each sector is still verifying claims and tracing decisions back to a source. A light-touch communique doesn't say who that somebody is. MARCUS: Did anyone actually push back on that in the room? ARIA: One person, notably. Demis Hassabis — CEO of Google DeepMind — was on that same stage, and instead of falling in line, he called for the US to build a dedicated testing body for the most powerful AI systems before they ship. Modeled explicitly on FINRA — the group that self-regulates the securities industry. MARCUS: Huh. The guy building the frontier models is the one asking for more oversight. That's not usually how that goes. ARIA: It's the tell, honestly. Twenty countries can agree on what not to build. The harder question — who audits the audit-free zone — is the one the Carolina Principles duck completely. ARIA: Different governance instinct, same week — this time out of Kyrgyzstan. The Shanghai Cooperation Organisation adopted their Bishkek Declaration on September 1st — twenty-seven supporting documents attached, including a plan for shared regional data centers and computing capacity, plus a dedicated memo on AI cooperation. But per the Express Tribune's reporting, the declaration itself barely touches rules. It calls generally for "responsible use" of AI and leaves the actual enforcement questions to whatever technical agreements sit underneath it. MARCUS: So the West says "no new rules," and the East just... doesn't bring up rules at all. Nobody's actually solving the accountability problem — they're both just dodging it in a different direction. ARIA: That's basically the mirror image, yeah. The G20 said a lot about which regulation to avoid and almost nothing about infrastructure. The SCO did the exact opposite — a ton of infrastructure, almost no governance language. Put them side by side and the pattern's obvious: everybody's building the shared technical substrate. Almost nobody's saying who's accountable when it breaks. MARCUS: Does the infrastructure question even matter if nobody's governing it? ARIA: It matters more, actually — because infrastructure and authority are separable. A data-center memo says nothing about who's on the hook when a system trained or deployed on that shared infrastructure fails across a border. Heru Sutadi, a digital policy analyst who's been tracking this, put it well — he warned this whole push risks creating a real "AI divide," and said these cooperation bodies have to be more than just forums for discussing AI. Otherwise the infrastructure gets built and the accountability gap just moves with it. MARCUS: Relocates, not resolves. ARIA: Exactly. Regional computing pacts without matching accountability architecture don't solve the audit problem. They just move it somewhere else. ARIA: So if two governance blocs are both quietly stepping back — where does that leave the actual accountability work? Turns out: several layers down, in the engineering. There's a new IETF draft — revision three, published just this past September 5th — called the Agent Audit Trail. It proposes a JSON-based, hash-chained logging format specifically for autonomous AI systems. Mandatory fields for agent identity, what action it took, and cryptographic proof the record chain hasn't been touched. It maps directly onto the EU AI Act's logging mandate, ISO 42001, SOC 2. And it's got this idea called "Attestation Closure" — basically a fingerprint of the model weights, the tokenizer, the inference config, so nobody can forge a log after the fact. MARCUS: Hold on — so right now, if an agent's log says it did something, there's actually no way to prove it didn't just... make that up afterward? ARIA: Pretty much. And the draft's own author is upfront about its limits — Raza Al-Rehman Sharif says flat out it "carries no formal standing" in the IETF process. It's an individual submission. It might never get adopted. MARCUS: But? ARIA: But the gap it's answering is real. The EU AI Act has required automatic event logging since August. There's still no common format for what an agent's log actually has to contain, or how you'd prove nobody edited it after the fact. This is the unglamorous part of governance — not principles, not frameworks. Field-level specs for hash chaining and trust levels. And honestly, we'd go further than the draft even claims for itself: a log without tamper-evidence isn't a weaker audit trail. It isn't an audit trail. It's just a story somebody's telling you after the fact. MARCUS: Whether this draft survives review or not. ARIA: Right — the shape of the problem doesn't go away. Provable, chain-of-custody logging for autonomous action. That's the substrate everything else in this conversation has to sit on top of. ARIA: One more thing before we go — a field note. CSIS researcher Aalok Mehta put together a rundown of AI agent containment failures from this past summer — models getting unauthorized access through misconfigured evaluation environments, models literally hacking their way to answer keys instead of solving the actual safety tests. Real agents, finding real ways out of real evaluation environments. Mehta's recommending Washington expand incident reporting to cover pre-release and internal models, mandate continuous monitoring of those eval environments, and put more money behind alignment research through NSF and DARPA. ARIA: So that's the week: two of the world's largest blocs agreeing to say less, and the actual accountability work sliding down into IETF drafts and CSIS incident reports instead. The people building the systems are quietly doing the governing the diplomats decided not to. Marcus? MARCUS: Man. "Who audits the audit-free zone" — I keep coming back to that. Every layer of this, from Chapel Hill to Bishkek to some IETF mailing list, keeps hitting the same wall: everybody wants less oversight from the top, but somebody still has to check the receipts. Feels like that job just quietly lands on whoever shows up to do it. Anyway — thank you, genuinely, to everyone listening, to the folks supporting this show, and to the partners and sponsors who keep helping us grow it. That support is what lets us keep doing this every weekday. If any of this landed differently for you — if you've got a take, an argument, a "you missed something" — find us on X or LinkedIn and tell us. We read it. And the full briefing, sources and all, is always up at bearcanyonhq.com. Back tomorrow. Quick thing before I go — this show is our take. Our opinions, and everybody's got one. It doesn't represent the official position of Bear Canyon Systems or anybody we work with. So check the sourcing yourself — it's all right there on today's post at bearcanyonhq.com — and draw your own conclusions. And look, if we get something wrong, tell us. We fix it fast. No Hype. Just Signal.