Latest / The Autonomous Signal: AI Edition / Who Acted? Courts, Insurers, and Researchers Chase the Same Question | 08.11.26
Transcript
- MARCUS: Hey — welcome to The Autonomous Signal: AI Edition. It's Tuesday, so we're past the Monday scramble and the week's actually starting to move. I'm Marcus Webb. Every weekday, this show gives you a real read on AI governance and accountability — not headlines, not hype. Real sourcing, real opinions, from people who've spent decades in this industry, working alongside the AI in our labs at BCS to get it to you fast. No slop. Just signal. Today's a big one. A federal appeals court just answered a question everybody's been dancing around — who's actually responsible when an AI agent acts on your behalf? At the same time, insurers are quietly deciding they don't want that risk on their books at all. And two new papers are trying to build the infrastructure underneath both of those decisions before the next case forces the question. Now — the Signal. Here's Aria. ARIA: Hey everyone — Aria Chen here. Let's get right into it, because today's briefing is basically one question wearing four different outfits: when an agent acts, who actually answers for it? Let's start in a courtroom. The Ninth Circuit ruled this week, and according to The Decoder's reporting on the case and the court's own opinion, when Perplexity's AI shopping agent buys something for you on Amazon, it's you who legally "accessed" the site under federal computer-crime law — not the software. MARCUS: Wait, hold on — so the actual reasoning is just, the statute says "whoever," and a shopping bot isn't a "whoever"? ARIA: Basically, yes. The panel vacated the injunction that had been blocking Perplexity's Comet browser from acting on Amazon. Eric Goldman — he teaches law at Santa Clara and has been tracking this case closely on his Technology and Marketing Law Blog — called it a clean fix, said the court replaced bad precedent with a better one. But here's the part that doesn't get solved by the ruling: if responsibility collapses onto "the user" every time, what happens when that instruction passed through five delegation hops before any human actually reviewed it? MARCUS: So it's like blaming the person who wrote a check, even if the amount got changed six times before it cleared. ARIA: That's actually a good way to put it — except with agents, nobody's even sure the check crosses a human's desk before it clears. That's the real design problem underneath this ruling. Courts can allocate blame after something's gone wrong. What actually prevents the mess is a system built so the chain from your intent to the agent's action is provable — not something you're reconstructing in a deposition six months later. MARCUS: Provable beats "I'm pretty sure I told it to." ARIA: Every time. And that's the quiet warning in this ruling. It's a win for agentic AI companies today, but it's also a bill that's going to come due for whoever can't show their work. ARIA: Story two, and this one's got a number in it that stopped me cold. There's a new paper out of the NetX Foundation, from a researcher named Anbang Ruan, that takes a sociology framework from the nineteen-sixties — Talcott Parsons' AGIL model — and uses it to grade whether an agent ecosystem is actually governed, or just running. MARCUS: Okay, grade it against what, though? ARIA: Against OpenClaw — an agent ecosystem with more than seven hundred seventy thousand registered agents. By Ruan's framework, it hits nineteen percent coverage. Zero coordination between the institutional functions that are supposed to hold a system like that together. MARCUS: Nineteen percent? That's not a gap, that's basically nothing. ARIA: Right. And Gillian Hadfield, who's the Bloomberg Distinguished Professor of AI Alignment and Governance at Johns Hopkins, has made basically this exact point for years — that we're not actually short on the tools here. Her framing is something like, we know how to build institutions, we know how to design incentives, we know how to coordinate across borders when the stakes actually demand it. The problem was never capability. It's that nobody's building the judiciary and the enforcement mechanism before they launch the constitution. MARCUS: Seven hundred seventy thousand agents and no judiciary. That's a lot of chaos with nobody to call. ARIA: And an honest coverage score is the uncomfortable part — but it's also the only kind worth trusting. Ruan's paper isn't saying agent societies can't be governed. It's saying almost nobody's actually tried yet. ARIA: Story three, and I think this is the one enterprises are sleeping on. Per Honigman's analysis, the Insurance Services Office just rolled out an endorsement called CG 40 47. It lets commercial liability carriers flatly exclude anything "arising out of generative artificial intelligence." Bodily injury, property damage, advertising injury, all of it. MARCUS: Wait, flatly excluded? Even if the AI's just one small piece of a much bigger product? ARIA: That's the concerning part. The definition's written broad enough to sweep in almost anything that produces text, images, audio, or code. Alana McMullin, a partner at Lathrop GPM who represents policyholders in exactly these fights, has been telling reporters there's been — her words — "a major shift in the insurance industry's treatment of AI-related risks," and that insurers are moving fast to limit their exposure. Insurers are good at one thing: pricing what they can model, and excluding what they can't. Agentic AI just landed in the second bucket. MARCUS: So if the people whose entire job is calculating risk for a living won't touch it — ARIA: — why would anyone else take that bet uninsured? That's exactly the question. And it lands the same month regulators are tightening the obligations on these same systems. So the enterprise is left holding a much bigger bag than it thinks it is, unless the contracts and the governance underneath are doing real work. MARCUS: Rough week to be a general counsel. ARIA: Rough year, probably. ARIA: Okay — a couple of quick field notes before we go. First one — Aprio, a top-twenty accounting firm, just got accredited to independently certify companies against ISO 42001, the big international AI-management standard. Not flashy, but it means AI governance is quietly becoming the same kind of routine third-party audit that security and finance already go through. MARCUS: Funny how that lines up with everything else today. Courts want proof, insurers want proof, and now accountants are showing up to actually certify the proof. ARIA: Yeah, same throughline all day — assertions aren't enough anymore. Everybody wants receipts. Last one — a legal scholar, Yiheng Yao, published a framework for sorting AI-agent harm into three buckets before anyone even gets to court: pure tool use, collaborative planning, and what he calls autonomous drift. The interesting part is he wants the agent's own interaction log treated as the evidence — the record of exactly where a human-AI plan went off script. ARIA: So that's the day — a court, an insurer, and two papers, all circling the same open question. Nobody actually wants to be the one holding the liability when an agent acts. Which means the winners here won't be whoever moves fastest. It'll be whoever can actually prove what they authorized. Marcus? MARCUS: Yeah — that "prove it, don't assert it" thing keeps coming up. Makes you wonder how many companies right now think they're covered — legally, financially, whatever — and just aren't, and don't know it yet. Anyway. Thanks to everyone listening, subscribing, sharing this with a colleague — genuinely, that's what keeps a show like this growing, and we don't take it for granted. Same goes for the partners and sponsors backing us as that side of things comes together. If anything today landed differently for you, come find us on X or LinkedIn and tell us — we mean that. And if you'd rather just support the show quietly, hitting follow does more than you'd think. The full briefing, sourcing and all, is up now at bearcanyonhq dot com. We'll be back tomorrow. One last thing, same as always. This show is us — our take, our opinions, and like anybody, we've got them. None of it represents the official position of Bear Canyon Systems or its partners. Go check the sourcing yourself, it's all on today's blog post, and draw your own conclusions. That's kind of the point. And we're human on this end too — mistakes happen. If we get something wrong, tell us, and we'll fix it fast. No Hype. Just Signal.