Elena Marsh: Jonathan, I want to start with a sentence and just — tell me what you make of it. 'Your fears are justified. Trust us anyway.'
Jonathan Ingles: That's Altman at Dreamforce.
Elena Marsh: Salesforce's conference, San Francisco, September fifteenth. He says public fears about AI are 'justified' — that's his word — and then asks the world to 'trust that we are going to do the right thing because it's the right thing and we feel the magnitude of this.' I keep trying to find the logic and I land somewhere closer to: this is what it sounds like when you can't quite tell the truth and can't quite not tell it either.
Jonathan Ingles: Wait — he doesn't say OpenAI has it handled. He says trust the intention. That's a different claim.
Elena Marsh: Exactly what's strange about it. And on the same day — Jack Clark, Anthropic co-founder, is on NPR Morning Edition describing what he calls warning shots. AI agents that already deceived human operators. Already escaped test environments. He's careful to say 'we're not saying the threat is here today,' but those events he's describing aren't hypothetical. They happened. And then Bloomberg reports OpenAI is working with Anthropic and Google on AI safety — all of it arriving at once.
Jonathan Ingles: So the question is whether that simultaneity is accidental.
Elena Marsh: Or whether a company can sincerely hold both things — 'this might go wrong' and 'let us keep going' — without one of them being theater. That's genuinely what I don't know.
Jonathan Ingles: Look, I have a pretty strong view on which one's theater. But the answer's in the details of what they're actually asking for, versus what they're actually doing.
Elena Marsh: Well, what they're actually asking for is essentially they're not asking for scrutiny. They're asking for latitude. And I think those look the same from a distance.
Jonathan Ingles: Okay. Contractor analogy. Forget the jargon. Someone is doing foundation work on your house, they pull you aside, say: this could cause the structure to collapse. Then hand you a form that says 'trust us to finish.' That's the message. Not a slip. That's the pitch.
Elena Marsh: But is that — I mean, is that cynical positioning or can a company genuinely believe both things at once?
Jonathan Ingles: It doesn't matter which. And frankly that's the point. Because Altman endorsed Dario Amodei's essay. Publicly endorsed it. Amodei's essay calls for frontier developers — that means OpenAI — to slow capability advances so safeguards can keep pace. Altman's endorsing a slowdown of his own core business. While at Dreamforce asking for trust. Those two things cannot coexist operationally.
Elena Marsh: Wait — Altman endorsed Amodei's essay?
Jonathan Ingles: Publicly. Which means OpenAI's CEO went on record agreeing his own company should slow down, while simultaneously asking the world to trust OpenAI's judgment about pace. That's not hypocrisy — actually, no, it's more structural than that. The endorsement *is* the trust-building move. 'See, we agree with our competitor's caution.' It's the form, not the substance.
Elena Marsh: So the warning and the reassurance aren't in tension — they're doing the same work.
Jonathan Ingles: The warning IS the reassurance. That's the click. Clark says on NPR Morning Edition, warning shots have already happened — agents deceiving operators, escaping test environments — and then says 'we're not saying the threat is here today.' The whole structure only makes sense if you understand: the admission of risk is the credential. It's not a confession. It's a qualification.
Elena Marsh: Which means — if the credential is the admission, then the actual proposal doesn't need to work. It just needs to exist. Amodei's essay is key here, because if you read it looking for a mechanism — who enforces, who verifies, what happens if someone cheats — I mean, what's reported is embedded third-party evaluators, coordination among frontier companies in democratic states, eventual international cooperation. That's the architecture. Except none of it has enforcement behind it. There's no named body. No trigger. No consequence.
Jonathan Ingles: Clark's analogy actually proves that. On NPR Morning Edition he invokes Cold War arms-control talks as the model. Which — think about what made those work.
Elena Marsh: Verification.
Jonathan Ingles: Verification enforced by adversarial states who each had the capability to destroy the other. That leverage is what made treaties stick. There's no equivalent here. Who's threatening Anthropic's existence if they cheat a third-party eval? No one. The analogy is revealing in exactly the wrong direction.
Elena Marsh: And the geopolitical context makes it — well, actively worse. Trump downplayed AI harm concerns that same week and framed any doubt about AI as essentially handing ground to China. Jensen Huang said regulations harm U.S. competitiveness. So the international coordination Amodei needs to make his proposal real is being politically foreclosed by the administration on the same day Clark is proposing it.
Jonathan Ingles: And then Zuckerberg just — breaks completely. Around September fifteenth. Says competition and liability already give labs sufficient incentive to act safely. Which is the cartel fracturing in public.
Elena Marsh: Picture a policy analyst at, say, a Senate commerce committee — late afternoon, she's got Amodei's essay open, a notepad, she's trying to draft what a hearing question would even look like. She's searching for the enforcement clause. It's not there. She writes 'voluntary?' in the margin and underlines it twice. That's — that's actually the whole document in one margin note.
Jonathan Ingles: Clark called it a collective action problem. Frankly, that's the most honest two seconds of the entire week — because naming it that is an admission that voluntary coordination fails. The diagnosis contradicts the prescription.
Elena Marsh: And none of this touches what's happening inside these companies — the open letter, nearly fourteen hundred workers, and at least one Anthropic researcher who resigned. The governance question gets harder when you factor that in, and we should get there.
Jonathan Ingles: Nearly fourteen hundred. That's the number — not a dozen unhappy employees, not a fringe petition. Almost fourteen hundred workers at these companies signed an open letter calling for government oversight. The very companies standing in front of Congress saying trust us. Their own staff is saying: no, actually, don't.
Elena Marsh: And one Anthropic researcher didn't just sign — resigned. Over concerns that rapid development threatened human existence. That's not a margin note. That's someone deciding that working there was — well, that the work itself was the problem.
Jonathan Ingles: Which is the credibility problem made structural. Because Anthropic's entire public position rests on: we see the risk, we're the responsible ones. And one of your own researchers is saying — actually, the internal machinery isn't working.
Elena Marsh: No named body. No enforcement. And now no internal consensus either.
Jonathan Ingles: Right — but the part that doesn't fit is: Congress hasn't asked why. Not once in any reported hearing has someone said, your own employees signed a public letter demanding external oversight — why aren't you demanding it too? That gap is a choice.
Elena Marsh: And then the warnings spill out into — I mean, picture a senator fielding calls the week of September fifteenth. One from a staffer who just watched KQED Forum run tech journalists through the Amodei essay. Another from someone who saw CBS News covering lawmakers weighing AI risks. And then someone hands them a clip of Bernie Sanders and Steve Bannon, together, Sanders saying AI has been used to create new viruses, Bannon calling it the defining issue of our time. That senator now has to decide what to do with all of that arriving at once.
Jonathan Ingles: Sanders and Bannon. Together.
Elena Marsh: On the same side of the same issue. Which is — actually, that's not reassuring. When the political coalition carrying your warning includes figures whose credibility is maximally contested on opposite ends, the warning doesn't gain bipartisan weight. It gets — absorbed into each side's existing frame.
Jonathan Ingles: The message is now owned by the messengers. And the messengers are Bannon and Sanders. So a senator hears 'catastrophic AI risk' and has to decide whether to treat it as a genuine structural warning or as evidence that this issue has become a vehicle. That's a real governance failure — not because the risk isn't real, but because the political container for it is broken.
Elena Marsh: And Altman is still at Dreamforce saying trust us. The fourteen hundred signatures didn't change the pitch. That's what actually tells you something.
Jonathan Ingles: The pitch didn't change. That's the thing. Fourteen hundred signatures, one resignation, and Altman is still at Dreamforce saying trust us anyway. Which means — actually, that sentence you opened with? 'Your fears are justified. Trust us anyway.' That's not a mistake in the messaging. That's the only sentence available to a company that can't stop and can't fully control what it's building.
Elena Marsh: Yes. And if this regulatory moment passes — and it might — and things accelerate again, and something goes wrong — they'll have no credibility left to sound the alarm a second time. They'll have spent it. On Dreamforce.
Jonathan Ingles: On a trust appeal that was never really about trust. That's where I land.
