Should We Let AI Vendors Shift The Responsibility to Their Users
This weekend, the CEOs of Anthropic, OpenAI, Meta, and xAI all issued public statements calling for AI development to slow down. In this episode, Brian Silverman, Campbell Robertson, and Michael Muhlfelder push back on the framing itself, arguing that the "slow down" conversation is obscuring a more urgent and more solvable problem: governance.
The core argument: Models don't act until prompted. The risk isn't a rogue AI deciding to do something dangerous, it's the absence of deterministic controls around how these systems get deployed, tested, and held accountable. That's a human governance problem, not a model alignment problem, and it's one businesses can address today without waiting on Washington.
In this episode:
The "frontier lab" framing as marketing: Campbell argues that this weekend's coordinated safety statements function more as competitive signaling than genuine disclosure, pointing to timing gaps between when incidents happen and when they're publicly acknowledged.
Who's grading the test? A look at Anthropic's self-reported "85% deception gap closure" and what it means when AI companies are the ones scoring their own safety metrics.
The JPMorgan question: When a company deploys a frontier model and something goes wrong, who's actually liable: the vendor or the deployer? The group works through why compliance obligations land on the deploying organization regardless of what the AI vendor claims about its own safety testing — and why that distinction matters for every business using these tools today.
The pacing pact as a moat: Roughly 1,400 AI researchers have signed onto a coordinated "slow down" pact. Campbell makes the case that this looks less like a safety measure and more like a way for incumbent labs to lock out new entrants.
The Fahrenheit 451 problem: Mike questions whether it makes sense to ask the same labs building these systems to also police them, and whether Congress has the technical literacy to regulate any of this effectively.
Where we land : Brian argues for deterministic governance and middleware as the practical answer, and makes the case that organizations building real governance now will be insulated from worst-case outcomes regardless of how the regulatory or "pacing" debate plays out.
Why it matters: If you're an enterprise leader deploying AI today, this episode is a case against waiting for regulatory clarity before building your own governance framework, and a warning against mistaking a vendor's safety claims for your own compliance coverage.
🎙️ Hosted by Brian Silverman, Campbell Robertson, and Michael Muhlfelder — Three Takes on AI
Comments