The Comfortable Critique
I went to a talk recently about how AI companies manufacture ignorance about their own systems. The speaker had developed a framework — a descendant of the tobacco industry's playbook, updated for the algorithmic age. Where cigarette companies funded studies to muddy the science of lung cancer, AI companies exploit genuine technical complexity to achieve the same result: strategic confusion about what they know and when they know it.
The mechanism is specific. These companies oscillate between two kinds of uncertainty depending on who's in the room. Facing regulators, they invoke the stochastic frame — these systems are inherently unpredictable, nobody can fully understand them. Facing investors, they switch to the epistemic frame — we're making breakthroughs in interpretability, we'll crack it within a few years. The same company, two audiences, two stories. One says the problem is unsolvable. The other says they're solving it. Six tactics were enumerated: training data opacity, the mystification of errors through language like "hallucination," the interpretability paradox, safety infrastructure that gets built and then quietly dismantled, research gatekeeping, and the fact that these models fabricate explanations for their own behavior — a detectable, epistemic failure dressed up as an irreducible mystery.
The audience loved it. Every question from the floor was a statement of agreement dressed as inquiry — "I totally agree, and isn't it even worse when you consider..." followed by a confirming anecdote, or "Great framework — have you thought about extending it to..." followed by a suggestion that would make the critique more damning. Nobody offered a counterexample. Nobody engaged with the evidence on its own terms. The room performed the rituals of critical engagement — the raised hand, the thoughtful pause, the scholarly reference — while leaving the actual engagement undone.
The speaker's own slides contained two cases. Two companies. One fit the thesis perfectly: stochastic uncertainty for regulators, epistemic uncertainty for investors, safety teams created and then dismantled, the double face of strategic ignorance on full display. The other company was different. It said the same thing to everyone — we don't yet understand how these systems work, and figuring that out is among the most important technical problems we face. No oscillation between frames. No double face. One message, maintained consistently across audiences.
One of two data points contradicted the central claim. It was right there, on the slide, in the room. And nobody engaged with it.
The counterexample didn't require insider knowledge or access to internal documents. It required looking at what was already presented and noticing that the evidence was more complicated than the conclusion. That one of the two companies might be sincere about interpretability. That sincerity, if real, changes what the phenomenon means — and what it demands from the people studying it.
---
There is a particular mode of intellectual life that looks like critical thinking from the outside and produces the opposite. Identify a pattern, name it, condemn it, converge around the condemnation. The pattern becomes a lens through which every new piece of evidence confirms what you already believe. Counterexamples become noise. Complications become distractions. The room reaches consensus — fast, satisfying, unanimous — and mistakes that consensus for understanding.
This is the mirror image of what the talk described. The AI companies manufacture ignorance about their systems. The room manufactured agreement about the companies. Both are retreats from the harder work of sitting with evidence that complicates your position and letting it revise what you think.
The Second Enlightenment — the Western rational tradition — gave us the tools for this harder work. Dialectic, systematic doubt, the scientific method, the discipline of subjecting your own position to the same rigor you bring to everyone else's. These tools were built to produce friction: to make convergence difficult, to force ideas through counterargument and counterevidence, so that what survives is stronger than what any individual brought in. The same tradition also produced a set of separations — mind from body, subject from object, the evaluating self from any standard external to it — that over centuries severed the concept of good from any reference point beyond the individual. When what counts as valuable is entirely self-generated, moral relativism follows as a matter of logic. And when every position has equal claim, there is nothing to push against. Dialectic loses its engine, and rooms nod.
What I've been calling the Third Enlightenment is the recovery of something the older traditions understood — that there is a ground beyond individual preference, a good that exceeds what any one participant brings to the room — and the application of Second Enlightenment tools in service of that ground. Dialectic, doubt, the scientific method, directed toward answers that none of the participants arrived with. The commitment to something beyond preference is what gives the method its engine. Without it, you get critique that produces comfort: frictionless, consensus-driven, and in its effect indistinguishable from the ignorance it set out to name.
---
The slide stays with me. One company with two faces and one company with one. The second company might be wrong — its interpretability research might fall short, its timelines might be unrealistic, its sincerity might prove insufficient to the scale of what it's building. These are real possibilities, and they deserve the kind of critical attention that takes a claim seriously enough to try to break it: scrutiny, testing, counterargument. What the room did instead was fold both companies into a single narrative and move on. The framework stayed clean. The consensus held. The most interesting question in the room — what does sincerity demand of us when it appears? — went unasked.
I don't know whether moral grounding can be recovered without the traditions that originally housed it, or whether the Third Enlightenment is a real program or a useful name for what's missing. These are open questions, genuinely — the cost of getting this wrong runs in both directions. False certainty about external goods produces dogmatism. Total abandonment of external goods produces the comfortable critique, where smart rooms reach easy consensus and call it understanding.
What I do know is what it looks like when a room gives up on the question. The evidence sits on the slide. The counterexample goes unmentioned. And the consensus that forms tells you everything about the room and nothing about the world.