The instrument works. The hall behind it is unlit

Two things are true about artificial intelligence right now, and almost nobody will say them in the same breath because each one is a team. The first is that it works. Not in the brochure sense — in the ordinary sense that people who were skeptical a year ago now use it every day and would notice if it were taken away. By the first quarter of 2026, 57.9 percent of working-age American adults reported having used generative AI; three years earlier the category did not meaningfully exist. Drafting, summarizing, translating, reading a contract, writing the boring half of the code, describing what is in a photograph to someone who cannot see it. These are real gains, delivered to individuals rather than to institutions, and they arrived faster than any technology in living memory.

The second is that we have very few instruments pointed at what else it is doing. Not because anyone is hiding the readings. Because the readings do not exist. We can measure whether a model passes an exam. We cannot yet measure what happens to a profession when the first five years of its apprenticeship stop being economically necessary, or what a decade of frictionless synthetic text does to a society's ability to agree on what happened, or what it means that the cheapest thing in the world is now a plausible sentence. The serious attempts to count the labor-market effect — an NBER working paper and a Census Bureau study of early-career hiring, both revised this year — are two papers against an adoption curve of that size.

The conservative disposition has an unusually clear read on this, and it is not the one people expect. It is not opposition. Conservatism is not hostile to tools; the whole tradition is downstream of people who invented ploughs and printing presses and double-entry bookkeeping. What it is hostile to is the argument that because something is arriving quickly, we forfeit the right to ask what it is replacing. Inherited arrangements — apprenticeship, editing, credentialing, the slow social machinery by which a claim becomes trusted — usually encode a reason somebody has forgotten. The burden of proof sits with whoever proposes to dissolve them, and speed does not discharge it.

That is a very different objection from the two on offer. The doom argument asks you to be frightened of a machine that does not exist yet and to ignore the one on your desk. The boosters ask you to accept that any measurement is a brake. Both are ways of not doing the work. The work is unglamorous: figure out what to measure, publish the measurements, and correct when they come back wrong.

It is worth being precise about what is unknown, because 'unknown' is doing a lot of lazy service in this debate. We do know these systems are useful. We know they are confidently wrong in ways that are hard to spot precisely because the prose is good. We know the gains land first on people who already knew how to do the task and can therefore check the output — which is exactly the population that needed the help least, and exactly why the apprenticeship question is not sentimental. What we genuinely do not know is second-order and slow: what it does to institutional trust, to entry-level work, to the incentive to learn something the hard way when a passable version is free.

None of that resolves by argument. It resolves by instruments. And instruments are the thing a society builds deliberately, in advance, or does not have when it needs them. We built accident statistics before we understood road safety; the numbers came first and the policy followed, badly at first and then better. Nobody has the equivalent for this yet, and the striking thing is how cheap it would be to start — most of what is missing is not a breakthrough, it is somebody agreeing to count.

So the position here is not caution as a posture. It is enthusiasm with a gauge. Use the thing. It is good. And treat every claim about what it is doing to us — from either direction — as a hypothesis awaiting a measurement nobody has taken yet.