On AI

The most interesting questions about AI aren't technical — they're about us. What we'll trust, what we'll automate, and what its confident wrongness reveals about our own.

Fluency is not truth

The defining feature of modern AI — and its defining danger — is that it produces fluent, confident, plausible output regardless of accuracy. We are evolved to trust fluency; a confident, articulate answer feels true. That instinct, useful for millennia, is now a vulnerability. Learning to separate ‘sounds right’ from ‘is right’ may be the essential literacy of the age — and it turns out we were never very good at it, even with each other.

Neither savior nor apocalypse

The public conversation about AI oscillates between utopia and extinction, and both extremes are, conveniently, forms of hype — they make the technology sound more total than it is. The more honest and more useful register is the mundane one: this is a powerful, flawed tool that will reshape specific tasks, industries and habits in specific, mixed ways. The interesting work is in the specifics, not the spectacle.

What we choose to automate

Every automation is a values statement about what we consider drudgery and what we consider precious. It's worth noticing that we often build machines to take over the creative, expressive parts of work while keeping the administrative parts — and asking whether we've got that exactly backwards. The technology doesn't decide this. We do, mostly without noticing we're deciding. The consequences for actual jobs and actual days at work are large enough to deserve their own essay — see On AI and work.

Delegated judgment

The quieter shift isn't that AI writes things for us; it's that we've begun handing it small judgments — which email matters, which résumé is worth reading, which route is fastest — and stopped noticing we've handed anything over at all. A delegated judgment doesn't disappear; it just moves somewhere less visible, governed by criteria you didn't set and usually can't see. That's less a story about intelligence than about where authority quietly relocates. The same mechanism, viewed from the feed rather than the assistant, is the subject of On algorithms.

Trust, but verify what, exactly

“Trust but verify” is good advice that's nearly impossible to follow with a system whose reasoning you can't see and whose training you can't audit. The honest version of AI literacy isn't a checklist; it's a standing posture — treat fluent confidence as a prompt to check, not a substitute for checking, especially in the moments it's most convenient to simply believe. The same asymmetry of visibility, applied to what companies know about you rather than what a model tells you, is the subject of On privacy.