Try to explain how you ride a bicycle.
Not the mechanics of a bicycle. How you do it. The actual process of maintaining balance while in motion. What you do with your body when you start to tip. How you initiate a turn. What happens in your hands, your hips, your gaze.
You can’t. Not fully. Not in a way that would let someone who has never ridden actually ride from your description alone. The knowledge is real, you have it, you use it constantly, but it resists complete articulation.
Michael Polanyi called this tacit knowledge and summarized it in one of the most important sentences in 20th century philosophy of mind, we know more than we can tell.
Winograd and Flores take this further. It isn’t just that some knowledge is hard to verbalize. It’s that the background from which all explicit knowledge emerges is structurally unarticulable. Not difficult to express. Not expressible in principle. Because the moment you articulate part of it, you’re already standing in a deeper background that made that articulation possible.
The background has no bottom.
This is the move that collapses the knowledge representation project entirely.
The dream of classical AI, and much of cognitive science, was that you could take expert knowledge, interview the experts, extract the rules they were using, encode those rules in a system, and reproduce the expertise. This was the foundation of expert systems. Get the knowledge out of the head and into the machine.
It failed. Repeatedly. Consistently. At the edges of every domain, the system would encounter a situation the rules hadn’t anticipated, and it would fail in ways no human expert would. Because the human expert wasn’t actually using rules. They were operating from a background that the rules only partially described, and that background was doing most of the work.
The rules were the visible tip. The background was the iceberg.
And here’s what makes this structurally irreducible rather than just practically difficult: to articulate any piece of background knowledge, you need to already be standing somewhere. You need a context from which the articulation makes sense. That context is itself background. You cannot get outside it to formalize it because you’re always already inside something.
Heidegger called this thrownness. You didn’t choose your starting position. You were thrown into a language, a culture, a body, a history, a set of concerns. That thrownness shapes every perception, every interpretation, every judgment you make, and almost none of it is available for inspection.
This is why genuine expertise takes years even when the explicit content could be transmitted in weeks. The explicit content isn’t the expertise. The background shaped by accumulated experience is the expertise. And background only develops one way, through time, through practice, through breakdown, through being a structural system that has maintained coherence across thousands of perturbations.
LLMs were trained on the outputs of human background. Every book, every conversation, every piece of writing that made it into the training data was produced by humans operating from backgrounds they couldn’t fully articulate. The model absorbed the surface. The texture of how background-informed minds express themselves.
And it’s remarkable how much that gets you. Genuinely remarkable.
But the background itself didn’t transfer. Couldn’t transfer. Because background isn’t content, it’s the structural condition that makes content meaningful. You can’t train on it because it was never written down. It was never written down because it couldn’t be.
So when you use AI as a thinking partner, and it’s a genuinely useful thinking partner, you’re working with something that has absorbed the shape of human knowing without the ground that knowing stands on.
Which means it’s most reliable exactly where background matters least. Explicit, formal, well-defined problems. And least reliable exactly where background matters most. Novel situations, genuine complexity, the edges where the rules run out.
Which is precisely where you need it most.
The background you can’t see is also the background you can’t give away.
@ Lee Greene