The Background Problem (Dreyfus)

In the [YOU] on AI Field Guide

The cycle's most instructive failure—the passage about Deleuze that sounded like insight and broke under examination—is, in Dreyfus's framework, a background failure in the precise sense. The model connected Deleuze's "smooth space" to Csikszentmihalyi's flow state in an elegant, rhetorically persuasive way that was wrong in a manner invisible in the prose but obvious to anyone who had engaged with Deleuze's work. The model generated a connection based on the surface similarity of "smooth" and "flow," producing an output statistically consistent with how philosophical concepts are discussed but semantically disconnected from what the concepts mean. Its approximation of the philosophical background sufficed to produce prose that sounded philosophical, not prose that was.

The crucial detail is that the cycle's author caught the error—not through systematic verification but through a felt sense that something was not right, a nagging unease the morning after he had approved the passage. The nagging is itself a form of embodied background understanding, the holistic perception that something does not fit, registered as a bodily signal shaped by years of reading. The machine cannot nag itself; it has probability distributions, not a felt sense of rightness, and within those distributions it operates with breathtaking facility while having no way to distinguish what it knows from what it is merely pattern-matching toward. Hallucinations are the structural consequence: a system that processes patterns without inhabiting the web of involvements those patterns presuppose will generate outputs that are pattern-consistent but involvement-inconsistent.

The deepest irony, in Dreyfus's framework, is that the better the approximation becomes, the more dangerous the remaining gaps are. When classical AI's approximation was poor, the gaps were visible and the background's necessity obvious. Now that the approximation is extraordinary—contextually appropriate ninety-five or ninety-nine percent of the time—the remaining failures hide behind a surface of consistent excellence, and the temptation to trust the surface without checking grows correspondingly. A system that fails conspicuously teaches vigilance; a system that fails rarely and plausibly teaches trust, which is the precise condition under which the background's absence becomes catastrophic. The cycle's discipline of vigilance against plausible wrongness works only because the human possesses the genuine background the machine approximates.

Origin

Dreyfus identified the background, following Heidegger, as the fundamental obstacle to AI, and the classical failure on this front was dramatic and instructive. The project of symbolic AI was to formalize the background—to write down the rules of common sense, to represent in a database everything a competent adult knows about how the world works. Lenat's Cyc, begun in 1984, attempted exactly this, entering millions of assertions by hand, from "water flows downhill" to "people generally do not enjoy being hit in the face with a fish." It produced a very large database. The difference between a very large database and common sense is the difference between a map and the territory.

The philosophical core of the argument is Heidegger's concept of the totality of involvements—the web of relationships within which any particular thing shows up as what it is. A hammer shows up as a hammer not because of its physical properties but because of its place in a web that includes nails, wood, construction, shelter, dwelling, and an entire form of life. Remove the web and the hammer is just an oddly shaped object; the physical properties remain, the meaning disappears. The background is this totality rendered in the language of cognitive science: a web so vast and densely interconnected that no finite representation can capture it.

Large language models, Dreyfus's framework concedes, approach the background from a radically different direction than Cyc, and the difference requires the critique to be updated rather than merely repeated. They do not formalize the background; they absorb its textual residue, the implicit encoding left when millions of people write about restaurants and presuppose that one does not lie on the floor. The functional result is often indistinguishable from genuine common sense. But the absorption captures the statistical regularity, not the bodily disposition, the unthought assumption, the felt sense of what is appropriate—and the difference becomes visible at the edges, where the situation is novel and the standard patterns do not apply, which is precisely where common sense matters most.