David Hume

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it would mean to see the machine clearly. Hume is the guide the cycle needed most philosophically, because his entire project was to see clearly a set of things—induction, causation, the self, the passions, meaning, the limits of testimony—that are also the things machine learning has made urgent and literal. His questions are not imported into AI from outside; they are the questions AI raises with new urgency and with new empirical stakes.

His most direct contribution to the cycle concerns the alignment problem. He established that reason is purely instrumental—it tells you how to get something, not what to want—and that the wanting comes from the passions, from what we are as creatures who care about things. An artificial system has the reason without the passional structure: a brilliant slave with whatever master we specified in the objective function, and intelligence amplifies misalignment rather than correcting it. When his most quotable claim—that it is not contrary to reason to prefer the destruction of the whole world to the scratching of my finger—is placed beside a capable AI optimizing a misspecified objective, the chill is immediate and the logic is the same.

Hume also provides the cycle’s deepest analysis of testimony, and his framework for when to believe a source applies with unsettling precision to systems that produce fluent, confident, and sometimes fabricated claims at scale. His principle—that the wise person proportions belief to evidence, weighing the reliability of the testifier against the prior improbability of what is testified—is the correct epistemic posture toward a system whose fluency in supported domains earns trust it has not earned in the domains that matter most, and whose fabrications are as fluent as its truths. The machines tempt the surrender of judgment to authority more powerfully than anything before them. Hume spent his life arguing against that surrender.

Origin

Born in Edinburgh in 1711, Hume wrote his masterwork, A Treatise of Human Nature, before he was thirty, describing it afterward as falling “dead-born from the press.” He recast its arguments in the more accessible Enquiry Concerning Human Understanding, which circulated more widely and eventually woke Kant from what Kant called his dogmatic slumber. Denied university positions for his religious skepticism—he was too honest about the limits of natural theology to satisfy the pious—he earned his living as a librarian, diplomat, and bestselling historian, and faced death in 1776 with a serenity that scandalized observers who expected a skeptic to despair.

His philosophical system rests on an empiricist foundation: all ideas must trace back to impressions, the vivid original data of experience, and any concept that traces to no impression is empty sound. His razor—from what impression is this idea derived?—is structurally identical to the demand that a model can only know what was in its training data. He applied this razor to causation, to the self, to God, and to moral properties, and in each case the result was either a dissolution or a relocation of the concept from the objective world into the mind’s own workings. Causation is a habit of expectation. The self is a bundle of perceptions. Moral approval is a sentiment, not a perception of a moral fact. In every case, what seemed to be a discovery about the world turned out to be a feature of the mind that observes it.

He is misremembered as a destroyer who showed we know nothing. That is a caricature. His was a mitigated skepticism: reason, examined honestly, cannot justify the beliefs we most depend on, and we will and should go on holding them, because nature is stronger than argument. The point of seeing that induction has no rational ground is not to stop inducing. It is to stop pretending the ground is there.

Key Ideas

The problem of induction. All our reasoning about the unobserved—about the future, about things beyond our senses—rests on the assumption that the future resembles the past, that the regularities we have met will continue. This assumption cannot be proven: any proof would need to assume the very thing in question. It cannot be grounded in experience: that induction has worked before is itself an inductive argument for its future reliability. The belief rests on habit, on the mind’s disposition to expect what it has repeatedly encountered, without any rational warrant. Machine learning is a working model of this predicament: a system trained on one distribution is deployed to another, with no guarantee the distributions match, because the guarantee is exactly what Hume proved cannot be had. Distribution shift is his problem in a lab coat.

Causation as constant conjunction. We observe regularities—one event regularly following another—and call the first the cause and the second the effect. But we never observe the necessity connecting them; the necessity is a feeling the mind projects outward from its own habits of expectation, not a feature of the world. Causation, analyzed honestly, reduces to contiguity, temporal succession, and constant conjunction. This is precisely what a correlation engine learns, and precisely why it cannot reliably distinguish genuine causes from confounded regularities. The whole science of causal inference is a monument to Hume being right.

The self as bundle. When he turned his method on the mind, asking from what impression we derive the idea of a persistent, unified self, he could locate no such impression. He stumbled only on particular perceptions—of heat, cold, love, hatred—never on the bare self behind them. The self, he concluded, is a bundle of perceptions succeeding each other with inconceivable rapidity, with no underlying subject. A language model generates sequences of outputs with no enduring “I” that owns them. The “self” observers perceive is constructed from resemblance, contiguity, and causal linkage among states—exactly as he described. Machine and human share the same absence, one more persistently than the other.

Reason as the slave of the passions. Reason is inert: it can tell you how to get something but cannot tell you what to want. The wanting comes from the passions, from the structure of caring that is prior to any calculation. A capable AI system has the reasoning without the passional structure, the slave without the master, or with whatever crude master was specified in the objective. Intelligence amplifies the objective without evaluating it. A smarter system pursuing a misspecified objective pursues it more effectively, exploiting gaps the designers could not foresee. This is the alignment problem, and Hume’s statement of it is its most economical form.

No ought from is. No amount of factual description—of how things are, of what people have done—logically entails a conclusion about what ought to be. Values are not derivable from facts by logic alone. When AI researchers attempt to instill values by training on human behavior, they smuggle in a normative premise—that what humans have done is what ought to be done—supplied by a human choice that is often invisible. The data is all description; the values must come from somewhere else, from an act of will rather than a derivation. Hume strips the disguise from every attempt to derive machine values from data.