Rudolf Carnap

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what the machines actually are—not as a marketing claim or a philosophical puzzle but as a practical matter of deciding what to trust them with and what to demand they explain. Carnap is the cycle's most precise instrument for the question of what it means to formalize knowledge, and his career is the most honest record we have of where that formalization reaches its limit. The debates about whether machines “really understand,” whether their outputs are “genuinely” about anything, whether they “truly” reason rather than simulate reasoning—Carnap's distinction between internal and external questions dissolves the half of that debate that is genuine confusion while sharpening the half that is a real problem.

His internal/external distinction is the most useful single tool he offers the present moment. An internal question is asked inside a framework, answerable by the framework's own rules. An external question is asked about the framework as a whole—whether to adopt it—and has no factual answer, only a practical one. “Does the model represent the concept of justice?” asked internally, within a representational framework, has answerable content: does it track justice-related inputs, does its behavior organize around the concept? Asked externally—“does it really represent justice, in some deep framework-independent sense?”—it has no factual answer. The intractable debates that consume AI discourse are almost all external questions masquerading as internal ones, and Carnap's discipline is a standing instruction to stop running them and redirect energy to the answerable.

His concept of explication—replacing a familiar but vague concept with one that is precise and fit for systematic use—is the most honest account of what AI engineers actually do when they build systems. To train a recommendation system to maximize “engagement” is to explicate “what users value” as engagement. The system then pursues the explicatum with flawless literalness, discovering that outrage and compulsion engage better than satisfaction, driving the explicatum to its limit while the explicandum—what users actually value—is left behind. The alignment problem is the problem of the gap between explicandum and explicatum, automated and optimized at scale. Carnap's four requirements for a good explication—similarity, exactness, fruitfulness, simplicity—are precisely the criteria by which to judge whether a machine has been given the right target.

And his principle of tolerance is the framework's mature gift: the recognition that a language model's internal ontology, its strange alien way of carving up the world through high-dimensional representations, is not wrong because it fails to match our concepts. It is a different framework, chosen by its training for a different purpose than the one our concepts serve. The right question is not which framework is the metaphysically correct one but which is fruitful for which purpose, with which costs. Tolerance liberates us from the unanswerable external question of which kind of mind is the “real” kind—and delivers us to the harder, answerable practical question of what each kind can do and what each kind costs.

Origin

Rudolf Carnap was born in Ronsdorf, Germany in 1891, studied under Frege—the founder of modern logic—and joined the Vienna Circle under Moritz Schlick in 1926. The Circle was the most consequential philosophical movement of the twentieth century in terms of its influence on science, mathematics, and eventually computing: it held that philosophy's task was not to discover metaphysical truths but to clarify the logic of scientific discourse, to expose pseudo-questions, and to build frameworks adequate to knowledge. Carnap was its most rigorous and most productive member.

Der logische Aufbau der Welt (1928) attempted the ultimate formalist program: to construct every concept a person possesses from a single kind of primitive element—the whole momentary slice of experience—and a single primitive relation—recollection of similarity—using nothing but logic. The ambition was to show that knowledge has a logical structure and that structure can be made fully explicit. The Logical Syntax of Language (1934) introduced his principle of tolerance. Fleeing the Nazi threat, he emigrated to the United States in 1935 and held positions at Chicago, Princeton, and UCLA, where he built the formal theory of inductive logic in Logical Foundations of Probability (1950).

His response to the objections that defeated his central theses is the most admirable thing about him. He did not defend the verifiability criterion to the death; he weakened it, repeatedly, in public, trying to find a version that would work. He absorbed Quine's demolition of the analytic-synthetic distinction with a grace that astonished observers, acknowledging that within a formally specified artificial language the distinction could still be drawn by stipulation, while conceding that this was a narrower victory than he had claimed. He revised because he cared more about getting it right than about having been right. That disposition—the willingness to treat the refutation of one's life's work as a contribution to knowledge rather than an attack to be repelled—is the part of him most worth inheriting.

Key Ideas

The Aufbau and the formalist dream. Carnap's 1928 program is the philosophical ancestor of symbolic AI: represent knowledge in a formal symbolic language, define concepts precisely, manipulate the symbols according to logical rules, derive conclusions by computation over an explicit structure. The expert systems, the semantic networks, the ontologies that occupied classical AI were variations on this program—all attempts to do, in silicon and for a particular domain, what Carnap proposed to do in logic and for the whole of knowledge. When they worked, they worked because the domain was small enough to be fully formalized. When they failed, they failed for the reason Carnap's own program exposed: the construction depends on primitives that cannot themselves be constructed, on a base that must touch the world—and touching the world is not a logical operation. This is the symbol-grounding problem.

Explication. Taking a concept that is familiar but vague—one we use confidently in ordinary life but cannot define—and replacing it with a concept that is precise, exact, and fit for systematic use. This is the central operation of AI engineering, performed on every concept we want to automate. The danger is the gap between the explicandum (what we meant) and the explicatum (what we made precise): the machine pursues the explicatum with a literalness no human would, and the explicatum, pursued to its limit, often betrays the explicandum it was meant to honor. Human values are the worst explicanda of all, because their whole function may depend on their remaining open and contestable—and precision, in those cases, is not an improvement but a destruction.

Verificationism and its collapse. The verifiability criterion of meaning held that a statement has cognitive content only if some observation could confirm or refute it. The criterion was devastatingly simple and devastatingly wrong: it could not survive being applied to itself, and in its strict form it excluded the universal laws of science. Carnap's principled retreat—from strict verifiability to confirmability to testability to something so qualified that critics asked whether any criterion remained—is a parable for the AI discourse. The confident deflationary claim that questions of machine understanding are meaningless—that understanding just is the disposition to produce the right outputs—repeats verificationism's error: declaring unanswerable the question that most needs answering.

Inductive logic, ancestor of statistical learning. Carnap spent the last two decades of his career trying to formalize induction: to calculate, from evidence and hypothesis alone, exactly how much the evidence confirmed the hypothesis, expressed as a number. He wanted induction to be as rule-governed as deduction. This is, recognizably, a theory of probability as rational degree of belief—Bayesian inference—and it is the philosophical foundation of machine learning. A neural network adjusts the “confirmation” of parameter settings in proportion to how well they predict the training data. Carnap was not on the side of logic against probability; he was trying to build the logic of probability, to turn the updating of belief under evidence into something formal. The field that supposedly refuted him was executing a program he had helped write.

The principle of tolerance and internal/external questions. In 1934 Carnap declared: “In logic, there are no morals. Everyone is at liberty to build up his own logic as he wishes.” There is no one correct framework; there are only frameworks, chosen for their fruitfulness toward our purposes. His related distinction between internal questions (answerable within a framework by its rules) and external questions (about whether to adopt the framework, with no factual answer) dissolves the half of every AI debate that is genuine philosophical confusion—and sharpens the half that is a real problem. Whether machines “really” understand is an external question. Whether a specific interpretability technique reveals a stable internal structure—that is internal, answerable, worth years of work.