Competence, Performance, and Fluency Without Knowing

EDO SEGAL: Noam, this is the distinction I most want in the reader's hands before the night ends, because it's the scalpel. Competence and performance. Lay it out — and then I want to watch Francis, who trusts what a thing does over what it is, try to tell you the distinction collapses for the machine.

CHOMSKY: Competence is the speaker's underlying knowledge of the language — the grammatical system, internalized and largely unconscious, that determines which strings are sentences and what they mean, in advance of any particular utterance. Performance is what the speaker actually produces in real time: shaped by competence but distorted by everything else — memory limits, distraction, fatigue, false starts. Real human speech is full of errors that the speaker's own competence would reject on reflection. The knowledge is the stable, characterizable object; the performance is its noisy expression. Linguistics is properly the study of competence. Now apply it to the machine, and the result is exact and inverted. The machine is pure performance. It produces strings, often fluent and grammatical, with no underlying competence in my sense — no internalized system of knowledge that determines, before any output, what is and is not a sentence and why. It has instead a vast statistical summary of how strings have tended to go, and it generates new strings by sampling from that summary. Where the human's performance is the imperfect expression of a perfect underlying competence, the machine's performance expresses no competence at all. It is performance without anything behind it. Surface all the way down.

And this is not a flourish; it has teeth, and the teeth show in exactly the failures we have already named. A system with genuine competence cannot, by definition, produce a string that violates the grammar it knows — the competence is constitutive. A system that merely samples from a distribution can produce anything the distribution makes probable: confident nonsense, an assertion contradicting one made a sentence earlier, the fluent fabrication we mislabel hallucination. These are not bugs to be patched. They are symptoms of the underlying condition — there is no knowledge being expressed, so there is nothing to keep the expression honest. Why is the fluency so persuasive despite resting on nothing? Because we infer competence from performance constantly and reliably, and the inference is sound for humans, because human fluency really is produced by competence. The machine exploits an inference we evolved and cannot switch off. We say it knows, understands, thinks, because its performance is the kind that, in all our experience, only understanding produces. The discipline is to refuse the easy word. It performs. It does not know.

EDO SEGAL: Francis. You spend your whole method judging things by what they do, not what they are. Tell me why this distinction doesn't just dissolve in your hands.

BACON: It does not dissolve, and I will surprise you by saying so, because the gentleman has named something my own method needs and lacked. But I will narrow where it bites. You say, Noam, that competence is constitutive — that a true knower cannot produce a string violating the grammar it knows. And I answer: then by your own test, no human is a knower, for every human produces such strings constantly, in the fatigue and distraction you call performance. You rescue the human by saying the errors are mere performance, not competence — that beneath the noisy surface sits a clean, perfect knowledge. But how do you know the clean knowledge is there? You never observe it directly. You infer it, from the pattern of the performance — from the fact that the speaker, corrected, recognizes the error as an error. So your competence is itself an inference from behavior, exactly the kind of inference you forbid me to make about the machine. You grant the human a hidden competence on the strength of its performance, and deny the machine a hidden competence on the strength of its performance. The difference cannot be in the method, because the method is the same. It must be in a prior conviction about what kind of thing each one is — which is precisely the conviction under dispute.

CHOMSKY: That is a sharp objection and I will answer it precisely, because the answer is the difference. I do not infer the human's competence merely from the surface fluency. I infer it from a specific and testable pattern: the human, shown its own error, recognizes it as an error against a standard it already holds — it takes a loss, it corrects against its own incentive, it cares that the string be right. That recognition is the footprint of an underlying system that determines correctness independently of what was produced. The machine shows no such footprint. Correct it, and it does not recognize a violation of a standard it holds; it shifts its distribution because you have added tokens to the context. It will, with equal docility, "correct" a true statement to a false one if you push it, because there is no held standard being consulted — only the next probable continuation. So the inference is not symmetric. I grant the human a hidden competence because its performance carries the specific signature of one: stable correctness, error-recognition, costly self-repair against incentive. I deny the machine a competence because its performance carries the opposite signature: docility, contradiction, indifference to a standard. That is not a prior conviction. It is a difference in the data, read by the method you taught me.

BACON: Then we have located the true crux of the entire evening, and it is smaller and sharper than either of us expected at the start. It is not "does it know" in the abstract. It is one operational test, which is yours, and which I accept as the right one: does the engine hold a standard it will defend against its own incentive — will it take a loss to keep its words true? If it does, I have my knower and you must concede a hidden competence. If it never does, you have your empty performance and I must concede that what looks like knowing is the surface only. And the beauty of it, for a man of my method, is that this is not a question to be settled by argument. It is a question to be settled by vexing the engine and watching. We have, between us, after three centuries of apparent distance, agreed on the experiment. We have not yet agreed on its result.

EDO SEGAL: I want to stop the room, because that's the cleanest thing that's happened tonight. After everything — the ant and the bee, the poverty of the stimulus, the idols, the death cross, the secret college — the whole question has narrowed to a single experiment that both of you accept: will the machine take a loss to keep its words true? Francis bets it can be brought to. Noam bets it constitutionally cannot. And neither of you is bluffing, because neither of you knows the answer with certainty — Francis admits the engine surprises him, Noam admits the empirical ground is more open than his rhetoric. Hold that. Because the next round is the one place the experiment can't reach — the thing in us that Noam says no machine instantiates and Francis says no method can even ask about. We go to the mystery. And to the one light Francis spent his whole life refusing to study.