Fluency Is Not What You Think It Is

There is a small experience that happens to most people occasionally and that nobody gives a name. You are listening to someone speak. They are speaking smoothly. You feel yourself agreeing, or persuaded, or impressed. And then, perhaps a few minutes later, perhaps an hour later, perhaps the next day, you try to reconstruct what they actually said, and you discover that you cannot. There is a residue of having been convinced, but the thing you were convinced of will not come back. The argument has dissolved between the time you heard it and the time you tried to recover it.

This is not always a sign that the speaker was empty. Sometimes you simply forgot. Sometimes you were tired. Sometimes the argument was real but uncongenial to your memory. But sometimes — and this is the experience that has no name — the argument was never there. There was only the surface of an argument, delivered in a register that produced the feeling of having heard one, and your mind responded to the register without checking the substance, and now the register is gone and there is nothing underneath.

The reason this happens is well understood by cognitive psychologists, even though it is not well known outside that field. It is called the fluency heuristic, and it is one of the most thoroughly replicated findings in the study of judgment. In its original formulation, due to Rolf Reber and Norbert Schwarz in the late 1990s, the finding is this: information that is easier to process is judged more truthful than information that is harder to process. The difficulty does not have to be in the content. It can be in the surface. The same statement printed in an easier-to-read font is judged more likely to be true than the same statement printed in a harder-to-read font. The same statement spoken without pauses is judged more credible than the same statement spoken with pauses. The same statement made by a speaker with a familiar accent is judged more accurate than the same statement made by a speaker with an unfamiliar accent.

This is not a moral failing in the listener. It is how cognition works. Daniel Kahneman, in Thinking, Fast and Slow, treats fluency as one of the operating principles of System 1 — the fast, automatic, unconscious mode of thought that produces our initial reactions to almost everything. System 1 has no separate truth-checking module. It uses ease of processing as a proxy for accuracy because, in the world we evolved in, the proxy worked. Information that was easy to process usually came from sources that knew what they were talking about. Smooth speech usually came from people who had thought about the subject. The heuristic was right often enough that it was worth keeping.

The heuristic is still right often enough that it is worth keeping. Most fluent speakers most of the time really are saying things that track. Most polished writing really does come from people who have something to say. The relationship between fluency and reasoning is positive. It is just much weaker, and much more situation-dependent, than the heuristic treats it as.

Here are the situations under which the heuristic fails systematically.

The first is when the speaker is rehearsed. Someone who has practiced a speech can deliver it fluently regardless of whether the underlying reasoning would survive an unanticipated question. The fluency tells you that the speech was prepared. It does not tell you that the speaker could defend the argument against an objection they had not thought through. Politicians at scripted events, executives at investor presentations, and interview candidates working from a list of likely questions are all in this category. Their fluency is real. It is also a function of preparation, not of cognitive structure, and it cannot survive contact with anything they did not anticipate.

The second is when the speaker has been socially trained to be fluent. Some populations are taught fluency from childhood. Expensive schools train it explicitly. Certain professional pipelines — law, consulting, broadcasting — select for it and then refine it. Certain class backgrounds reward it from the dinner table onward. People who have been raised in these environments produce fluent speech as a default state, regardless of whether the underlying thought is being done well or being done at all. Their fluency tracks their education and their class more reliably than it tracks their reasoning. This is one of the reasons educational background is so heavily weighted in informal evaluations of intelligence: the heuristic is reading the surface that schooling produces, and assuming the surface tells you about the underlying mind.

The third is when the speaker is operating in their first language, and the listener is unconsciously using fluency as a proxy for thought. Second-language speakers often have impeccable reasoning structure paired with halting delivery — pauses to retrieve words, false starts as they reformulate a phrase, occasional grammatical irregularities. First-language speakers can have shallow reasoning paired with effortless delivery. The fluency heuristic systematically penalizes the first and rewards the second. A reviewer or interviewer or teacher who is using fluency as evidence of thinking will overrate native speakers and underrate non-native speakers, and will not feel that they are doing this, and will be confident that their judgments track the construct they care about.

The fourth is when the speech is not produced in real time. A read-aloud script is maximally fluent, by construction. The speaker is not reasoning; they are performing. This was always true and was always a problem, but it has become a more acute problem in the last two years. The current generation of large language models can produce a script that an average reader cannot distinguish from the written work of a thoughtful person. A speaker reading that script aloud will sound exactly as fluent as a speaker delivering original thought, and nothing in the audio signal will reliably reveal the difference. The fluency heuristic, in the presence of model-generated scripts, has become a useless guide to whether the person speaking is actually thinking. This is not a technical observation. It is something the assessment of human intelligence is going to have to absorb.

The fifth is when the speaker has nothing at stake epistemically. Harry Frankfurt’s well-known distinction between lying and indifference to truth applies here. The liar knows the truth and is concealing it. The speaker who is indifferent to truth has no relationship to it at all — is neither concealing nor revealing it, simply producing speech that serves their immediate purpose. Frankfurt’s observation is that the second kind of speaker is often more fluent than the first, because the liar is constrained by the truth they are hiding and the indifferent speaker is constrained by nothing. Speech freed from the obligation to track reality is smoother than speech that is bound by it. The clearer your stake in being accurate, the more your speech struggles. The speaker who has nothing at stake epistemically can speak smoothly forever.

The implications of all this are larger than they look.

If fluency is a noisy proxy for reasoning, then evaluators who use fluency as a proxy are systematically wrong about who is thinking. They overrate the rehearsed and underrate the unprepared. They overrate the schooled and underrate the unschooled. They overrate the native speaker and underrate the second-language speaker. They overrate the indifferent speaker and underrate the careful person who hedges, who pauses, who says “wait, let me reconsider that,” who allows their speech to slow down when the question gets harder. None of these biases are conscious. None of them are corrected by the evaluator deciding to be more careful. They are baked into the heuristic itself, which runs faster than deliberate judgment and produces its verdict before the deliberate judgment is engaged.

If fluency is now decoupled from reasoning by the existence of model-generated scripts, then assessment has a new problem nobody has fully grasped yet. A polished essay submitted in an application file used to be evidence that the applicant could write well. It is no longer evidence that the applicant did anything. A fluent prepared statement in an interview used to be evidence that the speaker had thought about the topic. It is no longer evidence that the speaker has any relationship to the topic at all. The institutions that built their evaluation systems around the assumption that fluent output came from a fluent mind are about to discover that the assumption no longer holds.

And if the markers that genuinely indicate real-time reasoning are precisely the markers the fluency heuristic discounts — the pauses, the reconsiderations, the visible adjustments, the moments of stopping and starting again — then the people most likely to be underestimated by current evaluation are exactly the people whose minds are doing the most visible work. The hesitating speaker who is genuinely figuring out their position in front of you is being penalized by an evaluator who has been trained, by everything in their life up to this moment, to read hesitation as ignorance.

There is a quiet experience that is the opposite of the one this essay began with. You are in a room with a confident speaker who is producing a fluent monologue. Everyone is nodding. You begin to realize that the speaker has, after several minutes, said nothing — that the words have been arranged in the rhythms of argument without containing one. You glance around the room. Almost no one else has noticed. And then your eye lands on one other person who has noticed — usually quiet, often the youngest person in the room, sometimes someone whose own English is not perfect — and the look on their face tells you that they have been listening more carefully than anyone, and that their assessment of the speaker is calibrated by something other than the ease of the syllables.

That second person’s silence is the more reliable signal. You have to learn to look for it, because the room is not going to point you toward it.