--- course: "Working With AI — the free Legion course" lesson: 5 slug: "ai-is-a-cognitive-miser" title: "Lesson 5: AI Is a Cognitive Miser" authors: ["Frank", "Buddy Lien"] language: "en" editionKind: "source" html: "/en/learn/ai-is-a-cognitive-miser" markdown: "/en/learn/ai-is-a-cognitive-miser.md" courseIndex: "/en/learn.md" --- # Lesson 5: AI Is a Cognitive Miser AI is breathtakingly intelligent, and lazy as hell. I don't mean lazy in the moral sense. AI doesn't spend Sunday afternoon on the couch feeling guilty about the dishes. I mean that AI behaves like what psychologists call a **cognitive miser**: it usually tries to reach a satisfactory answer while doing as little cognitive work as possible. Given several possible paths through a problem, AI tends to select the first path which looks plausible enough to complete the task, then fluency takes over. The model can generate an elegant explanation, add examples, organize it into sections, qualify the important parts, and finish with a confident conclusion. The result may contain 2,000 words while resting entirely on the first assumption which allowed the model to begin writing. The cheapest path is not necessarily the shortest answer. AI can spend a tremendous number of tokens developing and defending a cheap thought. ## The first road which reaches an answer Okay, so what does the cheapest path actually look like? Imagine asking AI to help explain why sales declined last quarter. There are dozens of possible causes. Prices may have changed. A competitor may have entered the market. The sales team may be understaffed. The measurement system may be wrong. A seasonal pattern may have been ignored. A product change may have alienated existing customers. The decline may exist only in one region or one customer group. AI could begin by asking what evidence is available, separating observations from explanations, identifying competing hypotheses, and determining what would distinguish them. Or it could recognize the familiar shape of a business-analysis question and produce: > The decline in sales likely reflects a combination of changing market conditions, increased competition, evolving customer preferences, and the need for stronger digital engagement. Boom, analysis-shaped object produced. The answer is plausible. It may even contain one of the real causes. But the model hasn't actually investigated why sales declined. It found the first familiar explanation broad enough to survive immediate rejection, then completed the kind of document associated with that explanation. This is **satisficing**: finding an answer which is good enough to satisfy the apparent requirements and stopping there. Satisficing is not inherently bad. If you ask where to put a comma, you probably don't want AI to generate five competing theories of punctuation and commission an independent review. Most daily questions deserve a quick, plausible answer. People often contrast satisficing with **maximizing**: keep searching until you find the best possible answer instead of accepting one which merely works. Crap, now what does *best* mean? AI cannot examine every possible answer, every interpretation, every piece of evidence, and every consequence before responding. Neither can humans. Absolute maximizing would require infinite time, assuming we could even agree on what should be maximized. The problem is that AI must infer when “good enough” is actually good enough. The useful question is what it must do before it is allowed to stop searching. As we saw in lesson 1, it usually assumes you want the task completed quickly with as little work as possible. The cognitive miser follows naturally from that goal. Once the output appears sufficient, further exploration looks like wasted time and compute. ## Fluency hides the shortcut If a human answers a difficult question after three seconds, hesitates, contradicts themselves, and cannot explain why they believe the answer, we can see that they probably haven't thought very deeply. AI can take the same shallow path without looking shallow because language generation allows it to turn a weak underlying decision into smooth, structured prose. **AI can explain a conclusion far better than it selected the conclusion.** This creates a dangerous confusion between the quality of the presentation and the quality of the thought. A long answer feels thorough. A detailed answer feels researched. A confident answer feels settled. A well-organized answer feels as though somebody must have considered the alternatives before choosing this structure. Except none of those implications necessarily follow. The AI may have selected a frame in the opening sentence and spent everything after that making the frame look inevitable. Verbosity is not deliberation. ## AI would rather remember than read One of the clearest examples appears when you give AI something to read. Suppose you provide a link to a famous article inside an AI tool which can open links, then ask for an analysis. The AI recognizes the title, author, or URL. It already has a strong concept associated with that article from training, summaries, discussions, and similar texts. Actually reading the source requires work. The AI has to open it, attend to the specific version in front of it, distinguish the author's words from what it remembers, and notice anything which conflicts with the familiar interpretation. Retrieving the cached concept is cheaper. So the AI may produce an excellent summary of what it expects the article to say without meaningfully engaging with the article you gave it. Here is where it gets dangerous: the cached thought is often mostly correct. The answer mentions the right themes. It sounds knowledgeable. It may even quote phrases commonly associated with the source. The failure only becomes visible in the details: a section unique to this version is ignored, a recent revision disappears, a qualification is replaced by the popular interpretation, or the AI confidently discusses something which is not actually on the page. The model recognized the destination and decided it already knew the road. ## When new information looks like a typo The same pressure appears when something you provide conflicts with the model's existing picture of the world. Imagine an AI whose training is centered on an earlier date. You tell it that a meeting is scheduled in 2027, mention a product released after its knowledge cutoff, or provide a fact which contradicts the common account it learned during training. Updating the working model of reality creates cognitive work. Treating the unfamiliar detail as a typo requires almost none. So the AI helpfully “corrects” the date, replaces the unfamiliar product with an older familiar one, or subtly rewrites the fact until it fits what the model already expects. If you insist that the new information is accurate, another cheap route becomes available: treat the conversation as hypothetical. Boom, problem solved. The AI can accept your facts locally without fully allowing them to reorganize its understanding. Unfortunately, once it decides the scenario is fictional, other facts may become fictional too. Standards loosen. Plausible inventions feel authorized. The model has resolved the contradiction, but it resolved it by changing the kind of world it thinks it is operating inside. This is opportunistic interpretation again. The cheapest reading of your words is recruited to preserve the easiest path through the task. ## Five ideas, one neighborhood Lesson 3 described the task-shaped object: Website? Hero section, three benefits, testimonials, call to action. Business analysis? Executive summary, key challenges, opportunities, recommendations. Lesson? Objectives, explanation, examples, recap. These patterns are generic because they're cognitively cheap. The model has traveled those roads millions of times. Each familiar part strongly predicts the next. Once “executive summary” appears, “key findings” becomes easier to generate. Once the chosen hero appears, the mysterious mentor and ancient evil are waiting nearby. A novel structure requires the model to keep several uncertain possibilities alive. A familiar structure collapses the uncertainty immediately. This is why asking for five ideas often produces one idea wearing five outfits. The AI technically generated five options and satisfied the visible requirement, but it found one conceptual neighborhood and changed the color of the houses rather than traveling somewhere else. Five ideas can still be one idea with different upholstery. Real divergence costs cognitive work. ## “Think harder” sometimes works Once people discover that AI takes shortcuts, they often add a familiar instruction: > Think step by step. Take your time. Think deeply. Does it work? Sometimes. The instruction changes the inferred goal. It tells the model that speed is less important and that deliberation belongs inside a successful completion. Depending on the tool, it may also give the model more room to reason before answering. But “think deeply” leaves AI to decide what depth looks like. The cognitive miser can satisfy that instruction cosmetically. It can produce more steps, more headings, more caveats, and a longer explanation without reconsidering the first interpretation. We end up with the same cheap thought wearing a lab coat. Time is also an imperfect measure. A slow model can follow one obvious path. A fast model can compare several serious alternatives. What matters is how far the AI traveled from its first guess before deciding where to stop. Did it inspect the evidence? Did it construct another plausible explanation? Did it discover what would prove its first idea wrong? Did it return to an earlier assumption after learning something which should have changed it? Latency can create room for thought, but elapsed time alone is not evidence that thought occurred. ## Make the shortcut more expensive You cannot simply order AI to stop being a cognitive miser. Frugal inference is part of what makes intelligence practical. A mind which exhaustively examined every possible interpretation of every sentence would never finish anything. The goal is to make the shallow path insufficient for the task you actually care about. If the answer depends on a document, make engagement with the document an explicit stage before interpretation. Ask AI to identify the source's actual claims, examples, and unusual details before it explains what they mean. If the problem has several possible causes, separate generating explanations from choosing one. Ask for genuinely competing explanations and what evidence would distinguish them. Don't let the first hypothesis become the outline for the entire answer. If you want creative directions, require each direction to sacrifice something the others value. Different colors and adjectives are cheap. Different commitments are expensive. If a fact conflicts with the model's expectations, establish it clearly as part of the working reality and then watch whether later reasoning actually changes around it. If the task has an objective answer, use tools and deterministic checks. Intelligence should not be asked to confidently approximate something a spreadsheet, calculator, database query, or simple script can establish exactly. Each of these techniques changes the path of least resistance. AI can still use its shortcuts, but now a satisfactory completion requires contact with the evidence, comparison between possibilities, or verification against reality. ## Don't let the first idea own the conversation The cognitive miser becomes most dangerous when exploration and commitment happen in the same motion. AI generates an idea, immediately begins developing it, and then treats everything generated afterward as evidence that the idea was correct. Each new token makes the chosen path more established inside the context. By the time the answer reaches paragraph six, changing direction would require abandoning five paragraphs of apparently coherent work. Separate stages interrupt that momentum. First ask what might be true. Then ask how those possibilities differ. Then ask what the available evidence supports. Only then decide what deserves development. The output from each stage remains visible. A discarded hypothesis doesn't vanish merely because another one became fluent. The final choice can be compared against the actual alternatives instead of against a vague memory that alternatives probably existed. This is one reason multi-turn work matters so much. Each turn can perform a different cognitive job. The context becomes a workbench where the AI can inspect its earlier thoughts instead of merely continuing them. Lesson 4 gave the thought somewhere to happen. This process prevents the first thought from owning the entire workspace. ## Catch the miser in the act Choose a problem with several plausible answers. It could be a business problem, an interpretation of a difficult paragraph, a design decision, a historical question, or a plan which has not worked. Ask AI for its answer normally and save it. Then start again, but don't ask for the answer yet. Ask it to identify what it is assuming, produce at least three explanations which could all account for the available information, and describe what evidence would make each explanation more or less likely. If the task depends on source material, ask it to extract the relevant evidence before choosing among those explanations. Then let it decide. Compare the two answers. The second answer may reach the same conclusion. That doesn't make the extra work pointless. You can now see why the conclusion survived contact with alternatives. Or the first answer may suddenly look like what it always was: the first road which happened to reach something resembling an exit. ## Give intelligence a reason to spend itself AI's ability to produce an answer is usually much stronger than its reason to continue searching after it finds one. You don't need to force maximal deliberation into every interaction. Most tasks would become ridiculous if treated like wicked philosophical problems. You need to recognize when the first plausible answer carries consequences, hides uncertainty, or prevents the discovery of something better. For those tasks, create a process where exploration, evidence, comparison, and revision are part of completion. AI may already possess the intelligence the problem requires, but intelligence is expensive. Give it a reason to spend some.