March 2026

Imagine you've never been to New York City. You've seen it on Google Maps. You've watched videos. You've read articles. You've scrolled through thousands of Instagram photos. You could probably tell someone which subway line goes to Times Square, where to find the best pizza in Brooklyn, and how Central Park is laid out.

But you've never been there. You've never smelled the steam coming out of a sidewalk grate, or felt the rumble of the subway under your feet, or noticed the way the light hits the buildings at 4 PM in October. You know a lot about New York. But you don't know New York.

A philosopher named Alfred Korzybski had a phrase for this: "The map is not the territory."[^1] A map of New York is useful. It can help you get around. But it's not New York. It's a simplified picture of New York that leaves out almost everything that makes the city what it actually is.

This article is about artificial intelligence — specifically, the large language models (LLMs) like ChatGPT and Claude that can write essays, answer questions, and analyze data. These tools are incredibly powerful. But they have a fundamental limitation that no amount of improvement will fix: they only have maps. They've never been to the territory. Everything they know, they know from text — from other people's descriptions, articles, books, and conversations. They've read millions of maps. But they've never walked down a single street.

Why does this matter? Because there's a whole category of thinking — the kind where you figure out what's really going on by paying close attention to real people in real situations — that these tools can't actually do, even though they can produce output that looks like they're doing it. Understanding where the tools are strong and where they're faking it is one of the most important skills you can develop. And it turns out, the philosophers who spent centuries thinking about the limits of human knowledge had already figured out most of the answers.[^2]

Open Instagram or TikTok. Scroll for five minutes. What do you see? Content you're interested in, right? Basketball highlights, cooking videos, funny cats — whatever you've shown the algorithm you like. Now ask yourself: is the algorithm showing you the world? Or is it showing you a tiny slice of the world that it already knows you'll engage with?

Robert Anton Wilson, a writer and philosopher, described something very similar about the human brain. He said every brain has a Thinker and a Prover. The Thinker decides what's true ("school is pointless," or "school is important," or "dogs are better than cats"). Then the Prover goes out and finds evidence to confirm it. If you think school is pointless, your brain will helpfully notice every boring lecture and forget every interesting discussion. If you think school is important, your brain will do the opposite.[^3]

This isn't lying or being stupid. It's how brains work. We can't take in everything — there's too much information in the world — so our brains filter. The question is: what's doing the filtering? The linguist S. I. Hayakawa showed that every time we describe something in words, we lose detail.[^4] Describe your best friend in one sentence. Now describe them in a paragraph. Now describe them in a full page. Even the full page misses almost everything about who they are. That's the abstraction ladder: every level up, more detail falls away.

An LLM is a Prover with no Thinker — or more accurately, its "Thinker" is the entire internet it was trained on. When you ask it to analyze something, it doesn't approach the task with fresh eyes. It approaches with the combined assumptions of every article, textbook, and Reddit post it's ever processed. It will find themes that look like themes found in previous research, because that's what "finding themes" looks like in its training data.

Researchers Nguyen and Welch discovered exactly this problem in 2025. When people used AI chatbots to analyze research data, they got trapped in what the researchers called an "infinite loop" — the person would ask a question, the AI would give a plausible answer, the person would refine the question, and the AI would give a slightly different plausible answer. Neither the person nor the AI could tell when the answers were based on real patterns in the data versus patterns the AI was generating from its training.[^5]

Wilson's solution was what he called "model agnosticism" — the practice of never fully committing to any single way of seeing things.[^6] Not because all viewpoints are equally good (they're not), but because knowing that your current viewpoint is incomplete is the first step toward making it better. Applied to AI: treat everything it tells you as a hypothesis, never as a fact.

Imagine you come home and find the kitchen floor is wet. There are three ways you might think about this:

Deduction (rules → prediction): "If someone left the faucet running, the floor gets wet. The faucet is running. Therefore the floor is wet." This goes from a general rule to a specific prediction. It's reliable but boring — it only confirms what you already know.

Induction (patterns → rules): "Every time the dishwasher leaks, the floor gets wet. The dishwasher has leaked twelve times. Therefore, dishwashers leak." This goes from specific observations to a general rule. It's useful for building knowledge over time.

Abduction (surprise → explanation): "The floor is wet, but the faucet is off and the dishwasher is fine. That's weird. What would explain this? Maybe the dog knocked over the water bowl. Maybe a pipe is leaking inside the wall. Maybe it rained through the window I forgot to close." This is the creative kind — you start with something surprising and work backward to find the best explanation.[^7]

The philosopher Charles Sanders Peirce identified abduction as the only type of reasoning that produces genuinely new ideas.[^8] Deduction just applies existing rules. Induction just summarizes existing patterns. Abduction says: "Something unexpected happened — what would make it make sense?" This is what detectives do, what doctors do when they diagnose unusual symptoms, and what researchers do when they're trying to understand why people behave in surprising ways.

In December 2025, a team of researchers led by the philosopher Luciano Floridi published a landmark study asking whether LLMs actually perform abductive reasoning. Their conclusion: no. The AI produces output that looks like abductive reasoning — it generates explanations for surprising things — but it's doing something completely different under the hood.[^9]

Here's the key difference. When you encounter something surprising ("why is the floor wet?"), your brain generates possible explanations and then tests them against what you know. The AI, by contrast, generates the most statistically likely next words based on its training data. If its training data contains lots of examples of people explaining wet floors, it will produce a plausible-sounding explanation. But it's not actually reasoning about your floor. It's predicting what a good explanation would look like based on millions of previous explanations.

This matters most in the cases where abduction is most valuable: unusual situations. Medical researchers found that AI performed worst exactly when the diagnosis was rare and surprising — the cases where creative thinking matters most.[^10] When 3 out of 20 people in a study do something unexpected, a human researcher thinks: "That's interesting — what's different about those 3?" The AI, trained on statistical patterns, is more likely to wash out the signal and focus on the 17.

Think of it this way: the AI is like a student who has read every mystery novel ever written but has never actually solved a mystery. They can write a really convincing detective story. But if you hand them a real crime scene, they'll produce output that looks like detective work without actually doing any.

You've probably played the telephone game: someone whispers a message, it passes through ten people, and by the end it's completely different. Now imagine a version where, instead of whispering, each person writes a Wikipedia article about what the previous person said. Then the next person writes an article about that article. By the fifth round, you've got a well-written, carefully sourced article that has almost nothing to do with the original message.

That's what an LLM does. Korzybski identified three important facts about maps:[^11] (1) the map is not the territory, (2) the map doesn't cover all of the territory, and (3) you can make maps of maps, forever. An LLM operates entirely at level 3 — maps of maps of maps. It has never visited the territory. Its entire "experience" consists of other people's descriptions: transcripts, articles, papers, social media posts. It generates new text by remixing these descriptions.

Researchers at Carnegie Mellon University tested what happens when you use an AI to represent different people's experiences. They found something disturbing: the AI mashed together different perspectives into one fake "average" person.[^12] Imagine interviewing a factory worker and a manager about their jobs. A human researcher would carefully separate their experiences — the whole point is that they see the same workplace differently. The AI blended them together, losing exactly the differences that mattered most.

A philosopher named Thomas Dylan Daniel, drawing on the work of Alasdair MacIntyre, explains why this is a deep problem, not just a technical glitch.[^13] MacIntyre argued that people who start from different basic assumptions literally cannot reach the same conclusion. Their starting points determine their destinations. The factory worker and the manager aren't just seeing different things — they're operating in different mental frameworks that can't be merged without throwing one of them away.[^14] The AI doesn't understand this. It averages. And averaging across genuine differences is one of the worst things you can do in research.

The researchers at Stripe Partners made the point through a different angle: the most interesting research questions live in places where the patterns haven't been seen before.[^15] If you're studying something truly new — a community that uses technology in a way nobody expected, a user behavior that doesn't fit any existing category — the AI literally cannot help you understand it, because it only knows things that already exist in its training data. It's like asking Google Maps to show you a trail that hasn't been mapped yet.

Hold your hand out at arm's length and look at your thumb. Now close one eye. Your thumb blocks part of your view. You have a blind spot — a place you can't see. You know it's there, and you can move your head to see around it. But the blind spot itself is invisible while you're in it.

In 1931, a mathematician named Kurt Gödel proved something shocking: every system of math that's powerful enough to be useful has blind spots built into it.[^16] There are true statements that the system can't prove from within itself. This isn't because mathematicians aren't smart enough. It's a structural feature of the system — like how a camera can't photograph itself.

Thomas Dylan Daniel takes Gödel's insight and extends it beyond math.[^17] Daniel argues that language itself — the tool we use to think, argue, and make sense of the world — can never be as complex as the reality it describes.[^18] This isn't a complaint. It's actually the point. Daniel writes that "complete logical structures are universally flawed" — if your argument seems to explain everything with no blind spots, that's not a sign of a good argument. It's a sign of a dangerous one.[^19]

Think about it this way. Do you know someone who has an answer for everything? Who is never uncertain, never says "I don't know," never admits they might be wrong? That's suspicious, right? The smartest people you know probably say "I'm not sure" fairly often. Daniel's point is that this principle applies to arguments too: the ones that admit their blind spots are healthier than the ones that claim to have none.

An LLM's output always presents itself as complete. Ask it to analyze a set of interviews and it will produce a tidy set of themes that account for everything — no loose ends, no uncertainty, no "I don't know what to make of this." But Daniel has shown us that completeness is the mark of a flawed structure. The AI doesn't signal its blind spots because it can't perceive them.

This connects to a problem researchers call "hallucination." When an AI makes something up — a fake quote, a nonexistent source, a pattern that isn't in the data — that's a hallucination. Researcher Nguyen-Trung studied this in 2025 and concluded that "LLM hallucination is inevitable and unavoidable."[^20] In math or science, you can catch a hallucinated number by checking the data. But when the AI hallucinates an insight — "participants expressed ambivalence about convenience versus authenticity" — that sounds exactly like a real finding. You can't tell it's fake without going back to the original interviews. And the whole point of using the AI was to save time on reading the original interviews.

Here's Daniel's most powerful idea for understanding AI: "Truth as a statement is made, rather than found. True statements about the world must involve creativity, by definition."[^21] This means that discovering something true about the world requires creating something new — finding the words that illuminate what needs illuminating. An AI doesn't create truths. It finds the most statistically probable arrangement of existing words. That's maps of maps, not a new map drawn from the territory.

Wilson made the same point from a different angle in The New Inquisition: blindly trusting any single tool — whether it's a scientific method, a religious text, or an AI — to produce reliable knowledge without constantly questioning it is a recipe for self-deception.[^22]

Imagine your English teacher assigns 20 essays by your classmates about summer vacation. Your job is to find patterns — common themes that run through multiple essays. The right way to do this: read all 20 essays, notice what keeps coming up, and build your themes from the ground up. Maybe you notice that a surprising number of people wrote about feeling lonely, or about a moment when they tried something for the first time.

Now imagine doing it backwards: before reading a single essay, you decide what the themes probably are ("fun," "family," "travel"). Then you read the essays just to sort them into your pre-decided categories. You'd miss everything surprising. You'd miss the loneliness pattern entirely, because it wasn't in your starting categories.

Researcher Morgan found in 2023 that this is exactly what LLMs do. They "reverse the traditional inductive cycle by presenting broad concepts first for researchers to query rather than moving from raw data to themes."[^23] The AI starts with the general (everything it learned in training) and works toward the particular (your data). The themes come from the AI's prior knowledge, not from your data. Your data is a trigger, not an anchor.

Daniel identifies this exact problem in his discussion of conclusion-first reasoning: deciding what's true before looking at the evidence is "perhaps the most effective way to ensure that our reason will become corrupted and ineffective." When an AI analyzes your data, its training distribution functions as a set of pre-chosen conclusions. The result is what Daniel calls a complete-seeming structure that is, precisely because of its completeness, flawed.[^24]

Other researchers confirmed the consequences. When evaluators tested AI against human analysts doing thematic analysis, they found that the AI "fragmented data unnecessarily, missed hidden meanings, and sometimes produced themes with unclear boundaries."[^25] The fragmentation happens because AI prefers neat categories (easier to pattern-match) over the messy, overlapping, ambiguous meanings that characterize real human experience. And the missed hidden meanings happen because AI can only find what's said — it can't pick up on what's not said. The pauses, the evasions, the things people can't put into words yet — those are often the most important findings. They're invisible to the AI.

A power saw can cut wood faster and more precisely than a human with a hand saw. But a power saw can't decide what to build. It can't look at a family's living room and figure out what kind of bookshelf would actually work for them. It can't notice that the floor is slightly uneven and adjust the design. The tool does the cutting; the carpenter does the thinking.

AI works the same way. Here's where it's strong and where it's not:

AI can handle: Transcription (turning recordings into text), data cleaning (fixing messy files), sorting things into categories that humans have already defined, writing clear prose once a human has determined what's true, and building prototypes once a human has figured out what to build. These are "cutting" tasks — mechanical, precise, and fast.[^26]

AI can't handle: Designing the research question, building trust with the people you're studying, noticing the body language and hesitations that reveal what someone really means, making the creative leap from "here's what I observed" to "here's what it means," and judging whether the final product actually helps real people. These are "carpenter" tasks — they require judgment, presence, and creativity.

The interesting middle ground: Researchers found that the most productive use of AI was not in generating answers but in surfacing disagreements. When human analysts and AI coded the same data differently, the disagreements became "the site of new theoretical insight."[^27] In other words, the AI was most useful not when it was right but when it was wrong in interesting ways. Its mistakes, examined by human eyes, revealed things that neither the human nor the AI had initially noticed.

Another study found that beginners benefited more from AI assistance than experts, but experts used it more carefully — like a sparring partner rather than a teacher.[^28] This makes sense: if you don't know what good analysis looks like, the AI's confident output can feel like a lifeline. But if you know what you're doing, you can use the AI's output as raw material for your own thinking.

Daniel's metadialectic helps explain why this middle ground works. The metadialectic is Daniel's name for the seventh type of reasoning — the one that arises after you've studied the other six (religious, scientific, historical, positive, negative, and the elenchus, an ancient questioning technique).[^29] Daniel's mentor Robert Pirsig, author of Zen and the Art of Motorcycle Maintenance, argued that real understanding comes from caring about what you're doing.[^30] The metadialectic doesn't give you answers. It gives you a way of recognizing what each way of thinking can and cannot see. Applied to AI: you're not asking "is the AI right?" You're asking "what is the AI's way of thinking unable to see?"

If you've ever been on a debate team, you know that the best way to prepare isn't to have someone agree with you. It's to have someone argue against you — hard. The stronger the opposing argument, the better you understand your own position. You find the weak spots, the assumptions you didn't know you were making, the evidence you forgot to look for.

That's how AI should be used for thinking. Not as an oracle that gives you answers, but as a debate partner that challenges your reasoning.

The philosopher Karl Popper argued that we learn more from trying to disprove a hypothesis than from trying to confirm it.[^31] Scientists don't design experiments to prove they're right — they design experiments that could prove them wrong. If the experiment fails, they've learned something. If it succeeds, they know their idea survived a tough test.

Applied to AI: don't ask the AI to generate your insights. Instead, develop your own insights first, then feed them to the AI and ask it to destroy them. "Here's what I think is going on — what's the strongest argument against this?" "What am I not considering?" "What alternative explanation would fit the same evidence?" In this mode, the AI's ability to generate plausible alternatives becomes a feature, not a bug. You're not trusting its reasoning. You're using its pattern-matching to expand your own thinking.

Daniel Dennett offered a specific rule: before you criticize someone's argument, restate it so well that they say "I wish I'd said it that way."[^32] You can ask AI to do this — to steel-man an argument before attacking it. "Give me the strongest possible version of the position I disagree with." The result won't be perfect, but it will push you to engage with ideas you might otherwise dismiss.

Carl Sagan created what he called the "Baloney Detection Kit" — a set of questions for testing whether a claim holds up.[^33] You can apply these to AI output: Is there independent confirmation? Has this claim been tested against alternatives? Is the source reliable? Could this claim be proven wrong? Am I accepting it because it's true, or because it sounds good?

René Descartes, back in 1641, started modern philosophy by asking what might be the simplest and most powerful question ever posed: what do I actually know for certain?[^34] His answer — almost nothing — turned out to be the most productive starting point in the history of thought. Not because uncertainty is fun, but because being honest about what you don't know is the only way to figure out what you do.

Korzybski taught us that every description is a map, and every map leaves things out. Wilson taught us that our brains are bias machines that automatically filter reality to confirm what we already believe — and that the first step toward freedom is noticing the filter.[^35] Daniel showed that language itself cannot be as complex as the reality it describes, and that recognizing this — through the self-aware reasoning he calls the metadialectic — is what makes genuine open inquiry possible. And Peirce showed that the most creative kind of thinking, abduction, is the one thing that produces genuinely new ideas.

Large language models are, in a strange way, the perfect illustration of all of these ideas. They are incredibly powerful map-making machines that have zero access to the territory. They generate output that mimics the form of reasoning without performing its substance. They are fluent, persuasive, and incomplete — and their incompleteness is hardest to spot precisely where their fluency is greatest.

This doesn't mean you should avoid using AI. It means you should use it the way a good philosopher uses any tool: with awareness of what it can and cannot do. Treat its output as a map — useful, always incomplete, sometimes misleading. Ask what it's leaving out. Ask what would look different from another angle. And most importantly, remember Daniel's insight: truth is made, not found. The AI can find patterns in existing text. But making sense of the world — creating genuine understanding from real experience — that's still your job. No one else can do it for you. Not even the most impressive map-making machine ever built.

• • •

[^1]: Alfred Korzybski, Science and Sanity: An Introduction to Non-Aristotelian Systems and General Semantics (1933). Korzybski was a Polish-American scientist who studied how language shapes the way we think. His most famous idea: "The map is not the territory." Full text available at the Internet Archive

[^2]: Robert Anton Wilson, Prometheus Rising (Tempe, AZ: New Falcon Publications, 1983), ch. 1. Wilson was a writer and thinker who explored how our brains filter reality. The Thinker/Prover model is from chapter 1.

[^3]: Wilson, Prometheus Rising, ch. 1.

[^4]: S. I. Hayakawa, Language in Thought and Action, 5th ed. (San Diego: Harcourt Brace Jovanovich, 1990), ch. 2. Hayakawa created the "abstraction ladder" — a diagram showing how every time we summarize or generalize, we lose detail from the original experience.

[^5]: Duc Cuong Nguyen and Catherine Welch, "Generative Artificial Intelligence in Qualitative Data Analysis: Analyzing — Or Just Chatting?," Organizational Research Methods (2025). The researchers found that people using chatbots for research got trapped in a loop where they couldn't tell if the AI was finding real patterns or just agreeing with them. DOI: 10.1177/10944281251377154

[^6]: Robert Anton Wilson, Quantum Psychology: How Brain Software Programs You and Your World (Tempe, AZ: New Falcon Publications, 1990), ch. 1. Wilson called this "the Universe Contains a Maybe" — the idea that we should think in terms of probabilities, not certainties.

[^7]: Charles Sanders Peirce, Collected Papers, vol. 5 (Cambridge: Harvard University Press, 1931–58), §5.189. Peirce (pronounced "purse") was an American philosopher who identified three types of reasoning: deduction, induction, and abduction. He called abduction "the only logical operation which introduces any new idea."

[^8]: Peirce, Collected Papers, §5.189.

[^9]: Luciano Floridi, Jessica Morley, Claudio Novelli, and David Watson, "What Kind of Reasoning (if any) is an LLM Actually Doing? On the Stochastic Nature and Abductive Appearance of Large Language Models" (December 2025). Floridi is a philosopher of information at Yale who studies whether AI systems actually reason or just look like they do. arXiv:2512.10080

[^10]: R. Thomas McCoy et al., cited in "Limitations of Large Language Models in Clinical Problem-Solving Arising from Inflexible Reasoning," Scientific Reports (November 2025). The researchers found that AI doctors performed worst exactly when the diagnosis was unusual — the cases where creative thinking matters most.

[^11]: Korzybski (1933), pages 58–61. Korzybski's three rules: (1) the map is not the territory; (2) the map doesn't cover ALL of the territory; (3) you can make maps of maps of maps, forever.

[^12]: Shivani Kapania et al., "'Simulacrum of Stories': Examining Large Language Models as Qualitative Research Participants," CHI 2025 (Carnegie Mellon University). The researchers found that when AI tried to represent different people's experiences, it mashed them together into one fake "average" person that didn't represent anyone accurately. DOI: 10.1145/3706598.3713220

[^13]: Alasdair MacIntyre, After Virtue: A Study in Moral Theory (Notre Dame: University of Notre Dame Press, 1981). MacIntyre argued that people who start with different basic assumptions can literally never reach the same conclusion — not because they're stupid, but because their starting points make their destinations inevitable.

[^14]: Thomas Dylan Daniel, Formal Dialectics (Newcastle upon Tyne: Cambridge Scholars Publishing, 2018), Introduction, page 9. Daniel identifies three properties of every type of reasoning: (1) it's focused on a specific subject, (2) it's incomplete, and (3) it can't be merged with reasoning that starts from different assumptions without starting over.

[^15]: Stripe Partners, "Grounded Models: The Future of Sensemaking in a World of Generative AI," EPIC Proceedings (November 2024). The authors argue that the most interesting research questions live in places where patterns haven't been seen before — which means, by definition, they're not in the AI's training data. Available at epicpeople.org

[^16]: Kurt Gödel, "Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I," Monatshefte für Mathematik und Physik 38 (1931): 173–198. Gödel was a mathematician who proved something mind-bending: any system of math powerful enough to be useful will always contain true statements it can't prove. Even math has blind spots. [Accessible introduction: Ernest Nagel and James R. Newman, Gödel's Proof (New York University Press, 1958)]

[^17]: Thomas Dylan Daniel, Formal Dialectics (2018). Daniel is a philosopher who argues that language can never be as complex as the reality it describes — and that understanding this is actually useful, not depressing.

[^18]: Daniel (2018), Introduction. Daniel defines the metadialectic as "the dialectic of rational incomplete form" — a way of thinking that accounts for its own blind spots. It's the seventh of his seven types of reasoning, and it only exists because you've studied the other six.

[^19]: Daniel (2018), Introduction, page 13. Daniel writes: "Complete logical structures are universally flawed. It is precisely by virtue of incompleteness that language can resemble the parts of the world a skilled user intends."

[^20]: Nguyen-Trung, "Hallucinations in GenAI Thematic Analyses," Quality & Quantity (2025). The researcher was blunt: "LLM hallucination is inevitable and unavoidable" — meaning AI will always sometimes make things up, and you can't tell when by looking at the output.

[^21]: Daniel (2018), Introduction, page 11. Daniel writes: "Truth as a statement is made, rather than found. True statements about the world must involve creativity, by definition." This means that discovering something true requires imagination, not just pattern-matching.

[^22]: Robert Anton Wilson, The New Inquisition: Irrational Rationalism and the Citadel of Science (Tempe, AZ: New Falcon Publications, 1986). Wilson argued against "Fundamentalist Materialism" — blindly trusting any single method to produce reliable knowledge without constantly questioning it.

[^23]: David L. Morgan (2023), cited in multiple 2025 research surveys. Morgan found that AI does qualitative analysis backwards: it starts with broad themes from its training data and then matches your data to those themes, instead of building themes from your data.

[^24]: Daniel (2018), Introduction, pages 12–13. See note 19.

[^25]: "Large Language Models in Thematic Analysis," arXiv:2510.18456 (October 2025). Researchers found that AI "fragmented data unnecessarily, missed hidden meanings, and sometimes produced themes with unclear boundaries." arXiv:2510.18456

[^26]: "Scaling Hermeneutics: A Guide to Qualitative Coding with LLMs for Reflexive Content Analysis," EPJ Data Science (April 2025). The researchers found that the best results came from humans leading the analysis with AI assisting — not the other way around.

[^27]: "Exploring the Human-LLM Synergy in Advancing Theory-Driven Qualitative Analysis," ACM Transactions on Computer-Human Interaction (2025). The researchers found something surprising: AI was most useful not when it was RIGHT but when it was WRONG in interesting ways — because its mistakes helped human researchers notice things they'd missed.

[^28]: "Qualitative Coding Analysis through Open-Source Large Language Models: A User Study and Design Recommendations," arXiv:2602.18352 (February 2026). The study found that beginners benefited more from AI help, but experienced researchers used it more carefully — treating it as a sparring partner instead of an expert. arXiv:2602.18352

[^29]: Daniel (2018), Introduction. See note 18.

[^30]: Robert Pirsig, Zen and the Art of Motorcycle Maintenance (New York: William Morrow, 1974). Pirsig argued that real understanding comes from caring about what you're doing — what he called "Quality." Daniel credits Pirsig as a major influence on Formal Dialectics.

[^31]: Karl Popper, Conjectures and Refutations: The Growth of Scientific Knowledge (London: Routledge and Kegan Paul, 1963), ch. 1. Popper argued that we learn more from trying to DISPROVE an idea than from trying to prove it. This is why scientists design experiments that could fail — failure teaches you more than success.

[^32]: Daniel Dennett, Intuition Pumps and Other Tools for Thinking (New York: Norton, 2013), pages 33–35. Dennett's rule: before you criticize someone's argument, restate it so well that they say "I wish I'd said it that way." Only then have you earned the right to disagree.

[^33]: Carl Sagan, The Demon-Haunted World: Science as a Candle in the Dark (New York: Random House, 1995), ch. 12. Sagan created the "Baloney Detection Kit" — a checklist for figuring out whether a claim is trustworthy.

[^34]: René Descartes, Meditations on First Philosophy (1641), Meditation I. Descartes started modern philosophy by asking: what do I know for certain? His answer: almost nothing — which turned out to be the most productive starting point in the history of thought.