What Is Knowledge?

§1 The machinery

The two trapdoors under every Gettier case

Before we explore new rooms, look down. Gettier's three-page bomb only goes off because the floor has two trapdoors built into it — two assumptions so natural the main lesson never paused on them. Name them and the whole landscape reorganizes.

Trapdoor one: fallible justification. The classical picture lets you be justified in believing something that turns out false. Smith had excellent reason to believe "Jones will get the job" — the boss said so — and it was false. If justification had to guarantee truth, that step would be impossible and the case couldn't even start. Trapdoor two: closure. Justification (and knowledge) is assumed to travel across entailment: if you're justified in believing something, you're justified in believing what it obviously implies. Smith reasons from "Jones will get it (and has ten coins)" to the weaker "the winner has ten coins" — a valid inference — and carries his justification along for the ride. Knock out either plank and Gettier cases evaporate.

That hands us a tempting exit. Slam trapdoor one shut: insist that real knowledge needs infallible justification — reasons that make error literally impossible. No more Gettier cases, ever. This is the dream of infallibilism, and it is very old. Descartes went looking in 1641 for a single belief no demon could fake, and found exactly one that survives even the supposition that an all-powerful deceiver is fooling you about everything else: I think, therefore I am. You cannot be tricked into wrongly believing you exist, because the tricking requires a you to be tricked.

The trouble is what the demon takes with him on the way out. If knowledge demands that kind of certainty, then you do not know you have hands, that the sun will rise, that the person across the table is your friend and not an android — because a clever enough deception could fake any of it. Buy certainty and the price is skepticism: the bar is set so high that almost nothing clears it. Peter Unger argued exactly this in Ignorance (1975) — that "knows," used strictly, applies to virtually nothing, much as "flat" strictly applies to no real surface. So infallibilism doesn't dissolve the problem; it trades a small puzzle (the odd lucky belief) for a total one (you know next to nothing). Which is our cue to open the next door, where that skeptic is already knocking.

§2 The biggest omission

The skeptic at the door

Western epistemology has a recurring houseguest who refuses to leave: the figure who says you can't know anything about the world outside your own mind. The main lesson kept the door shut. Open it, because every modern theory of knowledge is partly built to deal with what's standing there.

The skeptic's tools are thought experiments, escalating in cruelty. First, the dream: right now, how do you know you're not asleep? Dreams feel utterly real from the inside; you've been fooled before. (The Daoist Zhuangzi, around 300 BCE, dreamt he was a butterfly and woke unsure whether he was a man who'd dreamt a butterfly or a butterfly now dreaming a man — the same wound the Buddhist Dharmottara reopened in the main lesson, proof again that minds keep tripping over this independently.) Descartes raised the stakes to an evil demon bent on deceiving you about everything. The twentieth century updated the hardware: you might be a brain in a vat, nerves wired to a computer feeding you exactly the experiences you're having now (Hilary Putnam, Reason, Truth and History, 1981). You cannot tell from the inside. That's the whole point.

Spelled out, the skeptic's argument is brutally clean — and it runs on the very closure principle from §1:

(1) You don't know you're not a handless brain in a vat being fed a hand-experience.
(2) If you know you have hands, then (since having hands entails not being a handless vat-brain) you know you're not one.
(3) So you don't know you have hands.

Each line looks reasonable; together they seem to prove you know nothing about the external world. The interactive below lets you try every way out — and discover that each "way out" is a named philosophical position with a price tag.

The doors are worth naming in full. G. E. Moore (1939) simply ran the argument backwards: I am far more sure that here is one hand (holding it up) than I am of any fancy premise the skeptic offers — so if the premises imply I don't know it, so much the worse for the premises. Cheeky, and strangely hard to beat. Fred Dretske (1970) and Robert Nozick (1981) took the surgical route: deny closure. On Dretske's relevant alternatives view, to know something you only need to rule out the relevant ways you could be wrong, not every bizarre one. At the zoo you know the animal is a zebra — you've ruled out "it's a horse," "it's a goat" — even though you haven't ruled out "it's a mule cleverly painted to look like a zebra," because in this context that's not a live possibility. Knowledge doesn't automatically transmit to every entailment. The cost is steep: closure is intuitive, and giving it up has consequences elsewhere. Contextualism (our next section) offers the diplomat's solution: maybe the skeptic and Moore are both right, because "know" means something stricter in the skeptic's seminar than in ordinary life.

The 2020s upgrade: are we in a simulation?

The vat got a software update. Nick Bostrom's simulation argument (Philosophical Quarterly, 2003) makes a careful probabilistic case that at least one of three things is true: civilizations almost never reach the technology to run ancestor-simulations; or they reach it but choose not to; or we are almost certainly living in one. David Chalmers, in Reality+ (2022), takes the next step and bites a bullet most people won't: he argues we can't know we're not simulated and should assign the possibility real probability — but that this isn't a catastrophe, because "virtual reality is genuine reality." A simulated tree, on his simulation realism, is a real digital object, not an illusion; if you've always lived in a perfect simulation, your belief "that's a tree" is true, just realized in silicon. The skeptic assumed a fake world means false beliefs; Chalmers denies the link.

Two honest labels before we move on. The simulation hypothesis — that we are in fact simulated — is, as it stands, untestable metaphysics, not science: there's no agreed observation that would confirm or refute it, which puts it on the wrong side of the demarcation line we'll draw tomorrow. simulation hypothesis · unfalsifiable The philosophical payoff is real all the same: it sharpens what we even mean by "real" and "know." And there's a famous reply that turns the screw the other way. Putnam argued that "I am a brain in a vat" is self-refuting: your words only mean what they do because of your causal history, so a lifelong brain-in-a-vat's word "vat" couldn't refer to real vats (it never causally touched one) — meaning that if you were a vat-brain, your sentence "I am a brain in a vat" would come out false. Whether that works is still argued, which is precisely why this thread runs straight into the AI block: when a system trained only on text outputs "Paris is in France," does it know that — or is it the purest brain-in-a-vat of all, with words that never touched the world? Hold the question for Days 138–145.

§3 The moving target

"Knows" on a sliding scale

Here is a possibility the main lesson never entertained: maybe sixty years of hunting the perfect definition of "knows" failed because the word was never aiming at a fixed point. Consider a pair of cases from Keith DeRose (Philosophy and Phenomenological Research, 1992) that have launched a thousand papers — the Bank Cases.

It's Friday. You drive past your bank, which has a long Saturday line, and decide to come back tomorrow. Your wife asks if it'll be open Saturday. Low stakes version: nothing much rides on it; you say, "Yes, I know it's open Saturdays — I was here two Saturdays ago." That sounds true. You know it. High stakes version: there's a check that must be deposited by Monday or you bounce your mortgage and lose the house, and your wife points out, reasonably, that banks do change their hours. Now the very same sentence — "I know it's open Saturday" — curdles in your mouth. "Well… I'd better go in and check." Same person, same memory, same evidence, same day. Only the stakes (and whether someone raised the chance of error) have changed. Yet the knowledge seems to come and go. The dial below lets you slide between the two and watch it flip.

Three camps, three diagnoses of the same data. Contextualism (DeRose; David Lewis, "Elusive Knowledge," 1996; Stewart Cohen, 1988) locates the shift in the word: "knows" is like "tall" or "here" — context-sensitive. Raising the stakes or mentioning error raises the standard a belief must meet for the sentence "S knows" to count as true. Both utterances are correct, in their own conversations. The skeptic is even right in the seminar — he's just jacked the standard sky-high. Pragmatic encroachment (Jason Stanley, Knowledge and Practical Interests, 2005; Fantl & McGrath; John Hawthorne, Knowledge and Lotteries, 2004) puts the shift in the knower: what you know genuinely depends on what's practically at stake for you, because knowledge is supposed to be the thing you can act on. High stakes really can deprive you of knowledge you'd have had when it didn't matter — a startling idea, since it lets practical pressure "encroach" on a supposedly purely factual state. Invariantism (the traditional holdout) digs in: "knows" means one fixed thing, the standards don't move, and one of your two verdicts is simply mistaken — you either knew all along or never did, and the stakes just changed how willing you were to say so. that the verdicts shift · agreed why they shift · unresolved The data is robust; its explanation is one of the most active fault lines in the field.

§4 The pattern behind the patches

The luck we were really chasing

Return to the patch wars from the main lesson — no-false-lemmas, sensitivity, safety, virtue. They looked like a grab-bag of clever fixes that each met a nastier counterexample. Step back and they snap into focus: every one was chasing the same ghost. Duncan Pritchard gave it a precise name in Epistemic Luck (Oxford, 2005). The enemy of knowledge is a specific species he calls veritic luck: your belief is true in the actual world, but in almost all the nearby ways things could have gone, you'd have believed the same thing and been wrong. The truth and your believing it are only accidentally in step.

This is the deep content of the "safety" idea, and it's worth seeing. Picture the actual world as a dot, ringed by the nearby possible worlds — the small, realistic variations on how things might have been. A belief is safe (knowledge-grade) when it stays true across that neighborhood, and unsafe (merely lucky) when a slight nudge flips it to false. Toggle the three scenarios below and watch the neighborhood light up.

That single picture retroactively explains the whole mess. The stopped clock fails hard — a minute either side and you're wrong, so the neighborhood is a sea of red. Fake-barn country is subtler: the barn you're looking at is genuinely there (the core is green), but you're surrounded by façades, so a glance a hundred meters either way would have fooled you — red neighborhood, no knowledge, even with a true justified belief and no false premise. The patches all failed because each tried to capture "green neighborhood" with a slightly different yardstick, and luck kept finding the gaps.

Two more patches the main lesson didn't name, now that we have the frame. Defeasibility theory (Lehrer & Paxson, 1969) said knowledge is un-defeated justified true belief: there must be no true fact out there that, if you learned it, would dissolve your justification. It handles many cases elegantly — until the "misleading defeater" twist, where there's a true-but-misleading fact that shouldn't rob you of knowledge but technically does, forcing ever-finer distinctions. And reaching back further, the causal theory (Goldman, 1967, before he turned reliabilist) demanded that the fact cause your belief — no causal chain, no knowledge. Beautiful for perception; fatal for mathematics, since the number 7 and the Pythagorean theorem don't cause anything (Paul Benacerraf pressed exactly this "access problem" in 1973). You can't shake hands with an abstract object.

And the deepest crack in reliabilism, which the main lesson only gestured at: the generality problem (Conee & Feldman, 1998). Reliabilism says a belief is justified if produced by a reliable process — but which process? Your belief that it's 9:12 was produced by "reading a clock," and also by "reading that clock," and "using vision in dim light," and "trusting instruments on Tuesdays" — each as real as the others, each with a different reliability score. Pick the type and you've picked the verdict. Specifying the "right" grain, in a principled way, has proven stubbornly hard.

Where does Pritchard land? At anti-luck virtue epistemology: knowledge needs both conditions, because they catch different failures. You need safety (a green neighborhood — no veritic luck) and you need aptness (the belief is true through your own ability — the archer's skill from the main lesson). Neither alone suffices: the stopped clock can fail safety, fake-barn country can have local skill but bad luck. It's not a tidy three-word formula — and that, by now, may be the lesson. Knowledge might just be the kind of thing that takes two independent guarantees, one about you and one about your world.

§5 The question under the question

Why is knowing worth more than just being right?

Step back from "what is knowledge?" to a question Plato asked first and nobody has fully answered: why do we care? If a true belief gets the job done, what does the extra machinery of knowledge buy you? Plato put it as a traveler's problem in the Meno (~380 BCE). Suppose you want to walk to the town of Larissa. A person who knows the road will get you there. But so will a person who merely has a true belief about the road — who's never been, but happens to be right. For the purpose of arriving, the two are worth exactly the same. So why has the entire tradition prized knowledge above true belief? This is the value problem, and it's a load-bearing question: a theory of knowledge that can't say why knowledge is better has arguably missed the point of the concept.

The value problem turns into a precise weapon against one of the main lesson's theories. It's called the swamping problem (Linda Zagzebski, 2003). Reliabilism says knowledge is true belief from a reliable process. But ask what the reliability adds in value. Reliability is good only because it tends to produce truth. So once you already have the truth, what does it add that this particular truth also came from a reliable source? Zagzebski's homely analogy: a cup of good coffee is no better to drink for having come from a reliable coffee machine rather than an unreliable one that happened to produce an identical cup. The good-making feature (deliciousness / truth) is already present; the source's reliability gets swamped, adding nothing. If that's right, reliabilism can't explain why knowledge beats lucky true belief — the very thing a theory of knowledge most needs to deliver.

This is where virtue epistemology earns its keep, and where the archer finally pays off. Its answer: knowledge isn't valuable as a better-stocked true belief; it's valuable as an achievement — a success that's yours, that came about through your own competence. And achievements carry a kind of worth that lucky successes never do, the way a bullseye you actually aimed is worth something a lucky gust-blown hit isn't, even though the arrow lands in the same spot. A true belief reached through your own cognitive skill is a cognitive achievement; a lucky true belief is not. That's the extra value — not in the result, but in the getting there. The road to Larissa you can actually find again is worth more than the one you stumbled onto, even on a day you both arrive.

§6 The other knowings

The kinds of knowing we ignored

Everything so far — the entire main lesson — was about propositional knowledge, knowledge-that: knowing that it's 9:12, that Jones got the job. But look how much of "know" in plain English isn't that at all. You know how to ride a bicycle. You know your mother's face. You know Lisbon. None of these is a stockpile of facts, and philosophers have argued for a century about how they relate.

Knowing-how. Gilbert Ryle, in The Concept of Mind (1949), insisted that knowing how to do something is not knowing a set of facts. A brilliant cyclist may be unable to state a single law of balance; a person who has memorized every fact about bicycles may topple on the first try. Worse, Ryle argued, reducing skill to facts triggers a regress: if every skilled act required first knowing the proposition describing the rule, you'd then need the skill of applying that rule, which would need another rule, forever. So skill must be its own kind of knowing. The twist: Jason Stanley and Timothy Williamson fired back in "Knowing How" (2001) with intellectualism — the claim that knowing-how just is a species of knowing-that after all (knowing, of some way to ride, that it is a way to ride), dressed in different grammar. Whether skill collapses into propositions is genuinely unsettled. knowing-how reducible? · contested

Knowing by acquaintance. Bertrand Russell (1911) drew a second cut: between knowledge by acquaintance — your direct, unmediated grip on a patch of red you're seeing, a pain you're feeling, a face you're looking at — and knowledge by description, the facts you know about things you've never directly met ("the first person to stand on the Moon," whom you know only as the one satisfying that description). You can know a stupendous amount about Bismarck and never have known him; you know the color red in a way the world's greatest blind physicist, who knows every fact about wavelengths, does not. That gap — facts about an experience versus the experience itself — is a quiet seed for the hardest problem in the entire course, the one waiting on Day 123: why there's something it is like to see red at all.

§7 The social turn

Almost everything you know, someone told you

The main lesson, like most of traditional epistemology, imagined a lone mind facing the world — one person, one clock. But run an audit of what you actually know. That the Earth is about 4.5 billion years old. That Antarctica exists. Your own date of birth. The boiling point of water. You verified essentially none of it first-hand; you were told, by teachers, books, parents, instruments, strangers. Testimony is the overwhelming bulk of any human being's knowledge — and for centuries epistemology treated it as an afterthought.

The central question is whether trusting testimony is something you have to earn or something you're entitled to by default. David Hume (1748) took the demanding line: testimony is only as good as your own inductive track record of when testimony has proved reliable — it reduces to evidence you've personally gathered. Thomas Reid (1764) found this absurd: no child could bootstrap a track record before trusting anyone, and in fact we're built with a "principle of credulity," a default disposition to believe what we're told, exactly as we're built to trust our senses. On Reid's anti-reductionist view, testimony is a basic source of knowledge, not a derived one — and it has to be, or knowledge couldn't get off the ground in a social animal. The modern field mostly agrees that some default trust is unavoidable; the fights are over how much, and when it's defeated.

Two newer rooms branch off this one, and both matter enormously in 2026. The first is disagreement. When someone you regard as an epistemic peer — as smart, as informed, as careful as you — looks at the same evidence and concludes the opposite, what should you do? The conciliationist or "equal-weight" view (Adam Elga, Noûs, 2007; David Christensen, 2007) argues you should move substantially toward them: to stay put is to claim, with no independent reason, that you're the one who got it right and they made the mistake. The steadfast view answers that sometimes you can rationally hold your ground, because your own reasoning is evidence too. It sounds abstract until you notice it's the whole epistemology of echo chambers, expert consensus, and what to do when half your sources contradict the other half. equal-weight view · live debate

The second is sharper still: epistemic injustice, named by Miranda Fricker (Epistemic Injustice: Power and the Ethics of Knowing, 2007). Because so much knowing runs on testimony, who gets believed becomes an ethical question, not just an epistemic one. Fricker isolates two wrongs. Testimonial injustice: a speaker's word is given less credence than it deserves because of prejudice about who they are — the patient whose pain is dismissed, the witness disbelieved for their accent or gender. Hermeneutical injustice: subtler and deeper — a person can't even make sense of their own experience, to themselves or others, because the surrounding culture hasn't yet developed the concept for it (her example: the experience we now call sexual harassment, suffered by people who had no word for it and so couldn't name the wrong). Knowledge, it turns out, has a politics: the tools for understanding are unevenly distributed, and that unevenness can itself be an injustice.

§8 The formal edge

Two more frontiers, beyond Bayes

The main lesson's formal frontier was Bayesian credence. Two further formal ideas deserve a place on the map, because both bite ordinary intuitions and both feed straight into computer science and AI.

The logic of knowing. Jaakko Hintikka, in Knowledge and Belief (1962), treated "knows" as a formal operator you can reason with, like "necessarily" — launching epistemic logic, now a workhorse in computer science (reasoning about what distributed agents and AI systems "know"). It immediately surfaces deep puzzles. The KK principle: if you know p, do you thereby know that you know p? Tempting, but Williamson (from the main lesson) argues it's false — you can know something without being in a position to know that you know it, because knowledge has blurry margins. And logical omniscience: the clean logic implies that if you know some axioms, you know every logical consequence of them — which would make every mathematician instantly aware of every theorem. Obviously false for real, bounded minds, and a central headache for modeling actual reasoners (and machines).

The preface paradox. A companion to the lottery paradox from the main lesson, and arguably nastier. You write a long, careful book. For each claim in it, you've checked your work and rationally believe it's true. Yet you also write, sincerely, in the preface: "no doubt errors remain, and they are mine alone" — because you know that across hundreds of claims, you've almost certainly slipped somewhere. So you rationally believe each individual claim, and also rationally believe that at least one of them is false (David Makinson, "The Paradox of the Preface," 1965). Those can't all be true together. The moral lands on the main lesson's open question with full force: ordinary all-or-nothing belief isn't closed under conjunction — believing each of many things doesn't license believing their grand conjunction — which is one more reason the field keeps drifting from yes/no belief toward graded credence. The dial, again, doing what the switch can't.

Open questions

What this appendix leaves unsettled

• Certainty or not? Is the infallibilist right that real knowledge needs error-proof reasons (inviting skepticism) — or is fallible knowledge the only kind worth wanting?
• Can closure be denied without disaster? Dretske and Nozick block the skeptic by giving it up; the cost elsewhere is still being counted.
• Does "knows" move? Context-sensitive, stakes-sensitive, or fixed — and if it moves, what exactly is moving, the word or the world?
• Can the value of knowledge be explained at all, or does every account leave knowledge looking no better than lucky true belief?
• Is knowing-how just knowing-that in disguise, or its own irreducible kind of grip on the world?
• Is testimony basic or earned? — and, downstream, when a peer disagrees, must you really meet them halfway?
• And the function-first wager: if the concept of knowledge exists to flag good informants, does that dissolve the analysis project — or just relocate it?

§1 The activity turn

Epistemology forgot to ask about inquiry

Here is a strange omission, and once you see it you can't unsee it. For a century, epistemology has been almost entirely a theory of states — of belief, justification, knowledge: snapshots of a mind that has already finished thinking. It has had remarkably little to say about the activity that produces those states — the messy business of inquiry: asking a question, deciding what evidence to gather, knowing when to stop. Jane Friedman put a name to the missing half and lit a fuse under the field. The norms of inquiry she calls zetetic norms (from the Greek zētein, to seek), and her landmark paper "The Epistemic and the Zetetic" (Philosophical Review, 2020) argues something genuinely destabilizing: the norms of inquiry and the norms of belief are not merely separate — they actively conflict.

The engine is a principle so obvious it sounds like a truism — the Zetetic Instrumental Principle: if you want to figure out the answer to a question, you ought to take the necessary means to figure it out. Now watch it collide with a bedrock epistemic norm — the evidentialist's command that you may believe whatever your evidence already supports. Suppose you're trying to count the windows on the building across the street. Good inquiry says: focus, go count the windows, don't get distracted. But at every passing instant your senses hand you sufficient evidence to form, and be permitted to believe, a thousand idle truths — the color of that car, the number of people on the corner, the shape of a cloud. The belief-norm permits all of them. The inquiry-norm tells you to ignore all of them and count windows. Conform to one and you flout the other. The diagram makes the squeeze concrete.

Why does this matter beyond the seminar? Because it suggests epistemology has been studying the wrong unit. If belief-norms and inquiry-norms genuinely clash, then a theory built only on belief is incomplete — maybe even backwards. The radical proposal, the "zetetic turn," is that all epistemic norms are ultimately norms of inquiry (suspension of judgment becomes a question-directed attitude; believing an answer is a way of closing a question). The field has not swallowed this whole — and that's the honest part. Arianna Falbo ("Should epistemology take the zetetic turn?", Philosophical Studies, 2023) and others argue the inquiry-norms are really practical, not distinctively epistemic, and that a purely zetetic epistemology can't explain why some beliefs are irrational even when believing them would help your inquiry. So: the puzzle is now taken with great seriousness across the field; the grand thesis that inquiry swallows everything is a genuine, unresolved fight. Either way, the question "what is knowledge?" is quietly being reframed as "what is it to inquire well?" — and that reframing reaches all the way forward to Day 2, where the scientific method is exactly a set of norms for collective inquiry.

§2 The cognitive-science turn

What if knowledge comes before belief?

Day 1 treated knowledge as something built up from belief: take a belief, add truth, add justification, screen out luck. Almost every theory we met assumed belief is the raw material and knowledge the finished product. A large interdisciplinary team led by Jonathan Phillips and Joshua Knobe dropped a target article in Behavioral and Brain Sciences — "Knowledge before belief" (2021) — arguing that, as a matter of how human (and animal) minds actually work, this may be exactly upside down.

The standard story in psychology is that our "theory of mind" — our capacity to model other minds — is centered on belief, and matures when a child finally passes the false-belief test (understanding that someone can hold a belief the child knows to be false) around age four. Phillips and colleagues marshal converging evidence that representing knowledge is the more basic feat. The threads: developmentally, infants and toddlers track who has seen or has access to information — who knows — well before they can handle false beliefs. Comparatively, non-human great apes show robust signs of tracking what others can perceive and know, while convincing evidence of belief-tracking remains elusive. And in adults, attributions of knowledge are made at least as fast as — sometimes faster than — attributions of belief, which is hard to explain if "X knows p" is computed by first building "X believes p" and then checking extra conditions. The picture they paint is a factive theory of mind: the mind's first and most basic tool for modeling others is "what do they know?", with the trickier, error-tolerant "what do they merely believe (perhaps falsely)?" coming later and costing more.

If it holds, the payoff for Day 1 is direct and large. It would be empirical ammunition for Timothy Williamson's knowledge-first program — the philosophical claim (which we met as pure armchair theory) that knowledge is the basic, unanalyzable state and belief should be explained in terms of it, not the reverse. Suddenly that's not just a logician's hunch; it's a candidate fact about the architecture of cognition, with an evolutionary rationale (a social animal urgently needs to track who has reliable information — echoing Edward Craig's "good-informant" genealogy from Appendix I). But honesty demands the asterisk, and here it's loud: a BBS target article arrives wrapped in dozens of peer commentaries, and many dissent hard. Critics argue the knowledge/belief line is blurrier than the authors allow, that "tracking who saw what" needn't be a full representation of knowledge, and that culture and language shape the whole picture. So: the behavioral findings (kids and apes track informational access early) are reasonably solid; the strong interpretation (knowledge-representation is metaphysically and computationally prior, with belief assembled from it) is very much live. The furniture of the mind is being rearranged in real time, and the movers don't yet agree on the floor plan.

§3 The machine turn, part one

Does a language model know anything — or just bullshit?

Day 1 ended on a needle of a question: when a system like the one that drafted these pages outputs a true, well-supported sentence, does it know anything — or is it the ultimate Gettier case, right for reasons that have nothing to do with the truth? The 2020s turned that closing flourish into one of the hottest debates in the field, and the most-discussed entry has a title that sailed past peer review unblunted: "ChatGPT is bullshit" (Hicks, Humphries & Slater, Ethics and Information Technology, 2024).

Their move is precise, not merely rude. They borrow Harry Frankfurt's technical sense of bullshit (from his 1986 essay On Bullshit): bullshit is speech produced with indifference to truth. The liar at least tracks the truth — he has to, in order to steer you away from it. The bullshitter doesn't care either way; he says whatever serves his purpose, and whether it's true is simply beside the point. Now consider what a large language model fundamentally is: a system trained to emit the statistically likely next token, to produce fluent, plausible-sounding text. It has no representation of truth that it is trying to honor. So when it states a real fact and when it "hallucinates" a fake citation, it is doing the very same thing — generating likely-looking text — and succeeding equally at its actual task in both cases. On this view, "hallucination" is a flattering misnomer that implies a malfunction; the truer description is that the system is indifferent to truth by design, which is bullshit in Frankfurt's exact sense. They distinguish soft bullshit (no intent to deceive, just truth-indifference) from hard bullshit (additionally posing as a sincere truth-teller), and argue an LLM is at minimum a soft bullshitter.

Why this could redraw the map: it cuts directly against the loose talk of machines "knowing," "understanding," or "believing." If the argument is right, an LLM's true outputs aren't knowledge and aren't even really assertions in the full sense — they're a new category of truth-apt-looking text with no one home who cares whether it's true. That reframes how we should trust, cite, and regulate these systems. And it is, predictably, contested — the rebuttals are already a small literature. Some argue the "bullshit" label smuggles in a stance on whether models have intentions at all (Sarah Fisher, "Large language models and their big bullshit potential," 2024; David Gunkel & Simon Coghlan, "Cut the crap," 2025); others that as models are trained with reinforcement to be truthful and to express calibrated uncertainty, "indifferent to truth" is too crude. What's settled is the unglamorous core: a base language model has no built-in commitment to truth, and fluency is not knowledge. What's open is whether "bullshit," "instrument," "testifier," or some entirely new epistemic category is the right home for what these systems produce. It is the brain-in-a-vat from Appendix I made of silicon and shipped to a billion users — words that may never have touched the world, now answering our questions.

§4 The machine turn, part two

The support beam you never noticed: the epistemic backstop

Appendix I established how much of what you know arrives by testimony — other people's say-so. Regina Rini's "Deepfakes and the Epistemic Backstop" (Philosophers' Imprint, 2020) identifies a hidden structural support that has quietly kept that whole edifice honest — and shows how a new technology is sawing through it.

The insight is subtle. Why is testimony as reliable as it is? Part of the answer, Rini argues, is a silent regulator: the ever-present possibility of a recording. When someone can be contradicted by a photograph, an audio clip, or a video, they have a standing incentive to testify truthfully — because a recording might surface and catch them out. Recordings function as an epistemic backstop: not because we constantly check them, but because their mere availability disciplines testimony, the way an unused referee still shapes a game. For roughly a century — since photography and audio became hard to fake — we have built our norms of public truth-telling on this backstop without ever naming it. Deepfakes — convincingly fabricated video and audio, generated by the same machine-learning wave as §3 — dissolve it from both sides. They flood the channel with convincing fakes and, just as corrosively, they hand every caught wrongdoer a new escape: that recording of me is probably a deepfake. The "liar's dividend." Once any recording can be waved away, the backstop stops disciplining testimony — and testimony itself, our single largest source of knowledge, loses a support we didn't know it leaned on. Don Fallis sharpens the same worry in information-theoretic terms ("The Epistemic Threat of Deepfakes," Philosophy & Technology, 2021): deepfakes reduce the information a video carries about what actually happened, degrading it as a signal.

This is impactful precisely because it reframes deepfakes as an epistemological problem, not just a fraud or privacy one: the threat isn't only the specific lies, it's the erosion of a background condition for trusting recordings at all. But — true to this appendix — the magnitude is contested, and the pushback is sharp and worth taking seriously. Joshua Habgood-Coote ("Deepfakes and the epistemic apocalypse," Synthese, 2023) argues the doom framing is overblown: we have never relied on recordings as infallible, we already cross-check testimony against many sources, and societies have absorbed media-manipulation panics before. Atencia-Linares and Artiga ("Deepfakes, shallow epistemic graves," Synthese, 2022) defend the residual epistemic robustness of photography and video. So the mechanism Rini names — recordings as a silent regulator of testimony — is a genuine and illuminating contribution; the prediction of an "epistemic apocalypse" or wholesale collapse of public knowledge is a live dispute, not a settled forecast. Bring this one forward to Day 2, where science's answer to "how do you trust a report you can't personally verify?" is a whole institutional machinery of replication and recording — and to the AI block, where it meets §3 head-on.

§5 The adversarial turn

Hostile epistemology: when the environment is built to fool you

Traditional epistemology pictured a lone, neutral mind facing a neutral world. C. Thi Nguyen's program — he calls it hostile epistemology — starts from a darker and more modern premise: your epistemic environment is not neutral. It is increasingly engineered, often by parties with an interest in what you come to believe, to exploit the predictable shortcuts your mind must use. Three of his post-2020 moves have reshaped how the whole field talks about online life.

The first is a distinction that sounds academic and turns out to be the key to everything: the difference between an epistemic bubble and an echo chamber ("Echo Chambers and Epistemic Bubbles," Episteme, 2020). They are not the same thing, and conflating them is why so many well-meaning fixes fail. In a bubble, outside voices are merely absent — you simply haven't been exposed to them (think a filter that only ever shows you agreeable sources). In a chamber, outside voices are present but actively discredited — you've been trained to distrust them in advance ("the mainstream media lies," "experts are corrupt"). The consequence is stark and counterintuitive, and the interactive below lets you feel it: the obvious intervention — expose people to the other side — pops a bubble but can strengthen a chamber, because inside a chamber, encountering the enemy's argument is exactly what the chamber predicted, and so confirms it.

The second move names a vulnerability inside your own head. In "The Seductions of Clarity" (Royal Institute of Philosophy Supplement, 2021), Nguyen argues that the feeling of clarity — that satisfying click when everything seems to fall into place — functions as a thought-terminating heuristic. We use the sense that a matter has become clear as a signal that we've inquired enough and can stop. Usually fine. But it means clarity can be weaponized: a manipulator who can manufacture an exaggerated sense of clarity — a tidy ideology that explains everything, a conspiracy theory where every fact slots satisfyingly into place — can get you to halt your inquiry early, before you notice the holes. Notice how this snaps together with §1: clarity is dangerous precisely because it terminates the zetetic process. The slickest, most "it all makes sense now" account is, for that very reason, the one to interrogate hardest. The third move completes the toolkit: in "Trust as an Unquestioning Attitude" (Oxford Studies in Epistemology, 2022), Nguyen analyzes trust itself as the stance of not questioning — of taking something as a settled background you build on without re-checking. Indispensable (you can't re-derive everything from scratch), and exactly therefore exploitable: capture what someone trusts unquestioningly, and you've captured where they'll never think to look.

The honest tag here is twofold. The conceptual contributions — bubble vs. chamber, clarity as inquiry-terminating, trust as unquestioning — have been rapidly and widely adopted because they're genuinely clarifying and action-guiding. But two cautions earn their chips. First, philosophers are already pushing on the construct itself (Carey & Ventham, "There is no fresh air: a problem with the concept of echo chambers," Episteme, 2025). Second — and this is a hype-filter point the course insists on — the empirical social-science picture of how prevalent real-world echo chambers actually are is genuinely mixed; several large studies find most people's media diets are more varied than the "sealed echo chamber" image suggests. So treat the conceptual machinery as a sharp and durable tool, and the empirical scale of the phenomenon as an open measurement question. The framework is the contribution; the size of the fire it describes is still being measured.

§6 The formal refoundation

Re-deriving Day 1's Dutch book — from truth, not money

On Day 1 we justified the laws of probability with a bribe. The Dutch book theorem showed that if your credences break the probability rules, a clever bookie can sell you a set of bets you each accept as fair but which together guarantee you lose money. Powerful — but faintly unsatisfying as epistemology. Who cares about money? Shouldn't a belief be irrational for some reason to do with truth, not with your wallet? A research program that matured through the 2010s and is in full bloom now — accuracy-first epistemology, also called epistemic utility theory — answers exactly that, and it's one of the most elegant results in the modern subject.

The idea (seeded by James Joyce's 1998 "Nonpragmatic Vindication of Probabilism" and built out by Richard Pettigrew's Accuracy and the Laws of Credence, 2016, with a wave of 2020–2023 papers refining and contesting it) is to measure how good a set of credences is by a single epistemic yardstick: accuracy, its closeness to the truth. Full confidence in a truth is perfectly accurate; full confidence in a falsehood, maximally inaccurate. Now the theorem. For any incoherent credence — one that violates the probability laws — there is guaranteed to exist a coherent credence that is more accurate in every possible world at once. The incoherent one is, in the technical term, accuracy-dominated: strictly beaten on truth-closeness no matter how things turn out. So you don't need the bookie at all. Incoherent confidence is irrational for a purely epistemic reason — it leaves guaranteed accuracy on the table; there's a better-aimed set of credences available that's closer to the truth whatever the world does. The interactive lets you see the domination geometrically.

What makes this a frontier rather than a footnote: it's an attempt to rebuild the foundations of rationality on a single epistemic value — getting close to the truth — and to derive not just probabilism but the update rule (conditionalization) and more besides from accuracy-dominance arguments. If it fully succeeds, the whole Bayesian edifice we started sketching on Day 1 rests on truth, not on betting behavior or psychology. The chip, though, is earned. The core dominance theorem for probabilism is established mathematics. The ambition — that all epistemic norms fall out of accuracy alone — is contested: the cleanest results lean on technical assumptions (additivity, finitely many propositions) that critics argue smuggle in more than pure "closeness to truth" warrants (Chad Marxen, "Epistemic utility theory's difficult future," Synthese, 2021), and rival measures of accuracy can deliver different verdicts. So: a beautiful, genuinely illuminating reframing with a rock-solid center and a contested perimeter — which is, fittingly, the exact shape of this entire appendix.

Open questions

What the edge of the map leaves blank

• Is inquiry the real unit? Do the norms of seeking truly conflict with the norms of believing — and if so, which is fundamental?
• Knowledge or belief first? Is "factive theory of mind" the basic cognitive tool, with belief a later, costlier add-on — or is the knowledge/belief line itself too crisp?
• What do machines produce? Knowledge, assertion, testimony, instrument-readings, or a genuinely new category of truth-apt text with no one home who cares?
• Friction or collapse? Do deepfakes merely add cost to verifying recordings, or dissolve a load-bearing condition of public knowledge?
• How engineered is your mind's environment — and how big, really, are the echo chambers we can now so clearly describe?
• Can truth alone ground rationality? Does accuracy-first reach all the way, or only as far as its technical assumptions carry it?
• And a quieter contender for a future day: several of these point past knowledge toward understanding as the thing we actually prize — a pivot we'll feel again whenever a model can predict without explaining.