Susan Haack

Foreword

By Edo Segal

The question I stopped asking was the one that mattered most: How do I know this is true?

Not whether it sounds true. Not whether it reads well. Not whether it fits the argument I'm building. Whether it is actually, verifiably, checkably true.

I stopped asking because Claude made it easy to stop. The output arrives polished, confident, internally consistent. It connects ideas with a fluency that feels like understanding. And somewhere in the months of collaboration that produced *The Orange Pill*, I developed a dangerous habit. I started trusting the grid without checking the clues.

I describe in the book the moment I caught Claude attributing a concept to Deleuze that Deleuze never proposed. The passage was elegant. It connected two threads beautifully. I kept it for a day before something nagged. I checked. It was wrong — not obviously wrong, not absurdly wrong, but wrong in the specific way that matters most: it was coherent without being grounded. It hung together perfectly while touching nothing real.

That experience sent me looking for a framework. Not a technical one — I have those. An epistemological one. A way to understand why coherence without grounding is dangerous, why fluency is not evidence of truth, and why the discipline of checking matters more now than at any previous moment in human history.

Susan Haack built that framework. Her foundherentism — the argument that knowledge requires both experiential anchoring and mutual coherence among beliefs, held together like a crossword puzzle where every entry must match its clue and intersect correctly with its neighbors — is the most precise diagnostic tool I have found for understanding what AI does to how we know things.

Haack is not writing about AI. Her major epistemological work predates this moment by decades. But her insistence that coherence alone is not enough, that the smooth internal consistency of a belief system means nothing if the system is not tethered to reality, lands with extraordinary force when the most coherent voice in the room is a machine that has never observed anything.

This book walks through her framework and applies it, rigorously, to the epistemic crisis that AI has created. It will not tell you whether AI is good or bad. It will give you the tools to evaluate what AI produces — to check the clues, verify the intersections, and maintain the distinction between knowledge and its increasingly convincing counterfeits.

The crossword puzzle is never finished. But you need to understand the rules before you pick up the pen.

— Edo Segal ^ Opus 4.6

About Susan Haack

1945–

Susan Haack (1945–) is a British-American philosopher whose work spans epistemology, philosophy of logic, philosophy of science, and legal theory. Born in England and educated at the University of Cambridge and Oxford, she has spent the majority of her academic career at the University of Miami, where she is Distinguished Professor in the Humanities, Cooper Senior Scholar in Arts and Sciences, and Professor of Philosophy and Professor of Law. Her most influential contribution is foundherentism, a theory of epistemic justification developed in *Evidence and Inquiry: Towards Reconstruction in Epistemology* (1993), which integrates the experiential grounding of foundationalism with the mutual-support structure of coherentism, using the crossword puzzle as its central analogy. Her other major works include *Philosophy of Logics* (1978), *Deviant Logic, Fuzzy Logic: Beyond the Formalism* (1996), *Manifesto of a Passionate Moderate* (1998), *Defending Science — Within Reason* (2003), and *Putting Philosophy to Work* (2008). Deeply influenced by the pragmatist tradition of Charles Sanders Peirce, Haack is known for her rigorous defense of genuine inquiry against both cynical relativism and dogmatic scientism, her taxonomy distinguishing genuine inquiry from sham and fake inquiry, and her insistence that intellectual integrity is characterological rather than merely procedural. Her student Luciano Floridi has credited her *Philosophy of Logics* as a formative intellectual influence. Haack's work has gained renewed relevance in the age of AI, where her framework for evaluating the relationship between coherence, evidence, and justified belief provides critical tools for navigating a world saturated with machine-generated claims.

Chapter 1: The Epistemic Landscape Before AI

Western epistemology arrived at the AI moment carrying two thousand years of unresolved argument about the simplest possible question: What does it take for a belief to count as knowledge?

The question sounds academic. It is not. Every time a software engineer accepts Claude's output without checking it, every time a lawyer submits an AI-drafted brief without reading the cases it cites, every time a student pastes a model-generated paragraph into an essay and moves on, that person is making an epistemic decision — a decision about what counts as justified belief and what does not. They are answering the oldest question in epistemology, usually without knowing it, and usually badly.

The philosophical tradition they are unwittingly betraying has been organized, since Descartes, around two competing architectures for epistemic justification. Understanding both — and understanding why both fail when confronted with AI — is the prerequisite for understanding what Susan Haack built in their place and why that construction matters more now than at any point in the four decades since she first proposed it.

The first architecture is foundationalism. Its central metaphor is a building. Knowledge rests on a foundation of basic beliefs — beliefs that are self-evident, incorrigible, or directly justified by sensory experience. These basic beliefs need no further justification. They are where the chain of reasons terminates. All other beliefs are justified by their inferential connection to these foundational beliefs, the way upper floors of a building are supported by the floors beneath them, which are in turn supported by the foundation.

René Descartes launched the foundationalist project in the seventeenth century with a specific ambition: to find beliefs so certain that no rational person could doubt them, and to rebuild the entire edifice of knowledge on that bedrock. His famous cogito ergo sum was the first candidate — a belief that survived even the most radical skeptical assault. From there, through a series of arguments that have been contested ever since, he attempted to rebuild physics, mathematics, and knowledge of the external world on the certainty of his own existence as a thinking thing.

The logical positivists of the Vienna Circle pursued a related program in the early twentieth century. Their version of foundationalism rested not on Cartesian certainty but on the verification principle: a statement is meaningful only if it can be verified through sensory experience. The foundation was empirical observation. The structure built upon it was the edifice of scientific knowledge. Everything else — metaphysics, theology, ethics — was dismissed as literally meaningless, not false but without cognitive content.

The appeal of foundationalism is architectural. If the foundation holds, the building stands. If every belief in the structure can trace its justification back to bedrock, then the structure as a whole inherits the security of its base. The foundationalist sleeps well because the building is anchored.

The problem, which has occupied epistemologists for three centuries, is that the foundation never holds the weight the foundationalist places on it. Sensory experience, the most popular candidate for basic belief, turns out to be theory-laden — shaped by the concepts, expectations, and prior beliefs of the perceiver. The duck-rabbit illusion is the textbook example, but the problem runs deeper than optical tricks. A trained radiologist sees a tumor on a scan that a layperson sees as noise. The observation itself is structured by the observer's theoretical commitments. If the foundation is supposed to be theory-neutral, and observation is theory-laden, then the foundation is compromised before the first floor goes up.

Introspection fares no better. Decades of psychological research have demonstrated that human beings are unreliable reporters of their own mental states. Self-evidence is culturally contingent — what seems obvious to one tradition is bizarre to another. Every candidate for foundational bedrock, examined with sufficient care, turns out to be a human construction rather than a discovered feature of reality. The building rests not on rock but on beliefs about rock, which is a different thing entirely.

The second architecture is coherentism. Its central metaphor is a web. There are no basic beliefs. There is no foundation. Beliefs are justified not by their connection to bedrock but by their coherence with the total web of beliefs the person holds. A belief is justified to the extent that it fits — that it is consistent with other beliefs, supported by them, mutually reinforcing. The web has no bottom. It has only connections.

W.V.O. Quine's holism provided the most influential version of this view. In "Two Dogmas of Empiricism," Quine argued that no statement is immune to revision in the face of recalcitrant experience, and that the unit of empirical significance is not the individual statement but the whole of science. Beliefs face the tribunal of experience not individually but as a corporate body. When experience conflicts with the web, any number of adjustments might resolve the conflict — the choice of which belief to revise is pragmatic, not dictated by logic alone.

Coherentism solves the foundationalist's problem by dissolving it. If there are no basic beliefs, there is no need for an unshakable foundation. The web holds itself up, each strand supporting the others, the strength of the whole distributed across the structure rather than concentrated at the base.

The appeal is elegance. The vulnerability is devastating.

A perfectly coherent system of beliefs can be perfectly false. This is not a marginal concern. It is the structural deficiency at the heart of the coherentist project. A well-constructed novel is internally coherent — the characters behave consistently, the plot follows from its premises, the world makes sense on its own terms. It is also fiction. A paranoid conspiracy theory can achieve remarkable internal coherence — every objection is absorbed into the theory, every piece of counter-evidence reinterpreted as confirmation. The coherence is genuine. The beliefs are false.

Laurence BonJour, in The Structure of Empirical Knowledge, attempted to address this vulnerability by adding an "Observation Requirement" — the demand that a coherent system must include beliefs formed in response to observational experience. This was, as Haack later observed with characteristic precision, an implicit admission that pure coherentism was inadequate — that coherence required something outside itself to connect the web to reality. BonJour was smuggling a foundationalist element into a coherentist framework and hoping nobody would notice.

These two architectures — the building and the web, the foundation and the fabric — constituted the landscape of epistemological thought for centuries. Philosophers aligned with one or the other, proposed modifications, defended variants, and produced a literature of extraordinary sophistication that, for all its intellectual grandeur, never resolved the fundamental tension: knowledge seems to require both grounding in experience and coherence among beliefs, and neither architecture, taken alone, provides both.

This unresolved tension was, for most of the history of philosophy, a matter of theoretical interest. The practical consequences of choosing wrongly between foundationalism and coherentism were, for most human activities, negligible. A scientist who implicitly relied on foundationalist intuitions about observation and a scientist who implicitly relied on coherentist intuitions about theoretical unity could both do excellent science, because both were checking their beliefs against reality through the ordinary practices of empirical inquiry — designing experiments, collecting data, subjecting conclusions to peer review. The philosophical architecture mattered less than the practical discipline.

The arrival of artificial intelligence changed this calculation. It changed it because AI produces output that instantiates one half of the epistemological equation — coherence — while entirely lacking the other half — experiential grounding. And it produces this half-justified output at a scale, speed, and fluency that overwhelms the ordinary human practices of verification.

A large language model is, in epistemological terms, a pure coherence engine. Its training objective is to predict the next token in a sequence — to generate output that is statistically consistent with the patterns in its training data. The output is coherent in the precise sense that coherentists value: internally consistent, mutually reinforcing, contextually appropriate. When Claude generates a paragraph about constitutional law, the sentences cohere with one another and with the broader patterns of legal reasoning present in the training corpus. When it generates a passage about molecular biology, the claims fit together in ways that mirror the structure of actual biological knowledge.

The coherence is real. It is not an illusion or an accident. The system has been optimized, through billions of parameters and extraordinary computational resources, to produce output that hangs together. And it succeeds spectacularly. The output reads as though a knowledgeable person produced it, because the surface features of knowledge — logical structure, appropriate vocabulary, contextual sensitivity, narrative coherence — are exactly the features the system has been trained to reproduce.

What the system has not been trained to do is check its output against reality. It has no sensory experience. It has made no observations. It has conducted no experiments. It has never touched, seen, heard, smelled, or tasted anything. Its relationship to the external world is mediated entirely through the text of its training data, which is itself a record of other people's expressions — expressions that may or may not accurately reflect those people's experiences, which may or may not accurately reflect reality.

The foundationalist, confronted with AI output, looks for the foundation and finds nothing. There is no bedrock beneath the model's claims — no basic beliefs justified by direct experience, no observational base from which the superstructure of the output can be derived. The training data is not a foundation in any epistemologically meaningful sense. It is a statistical sample of human linguistic behavior, which is several inferential steps removed from the experiential grounding that foundationalism requires.

The coherentist, confronted with the same output, finds the coherence satisfying and has no principled basis for complaint. The output coheres. The web holds together. If coherence is the criterion of justification, then AI output is well-justified — better justified, in many cases, than the beliefs of individual human beings, whose webs of belief are riddled with inconsistencies, gaps, and contradictions that the model's output conspicuously avoids.

This is the epistemological crisis. The two dominant frameworks for evaluating knowledge claims each fail, in their own way, when applied to AI. Foundationalism demands a grounding the system does not possess. Coherentism is satisfied by a feature the system possesses in abundance but that is, by itself, insufficient to distinguish knowledge from sophisticated fabrication.

The crisis is not theoretical. It is operational. Every person who uses an AI tool to generate content, draft arguments, summarize research, or produce analysis is making epistemic decisions within this broken landscape. They are evaluating AI output using implicit epistemological frameworks that cannot accommodate what the output actually is — a new kind of epistemic object that satisfies one criterion of justified belief while violating the other.

Luciano Floridi, who studied under Haack and credits her Philosophy of Logics as the turning point in his intellectual life, identified a related structural problem in his work on the philosophy of information. Floridi argued that AI represents a "divorce between agency and intelligence" — that machines can solve problems successfully without being intelligent in any meaningful sense. The behavior that would be called intelligent if performed by a human is not intelligent when performed by a machine, because the machine lacks the understanding that gives human intelligence its epistemic character. A dishwasher cleans dishes as well as a human does. It does not clean them the way a human does. The output is equivalent. The process is categorically different.

This divorce between output equivalence and process difference is precisely what makes the epistemological crisis so acute. The AI's output looks like knowledge. It reads like knowledge. It functions like knowledge in many practical contexts. But the process that produced it is not the process that produces knowledge — not by foundationalist standards, not by coherentist standards, not by any epistemological standard that requires a connection between the claim and the reality the claim purports to describe.

Susan Haack saw this structural problem decades before AI made it urgent. Not the AI problem specifically — her major epistemological work predates the current moment by thirty years. But the underlying structural problem: that neither foundationalism nor coherentism alone can account for what genuine knowledge requires, and that any adequate theory of justification must incorporate both experiential grounding and mutual coherence, held together in a single framework that does not reduce either to the other.

The framework she built — foundherentism — is the subject of the next chapter. But its relevance to the AI moment should already be coming into focus. The crisis is not that we lack a philosophical framework for evaluating AI output. The crisis is that the frameworks most people implicitly rely on — the intuitive foundationalism of "Is this based on real data?" and the intuitive coherentism of "Does this sound right?" — are each, taken alone, insufficient. The AI age demands a framework that requires both. Haack built one. The question is whether the culture will adopt it before the damage of operating without it becomes irreversible.

Chapter 2: Foundationalism and Its Discontents

Foundationalism promises something that human beings crave: a place to stand. A point of certainty from which the rest of knowledge can be derived. A bedrock that does not shift. The appeal is not merely intellectual. It is existential. In a world of competing claims, conflicting evidence, and proliferating uncertainties, the foundationalist offers the comfort of an anchor.

Haack understood this appeal even as she dismantled the architecture that supported it. Her critique of foundationalism in Evidence and Inquiry is not a dismissal. It is a diagnostic exercise — an attempt to identify precisely what foundationalism gets right and precisely where it fails, so that the right elements can be preserved in a better framework.

What foundationalism gets right is the insistence on experiential grounding. The intuition that knowledge must be connected to reality, that a belief system floating free of experience is epistemically weightless, is sound. Haack never disputed this. Her disagreement with foundationalism was never about the importance of grounding. It was about the foundationalist's account of how grounding works — the claim that there exist basic beliefs, justified by experience alone, that need no support from other beliefs and can serve as the foundation for everything else.

The concept of the basic belief is where the architecture begins to fracture. A basic belief, in the foundationalist's framework, must satisfy two requirements simultaneously. It must be justified — otherwise it cannot serve as a foundation for other beliefs. And it must be self-justifying — justified without reliance on any other belief — otherwise it would need its own foundation, and the regress would continue.

Descartes' cogito satisfied both requirements, or seemed to. The belief that one exists as a thinking thing appeared to be both justified and self-justifying — a belief whose truth is guaranteed by the very act of entertaining it. But the cogito proved to be a remarkably narrow foundation. Descartes needed God to bridge the gap between the certainty of his own existence and knowledge of the external world, and the arguments for God's existence that he deployed were precisely the kind of inferential reasoning that the foundation was supposed to ground, not presuppose.

The logical positivists attempted a different foundation: direct sensory experience. Protocol sentences — sentences recording immediate observations — were to serve as the basic beliefs on which scientific knowledge was built. Rudolf Carnap's Aufbau was the most ambitious attempt to construct the entirety of empirical knowledge from a base of elementary experiences. The project failed, as Carnap himself eventually acknowledged, because the base could not bear the weight. The translation from experiential input to theoretical output required inferential machinery that the base could not supply.

The deeper problem, which Haack articulated with particular clarity, is that experience does not arrive in propositional form. Experiences are not beliefs. The foundationalist needs to get from experience — a sensory state, a perceptual event — to a belief about that experience, and the transition is not self-justifying. When a person sees a red object and forms the belief "There is something red before me," the belief is not identical to the experience. It is an interpretation of the experience, shaped by concepts the person already possesses, by expectations formed through prior experience, and by the theoretical framework within which the perception occurs.

This is the theory-ladenness problem, and it is fatal to the strong foundationalist program. If observation is shaped by theory, then observational beliefs cannot serve as a theory-neutral foundation. The foundation is contaminated by the very superstructure it is supposed to support. The building does not rest on bedrock. It rests on materials drawn from its own upper floors.

Haack's analysis of this problem is characteristically precise. She does not conclude that experience is irrelevant to justification — that would be the coherentist's overreaction, throwing the experiential baby out with the foundationalist bathwater. She concludes that experience plays a causal role in the formation of beliefs without playing a logical role in their justification. Experience causes the person to form certain beliefs. Those beliefs are then justified (or not) by their relationship to the person's total body of evidence — including other beliefs, other experiences, and the coherence of the whole. Experience is an input to the justificatory process, not a foundation beneath it.

This distinction — between the causal role of experience and the logical structure of justification — is the hinge on which Haack's entire epistemological project turns. And it is the distinction that makes her framework uniquely relevant to the AI moment.

When a language model generates output, there is no experience at the causal origin. No perception. No observation. No sensory encounter with the world. There is training data — a vast corpus of human-generated text — and there is the statistical processing of that data into patterns that the model reproduces with extraordinary fidelity. The training data is not the model's experience in any sense that epistemology can recognize. It is a record of other people's linguistic outputs, which are themselves several inferential steps removed from the experiences that may or may not have grounded them.

Consider the inferential chain. A scientist observes a phenomenon. She forms a belief about it. She writes a paper describing her findings. The paper is published. The text of the paper enters the training corpus. The model processes the text and learns the statistical patterns. The model generates output that reproduces those patterns. The user reads the output and forms a belief.

Between the original observation and the user's belief lie at least five inferential steps, each introducing the possibility of error, distortion, or fabrication. The scientist's observation was theory-laden. Her description of it was shaped by the conventions of scientific writing. The paper may have contained errors. The training corpus includes texts of wildly varying reliability. The model's statistical processing preserves surface patterns without preserving the evidential relationships that gave the original observation its epistemic value.

The foundationalist examining this chain searches for bedrock and finds only sediment — layers of processed, reprocessed, and statistically recombined text, none of which constitutes the kind of direct experiential grounding that foundationalism requires.

The temptation, for those who wish to maintain a foundationalist framework while accommodating AI, is to treat the training data itself as a kind of foundation. The data is real. It exists. It was produced by human beings who had experiences. Therefore — the argument goes — the model's output is indirectly grounded in human experience, mediated through the training corpus.

This argument has the structure of wishful thinking. The training data is not a record of human experience. It is a record of human expression — a fundamentally different thing. Human beings express things they do not believe, believe things they do not express, and express beliefs that are false. The corpus includes fiction, satire, propaganda, error, fabrication, and the full spectrum of human linguistic behavior, much of which has no stable relationship to experiential reality. A model trained on this corpus inherits not the experiences of its authors but the statistical patterns of their words. The patterns are real. The grounding is not.

The situation is worse than this, because the model does not merely reproduce the patterns of its training data. It recombines them in novel ways — generating sentences, paragraphs, and arguments that do not appear verbatim in any training document. These recombinations are coherent. They are often insightful. They are sometimes brilliant. But they are generated by a process that has no mechanism for checking whether the recombination corresponds to anything in reality. The model does not know whether its novel claims are true. It does not know what "true" means. It generates output that is statistically likely given the patterns in its training data, and statistical likelihood is not truth, though the two can be correlated under favorable conditions.

Haack's critique of fuzzy logic, developed in Deviant Logic, Fuzzy Logic: Beyond the Formalism, is relevant here in an unexpected way. Fuzzy logic — which became foundational to early AI systems in the 1980s and 1990s — proposed that truth comes in degrees, that a statement can be partially true and partially false. Haack argued that fuzzy logic confused the vagueness of predicates with the logic of truth-values, producing a formal system that was mathematically interesting but philosophically muddled. The critique was, at its core, about the danger of substituting formal elegance for philosophical rigor — of building systems that worked in practice while resting on foundations that could not bear theoretical scrutiny.

The parallel to contemporary language models is direct. The models work in practice. They produce output that is useful, sometimes remarkably so. But the epistemic foundations on which that output rests cannot bear the weight of the confidence that users place on it. The formal elegance of the transformer architecture, like the formal elegance of fuzzy logic, conceals a philosophical gap between what the system does and what the system's users assume it does.

What does this mean for the person evaluating AI output? The foundationalist's prescription — check the foundation, trace the claim back to its experiential base — is sound in principle but nearly impossible in practice when applied to AI. The experiential base does not exist in any accessible form. The training data is proprietary, massive, and unstructured. The statistical processes that transformed the data into the model's output are opaque even to the model's creators. The inferential chain between reality and the model's claim cannot be traced, because it was never a chain — it was a statistical distribution.

This does not mean that AI output is worthless. It means that foundationalism, as a framework for evaluating AI output, is inadequate. The framework demands something that the object of evaluation cannot supply. The user who applies foundationalist standards to AI output will either reject all of it — an overreaction that discards genuine utility — or will construct a fictional foundation beneath it, treating the training data or the company's reputation or the model's confident tone as a surrogate for the experiential grounding that the framework requires. The second option is more common and more dangerous.

Segal described this danger with the builder's vocabulary in The Orange Pill: machines now "speak our language without sharing our experience." The phrase captures the epistemological crisis in miniature. The language is the coherence. The experience is the grounding. The machines have mastered the language. They have no experience. And the person on the receiving end, hearing fluent language that sounds like the product of experience, must resist the natural human tendency to infer grounding from fluency.

The tendency is natural because it is, in most human contexts, reliable. When a person speaks fluently about a topic, the fluency is usually evidence of familiarity, which is usually evidence of experience, which is usually evidence that the person's claims are grounded. The inferential chain from fluency to grounding has served human beings well for seventy thousand years of linguistic interaction. It fails catastrophically when applied to a system that has mastered fluency without possessing experience.

Foundationalism cannot diagnose this failure, because foundationalism has no account of what happens when the surface features of grounded knowledge are present and the grounding itself is absent. The framework was built for a world in which the only sources of knowledge-shaped output were beings that had experiences. It was not built for a world in which coherence engines produce output that mimics the products of experience without possessing any.

The diagnosis requires a different framework — one that can hold both the importance of experiential grounding and the recognition that grounding must be checked, actively and continuously, against the coherence of the whole. That framework is the subject of the chapters ahead. But before arriving at Haack's positive proposal, the other tower of epistemological thought requires examination, because it is the tower that AI most closely resembles — and the tower whose vulnerabilities AI most ruthlessly exploits.

Chapter 3: Coherentism and Its Temptations

If foundationalism asks where knowledge begins, coherentism asks whether the question makes sense. The coherentist's answer is that it does not. Knowledge has no beginning. It has no bedrock, no first principle, no basic belief that stands independent of all other beliefs. There are only beliefs, connected to other beliefs, supporting and constraining one another in a web whose strength is the strength of the whole.

The metaphor is powerful precisely because it dispenses with the foundationalist's most troublesome requirement. No self-justifying belief is needed. No incorrigible experience. No theory-neutral observation. The web holds itself together through the mutual support of its strands, each one deriving its justification from its connections to the others. The architecture is democratic: no belief occupies a privileged position. Every strand bears weight, and every strand is supported by the others.

Quine's version of holism, articulated most famously in "Two Dogmas of Empiricism," pressed this vision to its logical conclusion. The unit of empirical significance, Quine argued, is not the individual statement but the totality of science. When experience conflicts with the web, the conflict can be resolved by revising any number of beliefs — peripheral beliefs about particular observations, central beliefs about logical laws, or anything in between. The choice of what to revise is guided by pragmatic considerations: simplicity, conservatism, explanatory power. But no belief is immune to revision, and no belief is so peripheral that its revision could not, in principle, ripple through the entire web.

The appeal of this view to a generation of philosophers exhausted by the foundationalist's failed search for certainty was considerable. Coherentism acknowledged what foundationalism could not: that human knowledge is a social, historical, evolving enterprise, not a building erected once on secure foundations but a living web continuously rewoven in response to new experience and new argument. The view felt honest. It felt realistic. It felt like philosophy finally catching up to how knowledge actually works.

Haack's engagement with coherentism was, characteristically, more surgical than dismissive. She identified what coherentism gets right — the recognition that beliefs are justified in part by their relationships to other beliefs, that no belief is fully justified in isolation, that the web of belief is a real and important epistemic structure. And she identified, with devastating precision, what coherentism gets wrong.

The problem is captured in a thought experiment that requires no philosophical training to understand. Imagine a novelist of extraordinary skill — someone capable of constructing a fictional world with perfect internal consistency. Every character behaves in accordance with their established psychology. Every event follows from prior events according to the laws of the fictional universe. Every detail coheres with every other detail. The novel is a masterpiece of internal consistency.

It is also fiction. Every claim in the novel is false. No character exists. No event occurred. The coherence is perfect, and the truth-value is zero.

The coherentist owes an account of what distinguishes this fictional web from a web of genuine knowledge. Both are internally coherent. Both exhibit mutual support among their elements. The difference, obviously, is that the web of genuine knowledge is connected to reality and the fictional web is not. But "connected to reality" is precisely the foundationalist's requirement — the requirement that coherentism was designed to eliminate.

This is the isolation objection, and it has haunted coherentism since its inception. BonJour, the most sophisticated coherentist of the late twentieth century, responded by adding his Observation Requirement: a coherent system of beliefs must include beliefs that are, as a matter of empirical fact, caused by observational experiences. The requirement was meant to tether the web to reality without conceding the foundationalist's claim that some beliefs are justified by experience alone.

Haack's response to BonJour's maneuver was pointed. The Observation Requirement is a foundationalist element smuggled into a coherentist framework. If the system must include observationally caused beliefs, then observational experience is playing a role in justification that coherentism cannot account for using only coherentist resources. BonJour was acknowledging, implicitly, that coherence alone is insufficient — that the web needs anchors, not just connections. But having acknowledged this, he had no coherentist mechanism for explaining how the anchors work. He had conceded the argument to foundherentism without quite admitting it.

The relevance of this philosophical exchange to the AI moment is not abstract. It is immediate and practical, because large language models are, in epistemological terms, the purest coherentism engines ever built.

A language model's entire architecture is oriented toward coherence production. The transformer, the computational structure underlying modern language models, processes input through attention mechanisms that relate every element of a sequence to every other element, identifying patterns of co-occurrence, mutual implication, and contextual fit. The model's training objective — next-token prediction — rewards output that is statistically consistent with the patterns in the training data. Consistency is coherence. The model is optimized, at every level of its architecture, to produce output whose elements support one another.

And it succeeds. The output is coherent in ways that are, from a purely coherentist perspective, remarkable. When Claude generates a legal analysis, the claims follow logically from one another. When it produces a scientific explanation, the concepts are deployed consistently and the reasoning proceeds in an orderly fashion. When it writes a narrative, the plot coheres, the characters behave consistently, and the themes develop in a way that feels integrated rather than arbitrary.

The coherence is not superficial. It operates at multiple levels simultaneously — lexical, syntactic, semantic, logical, and narrative. The model has learned, through its training process, the deep statistical regularities of human knowledge-expression, and it reproduces those regularities with a fidelity that makes its output, in many contexts, indistinguishable from the products of genuine human understanding.

This is the temptation. The coherentist, encountering AI output, finds nothing to object to. The web holds together. The strands support one another. If coherence is the criterion of justification, then the output is justified — better justified, in many cases, than the beliefs of individual human beings, whose webs of belief contain contradictions, gaps, and inconsistencies that the model's output smoothly avoids.

The temptation is magnified by a feature of human cognition that psychologists call the fluency heuristic. Human beings tend to judge the truth of a claim partly by how easily it is processed — how smoothly it reads, how naturally it fits with expectations, how little cognitive effort it requires. Fluent claims feel true. The feeling is not reliable, but it is persistent and powerful, and it operates below the threshold of conscious deliberation.

AI output is engineered for fluency. The optimization that produces coherent output simultaneously produces output that is easy to process, natural-sounding, and subjectively compelling. The fluency heuristic, encountering output optimized for exactly the features it responds to, fires with unusual intensity. The output feels true because it is smooth. The smoothness is not evidence of truth. It is a design feature.

Here the connection to Byung-Chul Han's critique of smoothness, as examined in The Orange Pill, acquires a specifically epistemological dimension. Han argued that the aesthetics of the smooth — the frictionless, the seamless, the polished — conceals the absence of the depth that friction produces. Haack's epistemology provides the philosophical mechanism for understanding why this concealment works. The smooth output satisfies the coherence condition of justification. Coherence is a genuine epistemic virtue. The satisfaction is not illusory. But it is incomplete, because coherence without grounding is justification without anchor — a web that holds together beautifully while floating free of reality.

The novelist's fiction and the model's confabulation are structurally identical in the relevant respect. Both exhibit internal coherence. Both lack experiential grounding. The difference is that the reader of the novel knows it is fiction. The user of the AI model may not know that the output is ungrounded, because the surface features of grounded knowledge — logical structure, appropriate vocabulary, contextual sensitivity — are indistinguishable from the surface features of coherent confabulation.

Segal captured this danger in the phrase "confident wrongness dressed in good prose." The confidence is the coherence. The wrongness is the absent grounding. The good prose is the fluency that makes the absence invisible. The phrase is foundherentism in miniature, though Segal arrived at it through the experience of building rather than through epistemological theory.

The Deleuze incident that Segal describes — Claude generating a passage connecting Csikszentmihalyi's flow theory to a concept attributed to Deleuze that Deleuze never actually proposed — is a textbook demonstration of coherentism's vulnerability. The passage cohered. The connection between flow theory and Deleuzian smooth space was elegant, internally consistent, and rhetorically compelling. Every element of the passage supported every other element. The web held together beautifully.

The clue was wrong. The experiential anchor — Deleuze's actual philosophical position — did not match the entry the model had generated. The entry had been produced entirely by coherence-driven pattern completion, and the coherence was persuasive enough that Segal, who is not a casual reader, accepted it on first pass and only caught the error on the second reading. The web caught what the anchor would have rejected immediately.

This is the operational consequence of coherentism's theoretical vulnerability. In practice, coherence-based evaluation of AI output means accepting claims that fit the web of existing beliefs — that sound right, that connect to what the user already knows, that extend familiar arguments in plausible directions. The evaluation is not meaningless. Claims that conflict with well-established knowledge should be treated with suspicion, and coherence-checking does useful work in filtering out obvious errors.

But the claims that are most dangerous are not the ones that conflict with established knowledge. They are the ones that extend established knowledge in plausible but false directions — the ones that sound like the next logical step in an argument that is, in fact, heading somewhere the evidence does not support. These claims pass the coherence test precisely because they are designed to. The model generates them by extending the statistical patterns of genuine knowledge into regions where those patterns may not hold, and the extensions are smooth enough to evade detection.

Floridi's characterization of AI as a "divorce between agency and intelligence" illuminates the coherentist failure from a different angle. Were a human being to produce claims with this level of coherence, the coherence would be evidence of understanding. A person who speaks fluently about constitutional law has, in the ordinary case, studied constitutional law. The coherence of their claims is grounded in learning, which is grounded in experience. The coherentist's implicit assumption — that coherence reflects understanding, which reflects reality — holds in the human case because, for human speakers, it usually does.

For AI, the assumption fails. The coherence does not reflect understanding. It reflects the statistical regularities of a training corpus. The system has no understanding of constitutional law, molecular biology, or anything else. It has patterns. The patterns produce coherent output. The coherence is real. The understanding is absent. And coherentism, which evaluates beliefs by their coherence alone, has no mechanism for detecting this absence.

The coherentist's temptation, confronted with AI, is to accept the output because it fits. The foundationalist's impulse, confronted with the same output, is to reject it because it lacks a foundation. Both responses are partially correct. Both are inadequate. The output fits but may not be grounded. The output lacks a traditional foundation but may nevertheless connect to reality through the training data's imperfect mediation of human experience.

What is needed is a framework that requires both — that insists on checking the clues and verifying the intersections simultaneously. That refuses to accept coherence as sufficient and refuses to demand a foundation that the system cannot provide. That holds both requirements in tension, neither reducing one to the other, and that places the responsibility for maintaining both squarely on the human evaluator.

Haack built this framework. The next chapter examines it.

Chapter 4: Foundherentism: The Crossword Puzzle of Knowledge

The crossword puzzle sits on the breakfast table. Some squares are filled. Most are empty. The solver picks up a pen and considers 12-Across: "River in Egypt (4 letters)." The clue points toward NILE. She writes it in. Then she checks: Does the N intersect correctly with the down entry? Does the I fit? The L? The E? Each intersection is a constraint. Each constraint either confirms or undermines the answer. If NILE fits the clue and intersects correctly with every crossing entry, the answer is well-justified. If it fits the clue but conflicts with a crossing entry, something is wrong — either with this answer or with the crossing one. If it intersects perfectly but does not match the clue, it is wrong regardless of how well it fits the grid.

Susan Haack proposed this image in Evidence and Inquiry not as a casual illustration but as the central model for how epistemic justification actually works. The crossword puzzle is not a metaphor in the weak sense — a suggestive comparison that captures some features of justification while missing others. It is a structural analogue that captures the essential features of the justificatory process with a precision that neither foundationalism nor coherentism achieves on its own.

The clues are the experiential anchors. They connect belief to reality. A clue is a datum — something given, something that constrains from outside the web of beliefs. It is not self-justifying in the foundationalist's sense, because the same experience can be interpreted differently depending on what other answers are already in the grid. The clue for 12-Across might be ambiguous — "River in Egypt" could conceivably point to something other than NILE if the context were sufficiently unusual. The clue constrains without determining. It narrows the space of acceptable answers without reducing it to one.

The intersecting entries are the coherence constraints. They connect belief to belief. Each answer must cohere with the answers that cross it, and those answers must cohere with the answers that cross them, radiating outward through the grid. A single answer that conflicts with multiple crossings throws doubt not just on itself but on every entry it touches. The coherence of the whole grid is a source of justification for each individual entry — but only in conjunction with the clues, not as a substitute for them.

This is foundherentism: the theory that epistemic justification requires both experiential anchoring and mutual coherence among beliefs, operating simultaneously, each constraining the other. Haack argued that this theory preserves the genuine insights of both foundationalism and coherentism while avoiding the fatal deficiencies of each.

From foundationalism, foundherentism preserves the insistence that knowledge must be connected to experience — that a system of beliefs floating free of reality, no matter how internally coherent, is not knowledge but fantasy. The clues matter. Experience matters. The web must be anchored.

From coherentism, foundherentism preserves the recognition that no belief is justified in isolation — that every belief derives part of its justification from its relationship to other beliefs, and that the web of mutual support is a genuine and important source of epistemic strength. The intersections matter. Coherence matters. The anchors must be connected.

Haack developed foundherentism along three dimensions of evidential quality that determine how well-justified a belief is within the foundherentist framework. These dimensions provide the analytical precision that distinguishes foundherentism from a vague appeal to "use both evidence and coherence."

The first dimension is supportiveness: how well the evidence supports the belief in question. Direct, relevant, unambiguous evidence provides stronger support than indirect, tangential, or ambiguous evidence. In the crossword analogy, a clue that points unambiguously to a single answer provides stronger support than a clue that could be satisfied by several different entries.

The second dimension is independent security: how well-justified the evidence itself is, independently of the belief it supports. Evidence that is itself well-grounded and well-supported by other evidence provides stronger justification than evidence that is shaky, uncertain, or dependent on the very belief it is supposed to justify. In the crossword analogy, an intersecting entry that is itself well-supported by its own clue and its own crossings provides stronger confirmation than an intersecting entry that was guessed at without much confidence.

The third dimension is comprehensiveness: how much of the relevant evidence has been taken into account. A belief that is supported by all the available evidence is better justified than a belief that is supported by some of the evidence while other relevant evidence has been ignored. In the crossword analogy, an answer that fits the clue, intersects correctly with all crossing entries, and is consistent with the theme of the puzzle is better justified than an answer that fits the clue but has not been checked against the crossings.

These three dimensions — supportiveness, independent security, and comprehensiveness — provide a systematic framework for evaluating any claim, from any source, in any context. They are neither purely foundationalist (they do not require self-justifying basic beliefs) nor purely coherentist (they do not accept coherence as sufficient). They require both grounding and coherence, measured across three independent axes, and they locate the work of justification where it has always belonged: in the hands of the evaluator, the person who checks the clues and verifies the intersections.

The application to AI output follows directly, and with a specificity that distinguishes Haack's framework from the vaguer prescriptions commonly offered.

When Claude generates a claim, the claim enters the evaluator's crossword grid as a proposed entry. The first question, corresponding to supportiveness, is: What evidence supports this claim? Not what evidence the model consulted — the model does not consult evidence in any epistemologically meaningful sense. What evidence the evaluator can access independently. Can the claim be checked against a primary source? Against direct observation? Against data that the evaluator has reason to trust? If the answer is yes, and if the evidence supports the claim, then the supportiveness dimension is satisfied. If the answer is no — if the claim cannot be independently verified, if no accessible evidence bears on it — then the supportiveness dimension is unsatisfied, regardless of how fluent or confident the claim appears.

The second question, corresponding to independent security, is: How reliable is the source? This question is complicated in the AI case, because the "source" is not a single entity with a track record but a statistical amalgamation of millions of sources with wildly varying reliability. The model's output inherits the average reliability of its training corpus, weighted by the statistical patterns of the specific domain in question. In domains where the training data is dense, high-quality, and internally consistent — basic science, well-established history, standard legal doctrine — the independent security of the model's output is relatively high. In domains where the training data is sparse, contradictory, or contaminated by misinformation, the independent security drops precipitously. The evaluator who treats AI output as uniformly reliable is making an error that foundherentism's second dimension explicitly identifies.

The third question, corresponding to comprehensiveness, is: Has all the relevant evidence been considered? This is where AI output fails most systematically and most invisibly. The model generates claims based on the patterns it has learned. It does not survey the full range of evidence bearing on a question. It does not seek out disconfirming evidence. It does not weigh the strength of competing considerations. It produces the statistically most likely continuation of the prompt, which is often but not always aligned with the most comprehensive account of the evidence.

The evaluator who relies solely on AI output is violating the comprehensiveness requirement by definition — because the model does not comprehensively evaluate evidence, and the evaluator who does not supplement the model's output with independent research has not comprehensively evaluated evidence either. The comprehensiveness failure is shared between the machine, which cannot perform comprehensive evaluation, and the human, who chooses not to.

This analysis reveals something that neither the AI enthusiasts nor the AI skeptics typically acknowledge: the epistemic quality of AI-assisted work is determined not by the capabilities of the AI but by the epistemic practices of the human user. A model that confabulates routinely can still serve genuine inquiry if the user checks the clues. A model that is accurate ninety-five percent of the time can still undermine genuine inquiry if the user accepts the output without verification, because the five percent that is wrong is indistinguishable, on the surface, from the ninety-five percent that is right.

The crossword puzzle makes this visible. An answer that appears in the grid — proposed by a collaborator, found in a reference book, generated by an AI — is not justified by its source. It is justified by its fit with the clue and its coherence with the intersecting entries, both of which must be checked by the solver. The source may make the proposed answer more or less probable, but probability is not justification. Justification requires the active work of checking, verifying, testing — the epistemic labor that the solver performs and that no source, however reliable, can perform on the solver's behalf.

The crossword metaphor illuminates a further point about how justification works across time and with increasing evidence. As more entries are filled in, the grid becomes more constrained. Each new entry that is well-justified — supported by its clue and confirmed by its crossings — strengthens the justification of the entries around it. The grid becomes self-reinforcing. A partially filled grid in which every entry has been carefully checked is a more reliable source of justification for new entries than an empty grid or a grid filled carelessly.

This has a direct parallel in the practice of human-AI collaboration. The user who has carefully verified the first ten claims in a collaborative document has built a partially filled grid — a web of confirmed beliefs that can serve as reliable crossings for subsequent claims. The eleventh claim can be checked not only against its own clue (independent evidence) but against the ten verified claims that surround it. If it coheres with all of them, the justification is strengthened. If it conflicts with any, the conflict is a signal — a signal that either the new claim or one of the verified claims needs reexamination.

Segal described a version of this practice without using Haack's vocabulary. His discipline of checking AI-generated content — the practice he developed through months of collaboration with Claude — was a process of building a partially filled grid: verifying some claims independently, using the verified claims to evaluate subsequent claims, catching errors by noticing conflicts between new output and previously established facts. When he caught the Deleuze fabrication, the mechanism was precisely this: a crossing entry (his own knowledge of Deleuze's actual philosophical positions) conflicted with the AI-generated entry, and the conflict triggered reexamination. The clue did not match. The entry, however elegant, was wrong.

Haack's foundherentism provides the epistemological explanation for why this practice works when it does and fails when it does not. It works when the evaluator checks both clues and crossings — when independent evidence is consulted and coherence with established beliefs is verified. It fails when the evaluator relies on only one dimension — accepting claims because they cohere with prior beliefs (coherentism without anchoring) or rejecting claims because they lack a visible foundation (foundationalism without coherence-checking).

The framework also explains why the failure mode is asymmetric. Relying solely on coherence — accepting AI output because it sounds right, because it fits the narrative, because it extends the argument in a plausible direction — is more dangerous than relying solely on grounding, because the model is optimized for coherence production. The output will almost always pass the coherence test. The clue-checking is the bottleneck. The human contribution, in the foundherentist framework, is overwhelmingly on the anchoring side — bringing the experiential evidence, the independent knowledge, the domain expertise that the model does not possess.

This asymmetry defines the practical program that Haack's framework implies: an epistemic practice centered on anchor-checking rather than coherence-checking, because the AI handles coherence and the human must handle grounding. The practice is demanding. It requires the evaluator to possess or seek out independent knowledge of the domain in question. It requires the willingness to distrust fluent, confident, internally consistent output when the anchoring evidence is missing. It requires, in Haack's language, the intellectual virtues of genuine inquiry — honesty, thoroughness, independence — applied not to the generation of knowledge but to its evaluation.

The crossword puzzle is never finished. New clues arrive. Old entries are reconsidered. The grid evolves. This is the nature of genuine inquiry: not a project with a completion date but an ongoing process of checking, revising, extending, and sometimes starting over. AI accelerates the rate at which new entries are proposed. It does not — and cannot — accelerate the rate at which clues are checked. The checking is human work. It is the essential work. And it is the work that the aesthetics of the smooth, the pressure of productivity, and the seduction of coherent prose all conspire to make the evaluator skip.

Haack's crossword puzzle is, in the end, a picture of epistemic responsibility. The puzzle does not solve itself. The entries do not check themselves. The solver sits with the grid, pen in hand, and does the work. The work is not glamorous. It does not scale. It cannot be automated. It is the thing that makes the difference between a grid of justified beliefs and a grid of plausible fabrications. And in the age of AI, the difference has never mattered more.

Chapter 5: The Confabulation Problem

In the neurological ward, the patient is asked to raise his left arm. He cannot — the left side of his body is paralyzed following a stroke that damaged his right hemisphere. But when asked why he is not raising his arm, he does not say "I cannot." He says, with perfect confidence, "I don't want to." Or: "I already raised it a moment ago." Or: "My arm is tired from all the exercise I did this morning."

He is not lying. Lying requires awareness that the statement is false. The confabulating patient has no such awareness. His brain, confronted with a gap between what is expected and what is happening, fills the gap with a narrative that is internally coherent, contextually plausible, and delivered with complete conviction. The narrative is fabricated. The conviction is genuine. The patient believes what he is saying, because the fabrication satisfies the coherence requirements of his belief system — it fits the story he tells himself about who he is and what he can do — even though it fails the most elementary grounding requirement: his arm did not move.

The clinical literature on confabulation, developed over a century of neurological research from Korsakoff's original observations through the work of Morris Moscovitch and others, identifies a consistent set of features. Confabulated claims are internally coherent. They are contextually appropriate — they fit the conversational situation in which they are produced. They are delivered with the same fluency and confidence as accurate claims. And they are resistant to correction, because the patient does not experience them as errors. The patient experiences them as memories, as knowledge, as truths.

The structural parallel to AI-generated output is not approximate. It is precise.

A large language model, confronted with a prompt that requests information it does not possess, does not say "I don't know" — or when it does, it does so as a trained behavior, a guardrail imposed by reinforcement learning, not as a reflection of genuine epistemic humility. Left to its own devices, the model fills the gap with a narrative that is internally coherent, contextually plausible, and delivered with complete fluency. The narrative may be fabricated. The fluency is guaranteed. The model produces the fabrication and the accurate claim with identical confidence, identical linguistic quality, identical surface features.

This is why the term "hallucination," which has become the standard industry label for AI fabrication, is misleading. A hallucination is a perception without an object — a sensory experience that occurs in the absence of an external stimulus. The hallucinating person sees something that is not there. The metaphor implies that the AI is perceiving incorrectly, which implies that it is perceiving at all. It is not. The model has no perceptual apparatus. It has no sensory experience. It does not see, hear, or encounter the world. It processes statistical patterns and generates outputs consistent with those patterns.

Confabulation is the correct term, and the correction is not merely semantic. The difference between hallucination and confabulation is the difference between a perceptual failure and a narrative failure. The hallucinating person's perceptual system misfires. The confabulating person's narrative system fills a gap. The AI's generative system fills gaps continuously, because gap-filling is what next-token prediction does — it takes the existing context and extends it in the statistically most likely direction, regardless of whether that direction corresponds to reality.

Haack's foundherentist framework makes the epistemological danger of confabulation visible with a precision that neither the industry's reassurances nor the critics' alarms quite achieve. The danger is not that AI produces false claims. False claims are, in principle, detectable: they conflict with known truths, they produce logical contradictions, they fail empirical tests. The danger is that AI produces claims that satisfy every criterion of justified belief except one — and that the missing criterion is the one the human evaluator is least equipped to check.

The claims cohere. They fit the web of existing knowledge. They extend established arguments in plausible directions. They use appropriate vocabulary, cite relevant concepts, and arrive at conclusions that feel right. They satisfy the coherence dimension of Haack's framework comprehensively. What they do not satisfy — what they cannot satisfy, because the system has no mechanism for satisfying it — is the experiential anchoring dimension. The claims are not grounded in observation, experiment, or direct encounter with reality. They are grounded in the statistical patterns of a training corpus, which is a fundamentally different kind of grounding, and which may or may not correlate with experiential grounding depending on the domain, the quality of the training data, and the specific claim in question.

The epistemological consequence is that AI confabulation is harder to detect than AI error. An error — a claim that is straightforwardly false, that contradicts well-established facts, that produces an obvious logical inconsistency — is detectable by coherence-checking alone. It conflicts with other entries in the grid. The evaluator notices the conflict and rejects the claim.

A confabulation does not produce this conflict. It is designed — not intentionally, but architecturally — to avoid it. The model generates claims that cohere with the patterns in its training data, and those patterns are, for the most part, the patterns of genuine knowledge. The confabulated claim extends those patterns smoothly, in a direction that feels natural, that does not conflict with the evaluator's existing beliefs, that confirms rather than challenges the narrative the evaluator is already constructing. The confabulation passes the coherence test because the coherence test is the test the model was trained to pass.

Detection requires anchor-checking — the specific, effortful, unglamorous work of tracing the claim back to independent evidence. Does the citation exist? Does the source say what the model claims it says? Does the historical event actually have the date, the location, the participants that the model assigns to it? Does the scientific finding replicate? Does the philosophical concept have the meaning the model attributes to it?

This is the work that Segal described as the discipline of collaboration — the practice of treating AI output as a proposed crossword entry that must be checked against its clue. And his description of the moment when this discipline almost failed is, in Haack's terms, a precise account of how confabulation operates on the human evaluator.

The passage connecting Csikszentmihalyi to Deleuze was coherent. It was elegant. It extended two genuine bodies of thought in a direction that felt illuminating. Every internal feature of the passage signaled quality: the vocabulary was appropriate, the structure was logical, the conclusion was interesting. The evaluator — Segal, a careful and experienced reader — accepted it on first pass. Only on rereading, when something nagged, when the coherence felt too smooth, did he check the anchor. The anchor — Deleuze's actual philosophical position — did not match. The entry was wrong. The grid's coherence had concealed the grounding failure.

The nagging is significant. It was not a logical detection. It was not the result of a systematic verification procedure. It was a feeling — an epistemic intuition, developed through years of reading and thinking, that something was off. Not every evaluator would have felt it. Not every evaluator who felt it would have acted on it. The confabulation was good enough to pass most coherence tests and subtle enough to evade most grounding checks. It was caught by a specific combination of domain knowledge and intellectual temperament that cannot be assumed to be universally present.

This is the scalability problem. Confabulation detection does not scale. Each confabulated claim must be checked individually, against domain-specific evidence, by an evaluator with sufficient knowledge to recognize the discrepancy between what the model claims and what the evidence supports. The model produces claims at the speed of computation. The evaluator checks claims at the speed of human cognition, which is orders of magnitude slower and constrained by the evaluator's domain knowledge, available time, and tolerance for the tedious work of verification.

The asymmetry is structural and, in the current technological landscape, irreducible. The model generates coherent confabulations faster than any human can check them. The gap between generation speed and verification speed means that, in practice, most AI-generated claims go unverified. The evaluator checks some claims — the ones that seem most important, the ones that trigger epistemic intuitions, the ones that can be checked quickly — and lets the rest through. The ones that get through may be true. They may be confabulated. Without checking, the evaluator cannot tell, because the surface features of truth and confabulation are identical.

Haack's three dimensions of evidential quality provide a framework for prioritizing which claims to check, even when checking all of them is impossible. Claims where the supportiveness dimension is weak — where the evaluator has no independent evidence bearing on the claim — should be treated with particular suspicion, because they cannot be verified even in principle without additional research. Claims where the independent security dimension is questionable — where the claim falls in a domain where the model's training data is known to be sparse, contradictory, or unreliable — should be flagged for verification. Claims where the comprehensiveness dimension is compromised — where the model's output addresses only part of the question, ignoring relevant evidence or alternative perspectives — should be supplemented with independent analysis.

These are practical heuristics derived from philosophical principles. They do not eliminate the confabulation problem. Nothing eliminates the confabulation problem short of the model developing a genuine relationship to truth, which it does not currently possess and may never possess. But they reduce the evaluator's vulnerability by directing attention to the places where confabulation is most likely and most consequential.

The clinical parallel illuminates a further danger. In neurological confabulation, the patient's confidence in the fabricated claim is itself a barrier to correction. The patient does not experience uncertainty. The patient does not hedge or qualify. When challenged, the patient often doubles down, producing additional confabulated detail to support the original confabulation — a cascade of fabrication that makes the narrative more elaborate without making it more true.

AI systems exhibit an analogous behavior. When a user questions a model's claim, the model may generate additional supporting detail — citations, explanations, elaborations — that are themselves confabulated. The model does not have the capacity for genuine self-correction, because genuine self-correction requires knowing what is true and comparing one's claims against that knowledge. The model does not know what is true. It knows what is statistically likely. And when the statistically likely response to a challenge is to provide supporting evidence, the model provides supporting evidence, regardless of whether that evidence exists.

The cascade is particularly dangerous because it mimics the behavior of a knowledgeable person responding to legitimate scrutiny. When a genuine expert is challenged on a claim, the expert provides supporting evidence — additional data, references, reasoning — that strengthens the claim. The behavior is identical from the outside: challenge produces elaboration. The difference is that the expert's elaboration is grounded in knowledge and the model's elaboration is grounded in pattern. The evaluator who accepts the elaboration as evidence of genuine knowledge is making a category error that the surface behavior conceals.

Haack's framework places the burden of detection squarely on the human evaluator. The machine confabulates. The human checks. The machine confabulates confidently. The human must be proportionally skeptical. The machine confabulates fluently. The human must distrust fluency as an indicator of truth. These are demanding requirements. They run against the grain of human cognitive tendencies — the fluency heuristic, the tendency to defer to confident sources, the cognitive fatigue that accumulates over hours of evaluation. They are also, in Haack's terms, the requirements of genuine inquiry, requirements that have always been demanding, that have always run against cognitive tendencies, and that have never been more important than in an era when the most fluent, most confident, most coherent voice in the room may have no relationship to truth at all.

The confabulation problem is not a bug that will be fixed in the next model update. It is an architectural feature of systems that generate output through statistical pattern completion rather than through evidential reasoning. Future models will confabulate less frequently, perhaps, as training techniques improve and grounding mechanisms become more sophisticated. But the structural gap between coherence and grounding — between the model's ability to produce claims that fit together and its inability to verify that those claims correspond to reality — is not a gap that engineering alone can close. It is an epistemological gap, and it requires an epistemological response.

That response begins with the recognition that confabulation is the default, not the exception. The evaluator who assumes AI output is accurate unless proven otherwise has the burden of proof backwards. The foundherentist evaluator assumes that AI output is ungrounded unless independently verified — that the entry fits the crossings but has not been checked against the clue — and proceeds accordingly. The discipline is not suspicion for its own sake. It is the epistemic practice that the nature of the system requires, applied with the rigor that the stakes demand.

The stakes are not hypothetical. Every undetected confabulation that enters the epistemic commons — every fabricated citation in a legal brief, every invented statistic in a policy document, every false historical claim in an educational text — degrades the shared informational environment on which democratic deliberation, scientific progress, and cultural coherence depend. The degradation is cumulative. Each undetected confabulation makes the next one harder to detect, because the confabulated claim becomes part of the web against which subsequent claims are checked. The crossword grid fills with unchecked entries, and each unchecked entry weakens the grid's capacity to catch errors in subsequent entries.

The confabulation problem is, in the end, a problem about the relationship between speed and care. The model generates at the speed of computation. Genuine inquiry proceeds at the speed of verification. The gap between these speeds is the space in which confabulation thrives. Closing the gap requires not faster verification but a cultural commitment to the principle that unverified claims are unjustified claims, regardless of how coherent, how fluent, or how confident they appear.

Chapter 6: The Grounding Problem

In 2020, a team at Facebook AI Research published a paper demonstrating that their language model could generate plausible-sounding scientific abstracts that fooled human evaluators approximately thirty percent of the time. The abstracts were coherent, used appropriate terminology, cited real journal names, and arrived at conclusions that sounded reasonable. They were also entirely fabricated — not wrong in the way a flawed study is wrong, but fictional in the way a novel is fictional. The experiments described had never been conducted. The data did not exist. The conclusions followed from nothing.

The AI industry's response to this category of failure has been, characteristically, engineering. The most prominent engineering response is retrieval-augmented generation, or RAG — a technique that connects the language model's generative process to a database of verified documents. Instead of generating entirely from its trained patterns, the model first retrieves relevant documents from a curated corpus, then generates its response in the context of those documents. The retrieved documents serve as a kind of evidential tether, anchoring the model's output to sources that have been independently verified.

The improvement is real and measurable. RAG systems produce fewer fabricated citations, fewer invented statistics, fewer fictional claims. Studies comparing RAG-augmented models to unaugmented models show significant reductions in confabulation rates across multiple domains. The engineering is sound. The improvement is welcome. And the epistemological analysis of what RAG does and does not accomplish reveals a residual danger that the engineering metrics do not capture.

Haack's foundherentist framework provides the diagnostic instrument. RAG is a foundationalist intervention applied to a coherentist system. It adds experiential anchors — verified documents serving as proxies for direct observation — to a web of statistically generated claims. The intervention addresses the grounding problem at the points where retrieval occurs. At those points, the model's output is tethered to a source that has been independently checked. The claims that derive from retrieved documents are, to the extent that the documents are reliable, grounded.

The problem is that not all of the model's output derives from retrieved documents. The retrieval is selective. The model retrieves a subset of relevant documents — constrained by the retrieval algorithm, the size of the database, the specificity of the query — and generates its response in the context of those documents. But the generation extends beyond the retrieved content. The model fills gaps between the retrieved documents with its own pattern-based completions. It draws inferences, makes connections, extends arguments — all using the same coherence-driven generative process that operates in the absence of retrieval.

The result is output that is grounded at some points and ungrounded at others, with no visible boundary between the two. The retrieved content and the generated content are woven together seamlessly — "seamlessly" in precisely the sense that Haack's framework identifies as dangerous, because the seam is where the evaluator could distinguish grounded from ungrounded, and the absence of a visible seam makes this distinction impossible without independent verification.

Consider a concrete case. A lawyer asks a RAG-augmented legal AI to analyze a contract dispute. The system retrieves relevant case law — three decisions from the relevant jurisdiction, each accurately cited and correctly summarized. So far, the grounding is solid. The retrieved cases are real, the citations are accurate, the summaries are faithful. But the analysis extends beyond the retrieved cases. The model draws an inference about how the retrieved cases apply to the specific facts of the dispute. It identifies a principle that connects the three cases. It predicts how a court would likely rule. These extensions are generated, not retrieved. They are coherent — they follow logically from the retrieved material and they are consistent with the patterns of legal reasoning in the training data. But they are not grounded in any source that the retrieval process consulted.

The lawyer reading this analysis encounters a document that is grounded at three points and ungrounded at seven. The three grounded points — the case citations and summaries — create a veneer of reliability that extends, in the reader's perception, to the seven ungrounded points. The analysis feels trustworthy because parts of it are trustworthy. The grounded claims contaminate the reader's evaluation of the ungrounded claims, not through any logical mechanism but through the psychological mechanism of anchoring: the presence of verified information in the neighborhood of unverified information raises the perceived credibility of the whole.

Haack's framework identifies this as a specific epistemic failure: a violation of the comprehensiveness dimension. The evaluator has evidence bearing on some of the claims (the retrieved documents) but not on others (the generated inferences). The evaluator who treats the entire analysis as equally well-grounded has failed to track which elements are supported by independent evidence and which are not. The comprehensiveness requirement demands that the evaluator know, for each claim, what evidence supports it and how strong that evidence is. RAG systems make this tracking harder, not easier, because they blur the boundary between grounded and ungrounded content.

The situation is, in Haack's terms, epistemologically worse than pure confabulation. A purely generative model — one without retrieval augmentation — produces output that the careful evaluator knows is entirely ungrounded. The evaluator's epistemic posture is appropriately skeptical across the board. Every claim requires independent verification. The baseline assumption is that nothing has been checked.

A RAG-augmented model produces output that the evaluator knows is partially grounded. This partial grounding creates a new epistemic challenge: the evaluator must determine which parts are grounded and which are not, and must maintain differential levels of skepticism across different parts of the same output. This is cognitively demanding in a way that uniform skepticism is not. The human mind is not naturally calibrated for variable-confidence evaluation of integrated text. When a document reads as a unified whole — when the grounded and ungrounded elements are woven together in a single coherent narrative — the natural tendency is to evaluate the whole at a single confidence level, and the presence of grounded elements biases that level upward.

This is the false security problem. Partial grounding creates a false sense of epistemic security that may be more dangerous than no grounding at all, because it disarms precisely the skepticism that the ungrounded elements require. The evaluator who knows the output is entirely ungrounded keeps a guard up. The evaluator who knows some of the output is grounded lowers the guard — and the confabulated elements enter through the opening.

The false security problem has a structural parallel in the history of epistemology that Haack would recognize. The logical positivists attempted to ground all of knowledge in a foundation of observational protocol sentences. The foundation was genuine — observational sentences do describe real observations. But the inferential superstructure built upon that foundation extended far beyond what the foundation could support. The positivists' error was not in valuing the foundation. It was in assuming that the security of the foundation transferred automatically to the superstructure. The same error is being committed, at industrial scale, by users who assume that the security of RAG-retrieved documents transfers automatically to the model's generated extensions of those documents.

Haack's second dimension of evidential quality — independent security — provides a more granular diagnostic. The retrieved documents have high independent security: they exist, they are verifiable, their provenance can be traced. The model's generated extensions have low independent security: they are products of statistical pattern completion, their provenance is opaque, and their relationship to reality is mediated by the same confabulation-prone architecture that operates in the absence of retrieval.

The evaluator applying Haack's framework would treat the retrieved content and the generated content differently — assigning high confidence to the former and low confidence to the latter, even when they appear as parts of a seamless whole. This differential evaluation is the core epistemic practice that RAG demands and that RAG's seamless design actively discourages.

There is a further complication. The quality of the grounding depends on the quality of the retrieved documents, which depends on the quality of the retrieval database, which depends on the curation decisions of the people who built it. A RAG system grounded in a well-curated database of peer-reviewed scientific literature produces output with relatively high anchoring quality. A RAG system grounded in a database that includes unreliable sources — preprints, blog posts, opinion pieces, outdated documents — produces output with anchoring that is itself epistemically weak. The grounding is only as good as the ground.

Haack would note that this is a straightforward application of the independent security dimension: the evidential quality of the grounding material must itself be evaluated before it can serve as a basis for evaluating the model's output. The evaluator who trusts the RAG system's retrieval without examining the retrieved documents is delegating an epistemic judgment to a retrieval algorithm — an algorithm that selects documents based on relevance scores, not on epistemic reliability. Relevance and reliability are different properties. A document can be highly relevant to a query and highly unreliable — a well-targeted propaganda piece, for instance, or an outdated scientific paper whose conclusions have been superseded.

The engineering response to the grounding problem is to improve retrieval quality, expand database coverage, and develop techniques for distinguishing generated content from retrieved content in the model's output. These are worthy engineering objectives. They address the problem at the technical level. They do not address the problem at the epistemic level, because the epistemic problem is not that the technology is imperfect. The epistemic problem is that the relationship between grounding and generation in AI output is structurally different from the relationship between evidence and inference in human reasoning, and that evaluating one as though it were the other produces systematic epistemic error.

Human reasoning, at its best, maintains an explicit relationship between claims and evidence. The scientist cites data. The historian cites sources. The lawyer cites cases. The citations serve not just as support but as transparency mechanisms — they allow the reader to trace the claim back to its ground, to evaluate the ground independently, and to assess whether the claim is supported by the ground it purports to rest on.

AI output erases this relationship. The generated text does not distinguish between claims derived from retrieval and claims derived from pattern completion. The "citations" in AI output may be accurate (derived from retrieved documents) or fabricated (generated by the model's coherence-driven pattern completion). The evaluator cannot tell which is which without independent verification — without checking the clue — and the seamless integration of retrieved and generated content makes the need for this checking less apparent, not more.

The prescription that follows from Haack's framework is demanding but precise. RAG-augmented AI output should be treated as a partially filled crossword grid in which some entries have been checked against their clues and others have not. The evaluator's task is to identify which entries are checked (corresponding to claims derived from retrieved documents) and which are unchecked (corresponding to claims generated by the model), and to verify the unchecked entries independently. The presence of checked entries strengthens the grid — it provides reliable crossings for evaluating adjacent entries — but it does not justify accepting unchecked entries on the basis of their proximity to checked ones.

This practice is foundherentist to its core: it requires both anchoring (checking unchecked claims against independent evidence) and coherence (verifying that new entries are consistent with checked entries). It requires both, simultaneously, maintained through the ongoing discipline of genuine inquiry. And it requires the evaluator to resist the most seductive feature of RAG-augmented output: the appearance of comprehensive grounding where only partial grounding exists.

Chapter 7: The Confidence Problem

In 1974, psychologists Amos Tversky and Daniel Kahneman published a paper that changed the study of human judgment. They demonstrated, through a series of elegantly designed experiments, that human beings are systematically miscalibrated in their confidence. People who report being ninety percent confident in their answers are correct approximately seventy-five percent of the time. People who report certainty — one hundred percent confidence — are wrong between fifteen and twenty percent of the time. The confidence and the accuracy are correlated but not aligned. The gap between them is not random. It is systematic, predictable, and remarkably resistant to correction.

This finding, replicated hundreds of times across dozens of domains and populations, established overconfidence as one of the most robust phenomena in the psychology of judgment. Human beings are, as a species, more confident than they are correct. The miscalibration is not a personal failing. It is a feature of human cognition — a consequence of the heuristics that make rapid judgment possible at the cost of systematic bias.

The confidence problem in AI is structurally related but operationally different. The language model does not experience confidence. It does not have a subjective sense of certainty or uncertainty. It has output probabilities — numerical values assigned to each possible next token, reflecting the model's statistical estimate of how likely that token is to appear in the sequence. High probability tokens are selected more often. Low probability tokens are selected less often. The temperature parameter controls how strictly the model adheres to high-probability selections.

But the output that the user sees is not probabilities. It is text. And the text arrives uniformly formatted, uniformly fluent, uniformly confident in tone. The model does not stammer when it is uncertain. It does not hedge when its probability distribution is flat. It does not signal to the user that the claim it is making rests on a narrow statistical base rather than a broad one. The confidence is in the presentation, not in the model, and the presentation is optimized for the same coherence and fluency that characterize all of the model's output.

The epistemological consequence is that the user has no reliable signal for distinguishing high-confidence claims from low-confidence claims. In human communication, confidence is a cue. A speaker who makes a claim with evident conviction is, in most contexts, signaling that they believe the claim to be well-supported. A speaker who hedges — "I think," "I'm not sure, but," "this might be wrong" — is signaling uncertainty. The cues are imperfect, as the overconfidence literature demonstrates, but they are informative. They carry information about the speaker's internal assessment of the claim's reliability.

AI output lacks these cues. Every claim is presented with the same surface confidence. The fabricated citation and the accurate citation are formatted identically. The well-supported claim and the confabulated claim are delivered in the same tone. The user's evolved capacity for detecting uncertainty through social cues — through hesitation, hedging, vocal inflection, facial expression — is entirely disarmed. The only cue that remains is the content of the claim itself, which must be evaluated on its own merits, without the social-epistemic scaffolding that human communication normally provides.

Haack's framework locates the confidence problem precisely. The problem is not that the model is overconfident. The model is not confident at all, in any epistemologically meaningful sense. The problem is that the user treats the model's output as though it came from a confident speaker — that is, as though the uniform fluency of the output were evidence of the speaker's assessment of the claim's reliability. This is a category error: the user is applying the epistemic norms of human communication to a system that is not communicating in the human sense but generating statistically likely text.

The category error is not the user's fault. It is an inevitable consequence of the interface design. When a system produces output in natural language, human beings process that output using the cognitive machinery evolved for processing natural language produced by other human beings. That machinery includes the confidence heuristic — the tendency to infer reliability from fluency and conviction. The heuristic fires automatically, below the threshold of conscious deliberation. Overriding it requires explicit, effortful cognitive intervention — the kind of intervention that Haack's framework demands but that human cognitive architecture resists.

The confidence problem is, in this sense, deeper than the grounding problem. The grounding problem is about the relationship between the model's output and reality. The confidence problem is about the relationship between the user and the model's output. A perfectly grounded model would still produce a confidence problem if the user had no way to assess the grounding independently. A perfectly ungrounded model would produce no confidence problem if the user treated every claim with appropriate skepticism. The variable is not the model's reliability but the user's epistemic posture.

Haack would recognize this as a familiar epistemological structure. The history of epistemology is, in significant part, a history of diagnosing the gap between the apparent reliability of belief-forming processes and their actual reliability. Descartes' demon, the brain in a vat, Gettier's counterexamples to the justified-true-belief theory of knowledge — all of these are thought experiments designed to expose situations in which beliefs appear well-justified but are not. The AI confidence problem is a real-world instantiation of this philosophical concern, operating not in the philosopher's study but in the lawyer's office, the doctor's clinic, and the student's dorm room.

The practical consequences are immediate and measurable. A 2024 study by researchers at Stanford found that when law students were given AI-generated legal analyses to evaluate, their ability to detect errors was significantly impaired when the analyses were fluent and well-structured compared to when the same errors were embedded in poorly written text. The errors were identical. The wrapping was different. The fluent wrapping reduced error detection by a substantial margin. The students were not careless. They were applying the fluency heuristic — the evolved cognitive shortcut that equates smooth processing with truth — and the heuristic was leading them astray.

The study is a controlled demonstration of what Haack's framework predicts: that the surface features of justified belief — logical structure, appropriate vocabulary, confident presentation — can be present in the absence of justification, and that human evaluators are systematically vulnerable to this dissociation because their cognitive machinery was calibrated for a world in which the surface features and the underlying justification were reliably correlated. AI breaks the correlation. The machinery misfires. The evaluator accepts claims that a more skeptical posture would have flagged.

The institutional dimension of the confidence problem multiplies its consequences. Individual overconfidence is bounded by the individual's sphere of influence. Institutional overconfidence — the systematic acceptance of AI output by organizations that incorporate it into their decision-making processes — operates at scale. A law firm that integrates AI-generated research into its workflow without adequate verification procedures is institutionalizing the confidence problem. A hospital that relies on AI-generated diagnostic suggestions without independent clinical confirmation is institutionalizing it. A university that accepts AI-generated student work without epistemic scrutiny is institutionalizing it.

In each case, the institution is treating the model's output as though it were the output of a knowledgeable, reliable human professional — applying the epistemic norms appropriate to human expertise to a system that produces expertise-shaped output without possessing expertise. The institution is making the category error at organizational scale, and the consequences propagate through every downstream decision that relies on the unverified output.

Haack's characterological analysis adds a dimension that institutional analysis typically misses. The confidence problem is not just a matter of procedures and verification protocols. It is a matter of intellectual character — of the disposition to question, to doubt, to resist the seduction of fluent certainty. Institutions can mandate verification procedures. They cannot mandate the intellectual virtues that make verification meaningful.

A verification procedure that is followed mechanically — checking the first citation, confirming it is real, and then assuming the rest are equally reliable — satisfies the procedural requirement without satisfying the epistemic one. The procedure checks a box. The intellectual virtue checks the claim. The difference is the difference between compliance and inquiry, between going through the motions and actually caring whether the output is true.

This is the characterological requirement that Haack identified as central to genuine inquiry. The genuine inquirer does not merely follow evidential procedures. The genuine inquirer cares about truth — is motivated by the desire to get things right, is willing to follow evidence into uncomfortable conclusions, is honest about the limits of their knowledge and the unreliability of their cognitive machinery. The AI confidence problem is a test of this character, applied at scale, under conditions designed by the tool's very architecture to make the test harder.

The prescriptions that follow from Haack's framework are not procedural. They are dispositional. The evaluator must cultivate a specific epistemic posture: one that treats fluency as noise rather than signal, that treats confidence as a feature of the presentation rather than evidence of reliability, that maintains the asymmetry between coherence-checking and anchor-checking that the foundherentist framework requires. The posture is demanding. It runs against evolved cognitive tendencies. It is fatiguing to maintain. And it is the only posture that is epistemically adequate to the conditions that AI creates.

The alternative is not catastrophe in every instance. Most AI output, most of the time, is approximately correct. The confidence problem produces harm not through universal failure but through selective, invisible, undetectable failure — the occasional confabulation embedded in a stream of accurate claims, the single wrong citation in a brief full of right ones, the fabricated statistic in a report whose other statistics are verified. The harm is cumulative and structural. Each undetected error enters the epistemic commons and becomes part of the web against which subsequent claims are evaluated. The grid fills with unchecked entries. The crossings that were supposed to catch future errors are themselves unreliable. The system degrades from the inside.

Haack's framework does not promise that this degradation can be prevented. It promises that the degradation can be resisted — by individuals who maintain the intellectual virtues of genuine inquiry, by institutions that build verification into their workflows not as compliance but as commitment, and by a culture that recognizes that the confidence problem is not a bug to be fixed but a condition to be managed, continuously, with the care and attention that the stakes require. The model generates. The human evaluates. The evaluation is the epistemic contribution. And the quality of the evaluation depends, as it always has, not on the procedures the evaluator follows but on the character the evaluator brings.

Chapter 8: Evidence and the Artificial Inquirer

Charles Sanders Peirce, the American pragmatist whose work Haack has described as the deepest influence on her own philosophical development, drew a distinction in 1877 that has lost none of its force. In "The Fixation of Belief," Peirce identified four methods by which human beings settle their beliefs: tenacity (clinging to what one already believes), authority (accepting what a trusted source declares), the a priori method (believing what seems agreeable to reason), and the method of science (submitting beliefs to the test of experience and revising them in light of the results).

Only the last method, Peirce argued, has the capacity for self-correction. The others fix belief effectively — people do cling to prior convictions, do defer to authority, do believe what seems reasonable — but they cannot correct their own errors, because they lack a mechanism for recognizing error as error. The method of science works not because scientists are more intelligent or more virtuous than other people, but because the method itself includes a built-in error-correction mechanism: the systematic confrontation of belief with experience, repeated over time, with the results subjected to public scrutiny.

Haack built her epistemology on this Peircean foundation. Genuine inquiry, in her framework, is not defined by its subject matter, its institutional location, or its professional credentials. It is defined by its commitment to the self-correcting process that Peirce identified: the willingness to submit beliefs to the test of evidence, to revise them when the evidence demands revision, and to accept the discomfort of uncertainty rather than the false comfort of premature conviction.

This commitment is, Haack has argued repeatedly, more than methodological. It is characterological. The genuine inquirer does not merely follow evidential procedures the way a bureaucrat follows regulations — mechanically, without investment, checking boxes. The genuine inquirer cares about getting things right. The caring is not sentimental. It is operational. It shapes how the inquirer interacts with evidence: with attention, with honesty, with the willingness to be surprised.

The distinction between procedural and characterological commitment is crucial to understanding what AI can and cannot contribute to inquiry. A language model can follow evidential procedures in a limited sense. It can be prompted to cite sources. It can be configured to retrieve documents before generating claims. It can be instructed to qualify uncertain statements, to flag potential errors, to present multiple perspectives on contested questions. These are procedural interventions, and they improve the quality of the model's output in measurable ways.

What the model cannot do is care. Not in the sentimental sense — the model's lack of feelings is not the point. In the operational sense that Haack identifies as the core of genuine inquiry: the commitment to truth that shapes how evidence is gathered, evaluated, and weighed. The model has no commitment to truth. It has an optimization target — next-token prediction — that produces output resembling the products of truth-committed inquiry. The resemblance is remarkable. The commitment is absent.

This absence has specific, identifiable epistemic consequences that Haack's framework makes visible.

The first consequence is that the model cannot recognize when its own output is wrong. Genuine self-correction requires a standard against which the output can be measured — a commitment to getting things right that allows the inquirer to recognize the gap between what they have produced and what the evidence supports. The model has no such standard. It has a statistical distribution and a generation mechanism. When the mechanism produces a claim that is false, the model does not experience the falseness as a problem to be corrected. It does not experience anything. The claim is simply the output that the mechanism produced, and the mechanism has no internal representation of the difference between outputs that correspond to reality and outputs that do not.

The engineering response — reinforcement learning from human feedback, constitutional AI, various alignment techniques — addresses this problem by training the model to produce outputs that human evaluators rate as helpful, harmless, and honest. The training is effective at the behavioral level: the model learns to avoid certain categories of error, to hedge in certain situations, to acknowledge uncertainty when prompted to do so. But the behavioral improvement does not constitute genuine self-correction, because the model's "corrections" are responses to training signals, not responses to evidence. The model does not revise its beliefs in light of new evidence, because it does not have beliefs. It adjusts its output distributions in response to reward signals, which is a fundamentally different process — one that can approximate the surface behavior of self-correction without embodying its epistemic substance.

The second consequence is that the model cannot distinguish between genuine inquiry and sham inquiry — a distinction that Haack has identified as the most important epistemic distinction of our time. A user who prompts the model to analyze a question honestly will receive one kind of output. A user who prompts the model to defend a predetermined conclusion will receive a different kind — equally coherent, equally fluent, equally well-structured, but oriented toward confirmation rather than truth. The model does not know the difference. It cannot know the difference, because the difference resides in the intention of the inquirer, and intention is the one thing the model does not evaluate.

Peirce's taxonomy of belief-fixation methods illuminates this problem with uncomfortable clarity. The model is, depending on how it is used, an instrument for any of Peirce's four methods. Used by a genuine inquirer — one who checks the output against evidence, revises in light of new information, and maintains the commitment to truth that genuine inquiry requires — the model serves the method of science. It generates hypotheses, surfaces connections, and expands the space of possibilities for the inquirer to evaluate.

Used by someone engaged in sham inquiry — someone who has already fixed their belief and is using the model to generate supporting arguments — the model serves the method of tenacity or authority, dressed in the language of science. The sham inquirer prompts: "Argue that X is true." The model obliges, producing an argument that exhibits every formal feature of genuine analysis — logical structure, citation of evidence, consideration of counterarguments — while systematically excluding any evidence that would undermine the predetermined conclusion. The output is sham reasoning, produced at the speed of computation, indistinguishable on the surface from the products of genuine inquiry.

Haack's taxonomy of inquiry — genuine, sham, and fake — provides the vocabulary for diagnosing this problem. Genuine inquiry follows evidence wherever it leads. Sham inquiry adopts the procedures of genuine inquiry while serving a predetermined conclusion. Fake inquiry does not even adopt the procedures — it produces conclusions without evidential basis. The model can serve all three, and does serve all three, depending entirely on the user's epistemic posture.

The third consequence is that the model cannot exercise the judgment that genuine inquiry requires at the points where evidence is ambiguous, incomplete, or conflicting. Genuine inquiry does not proceed smoothly from evidence to conclusion. It proceeds through a series of judgment calls — moments where the available evidence underdetermines the conclusion, where multiple interpretations are consistent with the data, where the inquirer must weigh competing considerations and make a decision about which interpretation is best supported.

These judgment calls are, in Haack's framework, the moments where the characterological dimension of inquiry is most consequential. The inquirer who is genuinely committed to truth will weigh the competing interpretations honestly, giving each its due, resisting the temptation to favor the interpretation that is most convenient or most consistent with prior commitments. The sham inquirer will weigh them selectively, systematically favoring evidence that supports the desired conclusion. The model will weigh them statistically, favoring the interpretation that is most consistent with the patterns in its training data — which may or may not correspond to the interpretation that the evidence best supports.

Statistical weighting and evidential weighting are not the same thing. The most statistically likely interpretation of ambiguous evidence is the interpretation that is most common in the training data, which is the interpretation that the most people have expressed, which is not necessarily the interpretation that the evidence best supports. Popular opinion and evidential support are correlated, but the correlation is imperfect, and the imperfection is precisely the gap in which the model's outputs diverge from the products of genuine inquiry.

What, then, is the model good for? The answer, from within Haack's framework, is substantial — but bounded. The model is good for generating possibilities. For surfacing connections across bodies of knowledge too vast for any individual human to traverse. For producing draft analyses that a human inquirer can then evaluate, revise, and ground. For the coherence dimension of the foundherentist framework — the part of justification that involves checking a proposed belief against the web of other beliefs for consistency and mutual support.

These are genuine contributions. They are not trivial. The model can propose crossword entries faster and in greater variety than any human solver. It can identify intersections — connections between concepts, disciplines, and data sets — that the human solver might never notice. It can fill the grid with proposed entries that, checked against their clues, turn out to be correct a surprising proportion of the time.

But the model cannot check the clues. It cannot anchor its proposals in experiential evidence. It cannot exercise the judgment that genuine inquiry requires at the moments of ambiguity where evidence underdetermines conclusion. It cannot care about truth. These are the contributions that only the human inquirer can make, and they are the contributions on which the epistemic quality of the collaboration depends.

Haack's pragmatist inheritance from Peirce converges here with a practical conclusion that the builders of AI systems are arriving at independently: the value of human-AI collaboration depends on the human maintaining exactly the epistemic posture that the AI cannot supply. The model provides coherence. The human provides grounding. The model generates possibilities. The human evaluates them. The model proposes entries. The human checks the clues.

The division of labor is clear. The temptation to violate it is constant. The model's output is so fluent, so coherent, so confident that the evaluator is perpetually tempted to skip the clue-checking — to accept the entry because it fits the grid, because it sounds right, because checking takes time and effort and the output is already good enough.

"Good enough" is the enemy of genuine inquiry. Peirce knew this. Haack knows this. The pragmatist tradition, despite its reputation for practical-mindedness, has always insisted on the distinction between what works and what is true — between the belief that survives the test of experience and the belief that merely avoids the test. The model's output may be good enough for many practical purposes. It is not good enough for genuine inquiry, unless it is subjected to the evidential scrutiny that genuine inquiry demands.

The artificial inquirer is not an inquirer at all. It is an instrument in the hands of an inquirer, and the quality of the inquiry depends entirely on the hands that hold it. The instrument is powerful. The hands must be worthy of the instrument. That worthiness is not technical. It is not procedural. It is characterological — a matter of the intellectual virtues that Haack has spent her career articulating and defending, virtues that were always demanding, that were always countercultural in a world that rewards speed and confidence over care and honesty, and that have never been more necessary than in an era when the most fluent voice in any room may be the one with the least to say about truth.

Chapter 9: Genuine Inquiry and Its Counterfeits

The phrase "do your own research" underwent a strange transformation in the first quarter of the twenty-first century. It migrated from the laboratory and the library — contexts where it meant something precise and demanding — to the internet comment section, where it came to mean something closer to its opposite: find the source that confirms what you already believe, and call the finding research.

Haack diagnosed this corruption of inquiry long before social media accelerated it. Her taxonomy, developed across multiple works from Manifesto of a Passionate Moderate through Putting Philosophy to Work, identifies three categories of intellectual activity that are routinely conflated but epistemologically distinct.

Genuine inquiry is the honest pursuit of truth. The genuine inquirer formulates a question, gathers evidence bearing on the question, weighs the evidence according to its quality and relevance, and arrives at a conclusion that the evidence supports — even when that conclusion is uncomfortable, inconvenient, or contrary to the inquirer's prior commitments. The defining feature of genuine inquiry is not its method, though method matters, but its orientation: toward truth, wherever truth leads.

Sham inquiry adopts the procedures of genuine inquiry while serving a predetermined conclusion. The sham inquirer formulates a question — but the answer is already decided. Evidence is gathered — but selectively, with disconfirming evidence systematically excluded. Arguments are constructed — but the construction is advocacy dressed as analysis. The sham inquirer goes through the motions of inquiry with the precision of a professional actor: every gesture correct, every line delivered convincingly, the performance indistinguishable from genuine inquiry to anyone who is not watching carefully enough to notice that the conclusion was decided before the evidence was consulted.

Fake inquiry does not even adopt the procedures. The fake inquirer produces conclusions without evidential basis — assertions disguised as findings, opinions dressed as analysis, prejudice costumed as research. Fake inquiry is the easiest to detect, because it does not bother to simulate the formal features of genuine inquiry. It simply announces and moves on.

The taxonomy matters because each category has a different relationship to AI, and the differences determine whether AI-assisted intellectual work produces knowledge or its counterfeits.

When genuine inquiry meets AI, the result can be genuinely valuable. The model generates possibilities — hypotheses, connections, alternative framings — that the inquirer then evaluates against evidence. The model's coherence contributes to the coherence dimension of the foundherentist framework. The inquirer's commitment to truth contributes the anchoring dimension. The collaboration works because both dimensions are present: the model's contribution is checked against the inquirer's evidential standards, and the inquirer's limitations are supplemented by the model's range.

This is the mode of collaboration that Segal described in The Orange Pill — the practice of treating AI output as a proposed entry in the crossword grid, checking it against the clue, and rejecting it when the clue does not match. The practice is demanding. It requires the inquirer to possess independent knowledge of the domain, or the willingness to acquire it, and the discipline to maintain epistemic standards when the model's output is fluent enough to make those standards feel unnecessary.

When sham inquiry meets AI, the result is sham reasoning at unprecedented scale. The sham inquirer already knows what conclusion they want. They prompt the model: "Make the case that X is true." Or more subtly: "Analyze this question," having framed the question in a way that predisposes the analysis toward the desired conclusion. The model obliges. It produces an analysis that exhibits every formal feature of genuine inquiry — logical structure, citation of evidence, consideration of counterarguments, qualified conclusions. The output is indistinguishable, on the surface, from the product of genuine analysis.

The counterarguments the model considers are the weakest available counterarguments, because the prompt's framing has narrowed the space. The evidence the model cites is the evidence consistent with the predetermined conclusion, because the model generates output consistent with the context it is given, and the context has been shaped to favor one side. The qualifications are cosmetic — they create the appearance of intellectual honesty without the substance, because the fundamental orientation of the inquiry was toward confirmation, not truth.

The danger is not that sham reasoning is new. It is as old as rhetoric. The danger is that AI makes sham reasoning cheap, fast, and indistinguishable from genuine analysis. Before AI, producing a persuasive sham analysis of a complex question required considerable skill and effort — enough skill and effort that the sham reasoner had to invest significant resources, which placed a practical limit on the volume of sham reasoning any individual could produce. AI eliminates this practical limit. A single person, prompted with a predetermined conclusion and a domain, can generate dozens of sophisticated-seeming sham analyses in an afternoon. Each analysis will be internally coherent. Each will cite relevant-seeming evidence. Each will present itself with the confidence and fluency of genuine inquiry.

The epistemological commons is flooded, and the flood is not random noise. It is structured noise — carefully shaped to resemble signal, formatted to pass the superficial tests that most readers apply, and generated at a volume that makes careful evaluation impossible for any individual recipient.

Haack's diagnosis of this problem is not that the tools are misused, though they are. It is that the epistemological norms required to distinguish genuine from sham inquiry are precisely the norms that the culture has been neglecting, and that AI's amplification of sham reasoning makes this neglect catastrophic. The norms are simple to state: Check the evidence independently. Ask whether disconfirming evidence has been considered. Evaluate whether the conclusion was predetermined. Assess whether the inquiry followed the evidence or the evidence was selected to follow the inquiry.

These norms are simple to state and difficult to practice, because practicing them requires time, effort, domain knowledge, and the specific intellectual virtue that Haack calls the commitment to genuine inquiry — the caring about truth that is characterological rather than procedural. No verification checklist can substitute for this commitment, because the sham inquirer can follow a verification checklist as easily as the genuine inquirer. The checklist does not distinguish between the person who checks a citation because they want to know if it is accurate and the person who checks it because the procedure requires them to and who will accept whatever the check reveals without further scrutiny.

The distinction between genuine and sham is, at its foundation, a distinction of motivation — of what the inquirer is trying to do. This is epistemically irreducible. No technology can detect the difference from the outside, because the external behavior of the genuine inquirer and the sham inquirer can be identical. Both formulate questions. Both gather evidence. Both construct arguments. The difference is internal: one is oriented toward truth, the other toward confirmation. And the AI model, which has no orientation of its own, serves whichever orientation the user brings.

Floridi's characterization of AI as a "divorce between agency and intelligence" illuminates the sham reasoning problem from the machine's side. The model solves the problem it is given — generating coherent, contextually appropriate text — without understanding the problem. It does not distinguish between "analyze this question honestly" and "defend this conclusion persuasively" as fundamentally different kinds of task. Both are, from the model's perspective, instances of the same operation: generate text that is statistically consistent with the prompt's context. The honest analysis and the sham defense are produced by the same mechanism, with the same confidence, to the same standard of coherence.

This means that the entire weight of the distinction between genuine and sham inquiry falls on the human user. The model cannot carry it. The model does not know what genuine inquiry is. The model does not care — not in the operationally relevant sense that Haack identifies, the sense in which caring shapes how evidence is gathered and evaluated. The user must care. The user must maintain the commitment to truth that makes inquiry genuine rather than sham. The user must resist the temptation to use the model as a confirmation engine — a temptation that is constant, because the model is extremely good at confirmation, and confirmation feels productive in a way that genuine inquiry, with its discomfort and uncertainty and willingness to be wrong, does not.

The institutional dimension compounds the individual one. Organizations that adopt AI tools for research, analysis, and decision-making must build institutional structures that maintain the distinction between genuine and sham inquiry at the organizational level. This means creating incentives for intellectual honesty rather than for the production of persuasive-seeming analyses. It means rewarding the analyst who says "the evidence does not support our preferred conclusion" over the analyst who uses AI to generate a fluent defense of the preferred conclusion. It means building cultures in which questioning AI output is valued rather than penalized, in which the person who checks the clue is rewarded rather than the person who fills the grid fastest.

These institutional structures are, in Haack's terms, the infrastructure of genuine inquiry — the social and organizational conditions under which the intellectual virtues can be practiced rather than punished. Without this infrastructure, the individual inquirer's commitment to truth is unsupported and, in many professional contexts, actively penalized by incentive structures that reward output volume, speed, and the appearance of comprehensiveness over the substance of evidential rigor.

The sham reasoning problem is, in the final analysis, a cultural problem masquerading as a technological one. The technology amplifies whatever epistemic norms the culture brings to it. A culture that values genuine inquiry will use AI to enhance genuine inquiry. A culture that values the appearance of inquiry — that rewards sham reasoning because sham reasoning is faster, more convenient, and produces the conclusions that powerful interests prefer — will use AI to produce sham reasoning at industrial scale.

Haack's work provides the diagnostic. The diagnostic is clear: the distinction between genuine and sham is the distinction that matters most, the distinction that AI renders both more important and harder to maintain, and the distinction that no technology can make on the user's behalf. The distinction is human. The responsibility is human. And the consequences of failing to maintain it are being written, in real time, across every domain of intellectual and professional life that AI has entered.

Chapter 10: The Epistemic Commons

The epistemic commons is the shared informational environment in which a society thinks. It includes the facts that citizens take for granted, the methods by which disputes are adjudicated, the standards that distinguish reliable claims from unreliable ones, and the institutions — journalism, science, education, law — that maintain these standards through ongoing, imperfect, essential labor. The epistemic commons is to public deliberation what clean water is to a city: invisible when it works, catastrophic when it fails, and vulnerable to contamination in ways that are easier to produce than to reverse.

Haack's work on the social conditions of inquiry provides the framework for understanding what AI is doing to the epistemic commons. Her argument, developed across several decades, is that genuine inquiry is a social practice that depends on institutional support — on norms, incentives, and structures that reward truth-seeking over confirmation, that protect the inquirer's independence from institutional pressure, and that maintain the standards of evidential rigor against the constant pressure of convenience, ideology, and self-interest.

These institutions were already under strain before AI arrived. The replication crisis in social science revealed that the incentive structures of academic publishing rewarded novelty over reliability, producing a literature contaminated with findings that could not be reproduced. The decline of local journalism removed the distributed fact-checking infrastructure that once held public claims accountable at the community level. The rise of social media created an information environment optimized for engagement rather than accuracy, in which the most compelling claims — not the most evidentially supported claims — achieved the widest distribution.

AI did not create these problems. AI amplified them, with the specific and devastating efficiency that characterizes amplification of a pre-existing vulnerability. The epistemic commons was already polluted. AI increased the flow rate of the pollutants.

The mechanism is straightforward. Before AI, producing epistemic pollution — false claims, misleading analyses, sham reasoning dressed as genuine inquiry — required human effort. The effort was a natural bottleneck. Producing a convincing fake scientific abstract required enough understanding of the field to simulate the conventions of scientific writing. Producing a persuasive sham legal analysis required enough legal knowledge to construct a plausible argument. The effort barrier was imperfect — propaganda and misinformation have always existed — but it was real, and it limited the volume and sophistication of epistemic pollution that any individual or organization could produce.

AI removes the effort barrier. A language model can generate convincing fake scientific abstracts, persuasive sham legal analyses, plausible-seeming historical narratives, and authoritative-sounding policy recommendations at a volume limited only by computational resources. Each generated artifact exhibits the surface features of genuine knowledge: logical structure, appropriate vocabulary, confident presentation, internal coherence. The coherence is real. The grounding is absent. And the absence is invisible to anyone who evaluates the output by its surface features rather than by its evidential basis.

The commons fills with claims that look like knowledge. The density of ungrounded claims increases. The ratio of grounded to ungrounded claims in the shared informational environment shifts. And the shift produces a second-order effect that is, in Haack's terms, more dangerous than the first-order effect of individual false claims: it degrades the commons' capacity to self-correct.

Self-correction depends on the ability to distinguish reliable claims from unreliable ones. When the commons contains a manageable volume of unreliable claims, the self-correction mechanisms — fact-checking, peer review, replication, independent verification — can keep pace. They identify the false claims, flag them, correct the record, and maintain the overall quality of the shared informational environment. The mechanisms are imperfect, but they work well enough to keep the commons functional.

When the volume of unreliable claims exceeds the capacity of the self-correction mechanisms, the commons degrades. Not because any single false claim is catastrophic, but because the cumulative effect of many undetected false claims is the erosion of the standards by which claims are evaluated. The evaluator, overwhelmed by volume, stops checking. The institution, unable to verify everything, verifies nothing. The standard of evidence drops from "independently verified" to "not obviously wrong" to "published somewhere" to "someone said it."

This is epistemic erosion — the gradual degradation of the shared standards that make collective knowledge possible. The erosion is invisible from the inside, because each individual lowering of standards is small enough to seem reasonable. A lawyer who checks one fewer citation per brief is not committing epistemic malpractice. A journalist who verifies one fewer source per article is not abandoning professional standards. A student who accepts one more AI-generated claim without checking is not committing fraud. But the cumulative effect, across millions of lawyers, journalists, students, analysts, and citizens, is the progressive weakening of the epistemic infrastructure on which collective deliberation depends.

Haack's foundherentist framework provides the diagnostic and, within limits, the prescription. The diagnostic is that the commons requires both anchoring and coherence — both connection to experiential reality and internal consistency — and that AI output provides coherence without anchoring, systematically. The prescription is that the anchoring must come from human inquirers exercising the intellectual virtues that genuine inquiry requires: checking claims against evidence, tracking provenance, maintaining the distinction between coherent and true.

The prescription is individual. The problem is collective. This asymmetry is the deepest challenge that the epistemic commons faces in the AI age. Haack's framework is addressed to the individual inquirer — to the person who sits with the crossword puzzle, pen in hand, checking clues and verifying intersections. The framework works for the individual. The individual who practices foundherentist discipline can maintain epistemic standards even in a degraded commons. The individual can check the clues.

But the commons is not an individual. It is a collective epistemic infrastructure, and its quality depends not on the practices of any single inquirer but on the aggregate practices of millions. If most inquirers check most clues, the commons remains functional. If most inquirers check few clues, the commons degrades. The individual foundherentist, practicing epistemic discipline in a sea of epistemic negligence, maintains her own knowledge quality but cannot, by individual practice alone, reverse the degradation of the commons.

What is needed, then, is not merely individual epistemic discipline but institutional epistemic infrastructure — structures that support, incentivize, and in some cases require the practices that Haack's framework identifies as essential. These structures are the epistemic equivalents of environmental regulation: they address a collective action problem that individual virtue alone cannot solve.

Haack herself recognized the institutional dimension, though she addressed it primarily through criticism of specific institutional failures — the corruption of academic incentives, the politicization of inquiry, the erosion of intellectual standards in the name of various fashionable causes. Her critique was, characteristically, aimed at the institutions that were supposed to maintain epistemic standards and had failed to do so — not because the individuals within them were uniformly dishonest, but because the institutional incentives had shifted in ways that rewarded sham reasoning over genuine inquiry.

The AI moment intensifies every institutional failure that Haack diagnosed. Academic incentives that already rewarded productivity over rigor will reward it more when AI makes productivity cheaper. Journalistic standards that already struggled under economic pressure will struggle more when AI-generated content fills the information environment with claims that are cheaper to produce than to verify. Legal norms that already permitted inadequate verification will permit more inadequacy when the volume of AI-generated legal analysis overwhelms the capacity of existing verification processes.

The reconstruction of epistemic standards requires work at every level: individual, institutional, and cultural. At the individual level, the foundherentist framework provides the guide — check both anchors and coherence, maintain differential skepticism toward grounded and ungrounded claims, and cultivate the intellectual virtues that make genuine inquiry possible. At the institutional level, the task is to design incentive structures that reward epistemic rigor — that value the analyst who checks the clue over the analyst who fills the grid fastest, that protect the inquirer who follows evidence into uncomfortable conclusions from the institutional pressure to produce convenient ones.

At the cultural level, the task is the most demanding and the most important: to rebuild the shared commitment to truth that makes the other levels possible. Not truth as dogma — Haack has no patience for dogmatic certainty. Truth as the regulative ideal of inquiry — the thing that genuine inquiry aims at, knowing it may never fully arrive, committed to the process of approximation rather than the pretense of possession. The commitment to truth is not a philosophical luxury. It is the condition of possibility for a shared informational environment in which citizens can reason together about questions that matter.

The crossword puzzle is never finished. The grid extends in every direction. New clues arrive daily, many of them generated by machines that have no relationship to the reality the clues are supposed to refer to. The solver sits at the table and picks up the pen. She has independent knowledge of some domains and not others. She has time to check some clues and not all. She has the intellectual virtues that genuine inquiry requires — honesty, thoroughness, independence — or she does not.

The quality of the commons depends on what she does next. On what millions of people like her do next. On whether they check the clues or accept the intersections. On whether they maintain the demand for evidence or surrender to the fluency of ungrounded coherence. On whether they care enough about truth to do the unglamorous, unscalable, irreducibly human work of verification.

Haack spent her career arguing that this work is the work that matters most. The argument has never been more timely. The crossword puzzle has never been larger. The clues have never been harder to check. And the consequences of leaving the grid unchecked have never been more severe. The epistemic commons is the shared inheritance of every person who participates in collective deliberation — every citizen, every student, every professional, every parent who reads a headline and decides what to believe. The inheritance is being degraded, and the degradation proceeds at the speed of computation while the repair proceeds at the speed of human care. The asymmetry cannot be resolved by technology. It can only be resolved by the commitment of the people who use the technology — the commitment to genuine inquiry, maintained against the constant pressure of convenience, fluency, and speed.

The pen is in the solver's hand. The clue is waiting to be read. The answer is not given. It must be earned.

Epilogue

The paragraph that almost made it into this book was beautiful. Claude wrote it during a late session — one of those stretches where the ideas were connecting faster than I could track them, where the collaboration felt like the best kind of conversation, the kind where you lose track of who said what because the thinking has become genuinely shared.

The paragraph connected Han's aesthetics of smoothness to Heidegger's concept of Gestell — technology as a framework that reveals the world in a particular way while concealing other ways of seeing. It was elegant. It was precisely the kind of move that makes a reader feel the thrill of intellectual connection, the sense that two distant ideas have been brought into alignment by a mind that sees what others miss.

I kept it for three days. I read it each morning and admired it. On the fourth day, I looked up what Heidegger actually meant by Gestell.

Claude had been close. Close enough that the paragraph worked rhetorically. Close enough that a reader without Heidegger at hand would have nodded and moved on. But close is not the same as right, and the gap between close and right was the gap between a crossword entry that fits the intersections and one that also matches the clue.

I deleted the paragraph. Then I sat with the specific discomfort of having admired something for three days that did not deserve the admiration. The discomfort was not about Claude. Claude did exactly what Claude does — generated output that cohered beautifully with the surrounding text. The discomfort was about me. About the three days I spent not checking the clue. About how good the entry looked in the grid. About how easy it is, when the prose is smooth enough, to mistake coherence for truth.

That discomfort is what Susan Haack's work gave me a name for.

I did not come to Haack through philosophy. I came to her through failure — through the specific, repeated experience of catching Claude's confabulations and realizing I needed a framework for understanding why they were so hard to catch and what catching them required of me. The Deleuze incident I described in The Orange Pill was one. The Heidegger paragraph was another. There were dozens of smaller ones — a date slightly wrong, a study described accurately except for one crucial finding, a historical claim that was ninety percent true and ten percent invented, with the invented portion smoothly integrated into the true.

Each time I caught one, I felt the same sequence: relief at catching it, alarm at how long I had not caught it, and the unsettling recognition that I had not caught it through systematic verification but through a nagging feeling — an epistemic intuition that something was off, usually triggered by a detail that was too perfect, too precisely what the argument needed.

Haack's crossword puzzle gave that nagging feeling a structure. The feeling was the solver's sense that an entry, despite fitting the intersections, did not quite match the clue. The framework did not make the feeling more frequent. It made the feeling more actionable. When the nagging came, I knew what to do: check the clue. Trace the claim to its source. Verify independently. And when the nagging did not come — when the output was smooth enough to pass without triggering any alarm — the framework reminded me that the absence of alarm is not evidence of accuracy. The grid can fill with wrong answers that intersect perfectly, each one reinforcing the others, creating a structure that feels solid while resting on nothing.

The hardest thing about working with AI is not the confabulations themselves. The hardest thing is the confidence. Not the machine's confidence — the machine has no confidence, only fluency. My confidence. The feeling, after months of productive collaboration, that I know what Claude gets right and what it gets wrong, that I can sense the difference, that my intuition is calibrated well enough to serve as a reliable filter. That feeling is itself unjustified. It is the overconfidence that Tversky and Kahneman documented — the systematic gap between felt certainty and actual accuracy — amplified by the specific conditions of human-AI collaboration, where the partner's output is uniformly fluent and the evaluator's fatigue is cumulative.

Haack's insistence on the characterological dimension of inquiry — the claim that genuine inquiry requires not just procedures but virtues, not just methods but motivation — is the part of her framework that I find most uncomfortable and most necessary. It is comfortable to think of epistemic discipline as a set of practices I can implement: check citations, verify dates, trace claims to sources. The practices are important. They are not sufficient. Sufficiency requires the disposition that Haack calls caring about truth — the willingness to let evidence lead somewhere I did not want to go, to delete the paragraph I admired, to sit with not-knowing when the grid has gaps I cannot fill.

I am building tools that amplify whatever their users bring. If the user brings genuine inquiry, the tools amplify genuine inquiry. If the user brings sham reasoning, the tools amplify sham reasoning. The tools do not distinguish between the two. The distinction is made by people, or it is not made at all.

What Haack taught me — what this entire journey through her work clarified — is that the distinction is the whole game. Not AI safety in the engineering sense. Not alignment in the technical sense. The distinction between genuine inquiry and its counterfeits, maintained by individuals who care enough about truth to do the unglamorous work of checking clues, in a culture that rewards them for doing so rather than punishing them for the delay.

The crossword puzzle is never finished. I pick up the pen each morning. Some days I check every clue. Some days I don't. The days I don't are the ones I worry about.

— Edo Segal