The Crossword Puzzle (Epistemology)

In the AI Story

Haack introduced the crossword puzzle in Evidence and Inquiry (1993), Chapter 4, as the central analogy for her foundherentist framework. The image answered both foundationalism and coherentism. Against foundationalism: clues (experiences) do not self-justify—their interpretation depends on what else is in the grid. Against coherentism: intersections (coherence) do not suffice—the grid can be filled with answers that cohere perfectly while every entry is wrong because none match their clues. The crossword requires both dimensions operating simultaneously. Haack specified how the analogy maps onto epistemic structure. Clues = experiential anchors (observations, data, perceptual inputs). Entries = beliefs. Intersections = logical and probabilistic relationships among beliefs. The solver = the inquirer. The filled grid = the inquirer's total epistemic state at a given time. Checking an entry = evaluating whether a belief is justified. The solver's work—proposing entries, checking them against clues, verifying intersections, revising when conflicts emerge—is the work of inquiry.

The puzzle's dynamic properties illuminate how justification changes with new evidence. As more entries are filled (more beliefs justified), the grid becomes more constrained. Each new well-justified entry provides additional crossings for evaluating subsequent proposals. A partially filled grid in which every entry has been carefully checked is a more reliable structure for justifying new beliefs than an empty grid. This is the coherentist insight preserved: beliefs support one another, and the web's overall strength is a source of justification. But the support is conditional on the entries being individually grounded. A grid filled carelessly—entries accepted because they fit intersections without checking clues—becomes an internally coherent structure of unjustified beliefs. Coherence without grounding is a well-constructed crossword with every answer wrong. The puzzle makes visible what coherentism alone cannot explain: why a perfectly coherent system can be perfectly false.

Applied to AI, the crossword puzzle provides the operational model for human-AI collaboration. The model generates proposed entries—candidate beliefs, hypotheses, analyses. The entries cohere (the model is optimized for coherence). The human solver checks them: Does this entry match its clue? (Is there independent evidence supporting this claim?) Does it intersect correctly? (Is it consistent with other verified beliefs?) If yes to both, the entry is justified and goes into the grid. If no to either, the entry is rejected or flagged for further investigation. Segal's discipline—catching the Deleuze fabrication by checking the clue (Deleuze's actual work) when the intersection felt too smooth—is textbook foundherentist practice. The entry cohered beautifully. The clue did not match. Coherence-checking alone would have accepted it. Anchor-checking caught it. The practice is demanding: it requires the solver to possess (or acquire) independent knowledge, to resist the seduction of fluent coherence, to maintain the discipline of checking both clues and crossings when the model makes checking feel unnecessary by delivering outputs that appear already verified.

The crossword puzzle's third lesson—that partial completion is progress—provides the developmental model for building reliable epistemic structures in the AI age. Start with a few carefully verified entries (core domain knowledge, independently confirmed facts). Use those entries as crossings to evaluate new proposals. Each verified addition strengthens the grid, providing more reliable intersections for subsequent checking. The grid grows—slowly, carefully, with every entry earning its place. This is the only epistemically sound path from zero knowledge to comprehensive understanding in an environment saturated with ungrounded but coherent AI outputs. The alternative—accepting entries rapidly without checking, filling the grid at computational speed—produces a structure that looks complete and is fundamentally unreliable. Every unchecked entry is a point of potential failure. Every crossing built on an unchecked entry propagates the failure. The grid becomes a coherent fantasy, and the solver—surrounded by hundreds of intersecting entries that all support one another—has no internal signal that anything is wrong. The only signal is external: the clue, checked against reality, revealing the mismatch. The checking is the work that matters most. The crossword puzzle makes the work visible.

Origin

The crossword puzzle as epistemological model is Haack's original contribution—no prior epistemologist used the image with this structural precision. Earlier metaphors (Descartes's building, Quine's web, Otto Neurath's boat) captured aspects of justification but not the dual requirement of grounding and coherence operating simultaneously. Haack's puzzle does. The image appeared in Evidence and Inquiry (1993) and became the signature of her epistemology, reprinted in anthologies, cited across the epistemological literature, and recognized as one of the most successful philosophical analogies of the late twentieth century. Its pedagogical power lies in its accessibility—anyone who has solved a crossword understands the constraint structure—and its precision—the analogy maps onto justification's architecture without significant remainder.

The AI application is not part of Haack's original development but follows directly from the framework's structure. If justified belief is like a crossword entry (matching clue, intersecting correctly), then AI-generated claims are like proposed entries that have been checked for intersections but not for clues. The analogy extends cleanly: RAG is partial clue-checking (grounding some entries in retrieved documents). Confabulation is proposing entries that intersect perfectly but are made up. Verification is checking the clue. The Susan Haack—On AI simulation applies the crossword model to diagnose the epistemic structure of AI outputs and prescribe the evaluator's discipline—check both, continuously, with the care the stakes demand.

Key Ideas

Dual requirement. Justified belief must match its experiential clue (grounding) and intersect correctly with other beliefs (coherence)—neither alone suffices.

Clues constrain without determining. Experience narrows acceptable answers without reducing them to one—interpretation depends on grid context.

Grid becomes self-reinforcing. Each verified entry strengthens justification for neighbors—partial completion is epistemic progress.

AI generates intersections, not clues. Models optimize for coherence (crossings) without experiential anchoring—producing entries that fit the grid but may not match reality.

Checking is irreducible human work. The solver must verify both clue-matching and intersection-coherence—no source, however reliable, eliminates this responsibility.