# When Answers Are Cheap, Understanding Becomes the Scarce Good *Information Anxiety 2* was written in 2000, before search engines were good and a quarter-century before large language models made fluent answers nearly free. Read in 2026, the book splits cleanly: its **diagnosis strengthens** while its **cures partly invert**. This is the most valuable thing the book does for a present-day reader, and it is entirely a product of drift — Wurman did not see it coming; the passage of time revealed it. The diagnosis strengthens because Wurman's least intuitive move — reframing the "information explosion" as a **non-information explosion**, most of the flood failing to inform (PDF-01-06) — describes the generative-AI condition almost perfectly. When a machine can produce a plausible paragraph on any topic in a second, the volume of text that looks like information and isn't explodes by orders of magnitude. His core claim, that the bottleneck was never volume but understanding, becomes *more* true precisely as answers become *more* abundant: when any answer is a prompt away, the scarce and valuable thing is no longer the answer but the understanding that lets you know *which* answer, *why* it holds, *where* it fails, and whether it is even the right question. Cheap answers commoditize information and, by contrast, raise the premium on understanding. The cures partly invert because they were built for a scarcity that no longer exists. Wurman's Chapter 8, "Finding Things" — maps, search mechanics, the GILS standard — solved the problem of *locating* information when locating was hard. That problem is largely gone; retrieval is solved and over-solved. So the part of the book that was once most practical is now most dated. Meanwhile the part that sounded softest — organize around interest, reject most of the incoming flood, anchor the new to the known — becomes the operational core, because in an environment of infinite cheap generation, *aggressive interest-filtering and structure are the only things standing between you and synthetic noise*. And there is a twist Wurman could not have anticipated: the filtering and structuring he prescribed as human discipline is now partly **automatable** — this very analytical system is a machine doing Wurman's "organize and reject" at scale — which both validates his prescription and moves the human's job up a level, from doing the filtering to judging the filter's output. This is the same compression described in [[AI Compresses the Knowledge-Worker Pipeline From Both Ends]]: when the machine handles capture, summary, and first-pass synthesis, the human work concentrates at the ends — framing the right question going in, and exercising judgment on what comes out. Both ends are *understanding* work, not *information* work. Wurman's thesis, ported forward, says exactly this: the durable human contribution is understanding, and understanding is what does not get cheaper when answers do. <!-- atomicity: 1 idea ✓ --> ## Why This Matters For **DevRel**, this is close to a strategy. When an LLM can generate boilerplate, snippets, and surface answers about any product instantly, the differentiated value of a developer-relations function is not more content (the machine floods that) but *understanding scaffolding* — the authored paths, the "why," the failure modes, the right questions — that turn cheap answers into real comprehension. The moat is understanding, and it is defensible exactly because it does not commoditize. For **my PKM practice**, the implication is to stop optimizing for capture and retrieval (solved) and optimize for the two ends AI doesn't do: choosing what is worth understanding, and judging what the machine returns. The vault's job shifts from storage to a thinking partner for synthesis. For **knowledge work broadly**, the implication is a revaluation: roles defined by producing or finding information face commoditization; roles defined by understanding — framing, judging, deciding, teaching — appreciate. ## Counterpoints to Hold - **AI may erode understanding too, not just information.** If models get good enough at judgment and framing, the "scarce good" frontier moves again; understanding's premium may be temporary. - **Cheap answers can build understanding.** Used well, instant answers are anchoring material (see [[You Understand the New Only by Anchoring It to the Known]]); the dichotomy of "answer vs understanding" can be overdrawn. - **Over-reliance risk.** Outsourcing synthesis to AI can *prevent* the understanding this note prizes — the externalization failure mode ([[Information Capture Without Synthesis Is a Form of Externalization]]). The scarcity of understanding could deepen unequally. - **"Understanding" is under-defined here.** The note treats it as obviously valuable without specifying which kinds resist automation; that's the load-bearing unknown. ## Links - [[3. Resources/Permanent Notes/Relational Linking Is a Distinct Organizing Family From LATCH's Attributes]] — applies-in: computed retrieval can cheaply order answers, while explicit relational links preserve the understanding behind why ideas belong together - [[3. Resources/Permanent Notes/Rewarding AI Usage Manufactures Usage Not Value]] — evidence-for: abundant AI activity makes usage easy to count while leaving the scarce outcome, demonstrated value, unmeasured - [[3. Resources/Permanent Notes/Implementation Is Cheap So Taste Becomes the Bottleneck]] — applies-in: cheap implementation raises the value of the judgment that distinguishes a coherent product from abundant output. - [[Adversarial Practice Builds Transferable Confidence Through Worst-Case Exposure]] — applies-in: when encouraging rehearsal is cheap, honest resistance is the scarce input that tests whether skill transfers. - [[AI Search Inherits the Trust and Fragility of the Sites It Cites]] — extends: cheap retrieval makes provenance and source judgment scarce in the same way cheap generation makes understanding scarce. - [[AI Literacy Is Relational Not Just Technical]] — extends: when answer production is cheap, literacy shifts toward judging when to trust, doubt, and repair after a misleading answer. - [[3. Resources/Permanent Notes/A Project Is the Community Not the Code]] — extends; when code becomes cheap, the hard-to-copy understanding and relationships of a project community become the scarce complementary asset. - [[3. Resources/Permanent Notes/A System Failure That Looks Like a Performance Failure Never Gets Fixed]] — applies-in; cheap answers can hide a broken knowledge system by making its symptoms look like individual judgment failures. - [[AI Compresses the Knowledge-Worker Pipeline From Both Ends]] — the pipeline-compression twin. - [[You Understand the New Only by Anchoring It to the Known]] — what understanding consists of, and why AI doesn't supply it for free. - [[Information Overload Is at Least as Much a Structure Problem as a Volume Problem]] — the filtering half, now partly automatable. - [[Information Capture Without Synthesis Is a Form of Externalization]] — the failure mode of outsourcing synthesis to the machine. - [[Personal Knowledge Management MOC]] — cluster context. - [[The Essential Role of Deep Work in Learning Hard Things Quickly]] — applies-in; distraction-free attention is the practice that converts available information into understanding. - [[3. Resources/Permanent Notes/Bubbles End From Cost Not From Lost Belief]] — contrasts: cheap answers can preserve the demand-side value story even when the supply-side financing required to produce them breaks. - [[3. Resources/Permanent Notes/Cheap Inspection Collapses the Signal-Substance Gap]] — extends: cheap generation raises the premium on understanding, while cheap verification raises the premium on inspectable substance - [[3. Resources/Permanent Notes/Curated Mastery Conferences Serve Practitioners That Volume Conferences Cannot]] — applies-in: curated practitioner depth becomes more valuable when broad introductory answers are abundant. - [[Perm · Content publishing at scale now requires automated curation filters, not just volume]] — applies-in: abundant production makes editorial understanding and selection more valuable than publishing frequency - [[3. Resources/Permanent Notes/Shared AI Tools Homogenize Output While Feeling Original to Each User]] — applies-in: shared-model fluency makes output abundant and similar, increasing the value of distinct judgment and understanding - [[3. Resources/Permanent Notes/Fluency Makes Small Errors Harder to Catch Than Large Ones]] — extends (added 2026-08-01): the mechanism by which judging got harder. Cheap generation severed the fluency-competence correlation readers spent a lifetime calibrating, so the surface no longer carries information about accuracy. - [[3. Resources/Permanent Notes/A Demo-Only Rubric Selects for Developers Who Cannot Explain What They Shipped]] — has evidence from: score the explanation as part of the rubric, not as a follow-up (Saha). - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists it under judgment, review and taste as the new bottleneck. - [[3. Resources/MOCs/Open Source Governance MOC]] — applies-in: the MOC cites it as the open-source twin of the community-over-code premise. ## Open Questions - Which specific kinds of understanding are most automation-resistant (framing? causal models? value judgment? tacit/embodied knowledge?), and for how long? - If filtering is automatable, what exactly is the human's residual judgment task, and how is it taught? - Does the understanding premium accrue broadly or concentrate (another Matthew effect), and what would broaden access to it? ## Source - [[4. Archives/Book reviews/Information Anxiety 2]] — §10.4 (drift re-score, primary finding) - [[4. Archives/Book reviews/Information Anxiety 2 — Source Passages]] — PDF-01-06, PDF-00-01 - Wurman, *Information Anxiety 2* (2000), Ch 1, 8 (read against the post-2022 condition)