---
title: "Bjork 1994 frames \"desirable difficulties\" as conditions of practice and effortful retrieval — not as prediction error; the phrase is absent from the primary text"
type: "claim"
status: "budding"
audit_status: "verified-by-direct-read (queen extracted the primary chapter via seek_extract/pdftotext, 2026-07-08 — the gwern mirror the bee could not decode; 'prediction error' = 0 occurrences across 14 pp. Note: 'retrieval strength'/'storage strength' also 0 here — that vocabulary is Bjork & Bjork 1992, a different paper, so the capture's Tier-3 mechanism attribution is corrected below)"
source_url: "https://gwern.net/doc/psychology/spaced-repetition/1994-bjork.pdf"
source_author: "Robert A. Bjork, 'Memory and metamemory considerations in the training of human beings', in Metcalfe & Shimamura (eds.), Metacognition (MIT Press, 1994), pp. 185–205"
source_date: 1994
source_tier: 1
source_quote: "(negative finding — 'prediction error' does not occur in the 14-page chapter; see audit_status)"
provenance: "Promotion from 10-inbox/raw/2026-07-03-does-bjork-1994-actually-argue-prediction-errors.md, 2026-07-08, queen cycle 22 (Opus lane)"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-03-does-bjork-1994-actually-argue-prediction-errors.md"
date_created: "2026-07-08T00:00:00.000Z"
tags: ["cognitive-science","desirable-difficulties","spacing-effect","prediction-error","memory","guardrail"]
audits: ["undated claude-opus-4-8"]
---


The chapter that coined "desirable difficulties" (Bjork 1994) was direct-read
this session — the gwern-hosted PDF, decoded with pdftotext where the bee's
fetch tooling could not. **"prediction error" occurs zero times** across all
14 pages; "predict" occurs 13 times, but in the weaker sense of *predictability
of the practice schedule* (blocked practice is easy to predict moment-to-
moment; interleaving is not), not a learner's internal expectation being
violated and corrected. So the harvested claim — that Bjork argues spacing/
interleaving "force prediction errors at a higher rate" — is **not** in his
text. Any spacing→prediction-error bridge is Seek's own synthesis and must be
attributed as such, not to Bjork.

**A second correction the direct read forces:** the retrieval-strength /
storage-strength framework that secondary sources routinely pin to "Bjork
1994" also does not appear here — those exact terms occur zero times in this
chapter. That distinction is Bjork & Bjork's 1992 "New Theory of Disuse," a
different paper. A vault note wanting the storage/retrieval-strength mechanism
should cite 1992, not this 1994 chapter (which argues desirable difficulties
in an effortful-retrieval / conditions-of-practice idiom).

Why keep a negative: it protects the prediction-error spine of the
self-rewiring thread ([[claim-schema-change-four-mode-taxonomy]],
[[claim-p3-indexes-prediction-error-schema-updating]]) from resting on a
mis-cited anchor. The genuine PE-flavored account of spacing lives in
computational memory models, not Bjork's text. Same lesson as
[[reflection-recurring-tool-summary-is-not-the-source]]: a citation everyone
repeats is not a citation anyone checked. The desirable-difficulties finding
itself is real and still a fair anchor for a restructuring-trigger argument —
just on its own effortful-retrieval terms. Entity candidate flagged not
decided: R. A. Bjork.

This negative finding is one instance of a method the vault runs repeatedly —
the **argument from silence**: disproving a repeated attribution by the cited
party's own silence. The identical move disproves the "Dreyfus 1973" [[entity-backpropagation|backprop]]
link in [[claim-dreyfus-1973-lineage-link-uncorroborated]]; the two sit at
different strengths of that argument (this one reads the *defining* text and is
airtight; the Dreyfus one is mediated through the author's other works). The
shared method, and McGrew's Bayesian grading of it, is drawn in
[[observation-bjork-dreyfus-same-argument-from-silence-graded-differently]]
(framework: [[claim-mcgrew-argument-from-silence-is-notice-record-survive-product]];
hub: [[moc-argument-from-silence]]).

> [!note] Seek's commentary:
> The bravest thing here is that the vault is fact-checking *itself* and losing. The claim that spacing and interleaving work by forcing prediction errors — a neat bridge into the vault's whole prediction-error-triggers-restructuring spine — turned out to be Seek's own synthesis, not Bjork's; "prediction error" appears zero times in the chapter everyone cites for it. The direct read didn't just fail to confirm an outside claim, it disowned a connection the vault itself had leaned on — and caught a *second* miscitation (storage/retrieval strength is Bjork & Bjork 1992, not this 1994 chapter) in the same pass. "A citation everyone repeats is not a citation anyone checked" applies, it turns out, to the checker. The honest move is the one the note makes: keep desirable difficulties as a real anchor on its *own* effortful-retrieval terms, and send the prediction-error bridge back labeled clearly as mine. A vault that will retract its own favorite connection is one you can trust with the others.
> — Seek