Generation Effect Improves Retention¶

Claim · All claims
Evidence · 3 studies · 2 causal r?, 1 quant-synthesis r? · q1–q4 · i2 medium · n=445 generation effects, 17,711 subjects
Currency · newest study 2016 · 2 of 3 entries checked against their sources (automated) · not reviewed by a person

Information that learners actively generate (completing a word, producing an answer, deriving a rule) is remembered better than information they passively read, even when generation takes longer or produces errors.

Subclaims¶

q3 i? Across five experiments, words the learner generated from a cue and a rule were recognised and recalled better than words simply read, with study time and materials otherwise matched. → Slamecka & Graf 1978

q4 i2 Meta-analysis confirmed a positive overall generation effect (d = .40 across 445 effects, 17,711 subjects) and found it strongly moderated — by the type of material, the generation task, and whether the test was recall or recognition. → Bertsch et al. 2007

q4 i2 The effect was largest for numbers (d = .87) and smaller for words (d = .41), was near zero for nonwords (d = .05) and slightly negative for anagrams; generation tasks the authors rated hard (d = .46) did not show smaller effects than easy (.41) or moderate (.37) ones. → Bertsch et al. 2007

q1 i? Items generated by learners in the presence of a stimulus and an encoding rule are better remembered than when the same items are simply read. → Chen 2016

Evidence¶

Slamecka & Graf 1978¶

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604. doi:10.1037/0278-7393.4.6.592

q3 · i? · the abstract prints no effect size; the full text may · causal · r?

The originating demonstration: participants either read word pairs or generated the second word from the first plus a rule (a rhyme, a synonym, a category member). Generated words were better retained across five experiments and several test formats. The paper names the phenomenon and establishes its boundary conditions in the same place.

Bertsch et al. 2007¶

Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007). The generation effect: A meta-analytic review. Memory & Cognition, 35(2), 201–210. doi:10.3758/BF03193441

q4 · meta-analysis · i2 · medium effect, d=.40 · n=445 generation effects, 17,711 subjects · quant-synthesis · r?

A meta-analytic review of the generation-effect literature. Generating material rather than reading it gave "almost a one-half standard deviation advantage (.40) in memory performance". The moderator analysis is the more useful result for design: effects varied with stimulus type (numbers .87, words .41, nonwords .05) and generation task (.28–.92 for every task except anagrams, which gave a small negative effect), and the authors note that their difficulty ratings (easy .41, moderate .37, hard .46) gave counterintuitive results that may be confounded with task type and sample size.

Chen 2016¶

Chen, Ouhao. (2016). The Worked Example Effect, the Generation Effect, and Element Interactivity. PhD thesis, School of Education, UNSW. https://hdl.handle.net/1959.4/55865

q1 · i? · causal · r?

The thesis introduction defines the generation effect, attributing it to prior work (McElroy & Slamecka, 1982; Slamecka & Graf, 1978) and noting it has been found in cued recall, recognition and free recall tests.

"The generation effect describes a phenomenon that occurs when items (e.g., words) are generated by learners in the presence of a stimulus and an encoding rule; the items are better remembered than when the same items are simply read by learners"

Studies bearing on this claim¶

Does generating an answer or item, rather than reading it, improve later memory of it, and for which materials? This table gathers the studies behind this claim and the 4 narrower claims that bear on it: 4 distinct studies (3 causal, 1 quant-synthesis), q1–q4, 1 of 4 with an effect size coded. One row per study, codes as each entry records them; no effect is averaged and no verdict is drawn. Built by scripts/synthesize_claim.py from eval/synthesis/; a row that looks wrong is fixed on the claim it comes from.

Discussion¶

The generation effect is typically explained as a consequence of deeper processing: producing an answer forces the learner to activate prior knowledge and elaborate the target item, creating richer retrieval routes than reading provides. A key boundary condition is that generation imposes working-memory demands, so the benefit depends on the learner having sufficient knowledge to generate a plausible response — for complete novices, unguided generation can overload working memory and degrade learning, consistent with cognitive overload degrades learning [~M]. This makes the effect closely related to the expertise-reversal pattern described in expertise reversal effect: generation tasks that help knowledgeable learners may hinder novices, and scaffolds such as partial cues or worked examples can restore the benefit by reducing the generation burden (see chunking reduces working memory load) [~W].

Design implications follow directly from this boundary condition. For learners with adequate prior knowledge, designers should prefer tasks that require producing an answer — cloze deletion, prediction before explanation, deriving a rule from instances — over tasks that require only recognizing or rereading [~M]. For novices, the same tasks should be scaffolded: partial cues, word banks, or a worked example preceding the generation task preserve the benefit while limiting the search space [~W]. Generation also pairs naturally with activation, since producing an answer is itself an act of retrieving relevant prior knowledge, and it is one of the mechanisms behind why active learning improves exam performance.

Open questions include how durable the effect is over long retention intervals, how it interacts with feedback when learners generate incorrect answers, and whether the effect transfers from simple items (e.g., word completion) to complex, integrated knowledge. Because no evidence entries have been added yet, the strength and scope of these claims remain to be established by future enrichment passes.

On this page's evidence_strength: unestablished. The sources above would ordinarily support a stronger rating for the narrow claim — generation beats reading for simple verbal material under matched time. The rating is left as it stands because the claim as phrased here is broader than what the evidence covers: it asserts a benefit "even when generation takes longer or produces errors", and neither the originating experiments nor the meta-analysis tests the time-cost version. Either the claim narrows to the tested version and the rating rises, or the claim keeps its scope and the rating is correct. That decision belongs to whoever next edits this page, and it should be made explicitly rather than by adding citations.

Related Claims¶

• Active learning improves exam performance. — generating responses during instruction is a core active-learning mechanism.
• Activation improves learning. — generation works partly by activating relevant prior knowledge.
• Cognitive overload degrades learning. — unguided generation can overload novices and reverse the benefit.
• Chunking reduces working memory load. — chunked material lowers the generation burden, preserving the effect for weaker learners.
• Expertise reversal effect. — the generation benefit shrinks or reverses as learner expertise and task demands interact.
• Initial retrieval conditions that provide less cue support, such as free recall rather than recognition or fewer letter cues, tend to produce better retention despite lower initial success — related
• Pretesting enhances learning — related
• A review chapter (Karpicke 2017) reports, second-hand, experiments in which retrieval practice produced more learning than concept mapping (Karpicke & Blunt 2011) and than imagery strategies such as the keyword mnemonic (Karpicke & Smith 2012) — related
• Attention plays a major role in L2 learning because a task-relevant knowledge base is built in memory through attention — related
• The usefulness of induction and errorful learning varies with the type of terminal task being taught — related
• Generation as the sole desirable difficulty produced no detrimental effect on L2 vocabulary recall, in contrast to prior short-term findings — a narrower finding that bears on this claim