Generation effect
A word a learner produces from a cue and a rule is remembered better than the same word read off the page.
A word a learner produces from a cue and a rule is remembered better than the same word read off the page.
The paradigm holds the words constant and varies only how they arrive. A learner sees rapid-fast and reads it, or sees rapid-f under the rule synonym and says fast. The same item serves both conditions across learners, so nothing about word choice can explain the difference. Two features separate this from the testing effect, which the atlas treats as a neighbouring entry.
Generation is a manipulation of encoding, not of practice after encoding, and the item being produced comes out of semantic memory rather than out of the study episode. Slamecka and Graf raised that second point themselves, noting that generating might be an instance of recall with semantic memory as the source, and that such an account comes close to being only a restatement of the finding it seeks to explain.
The effect is measured against reading, and what reading means is set by the experimenter. Begg, Vinski, Frankovich and Holgate told their readers how to study, and the advantage of generating held against readers who pronounced the words and vanished against readers who imagined them.
Replicated McCurdy and colleagues pooled 1,653 estimates from 310 experiments and found generated items remembered .102 more often than read items, 95% CI [.087, .117].
Two meta-analyses reach the same place by different routes. Bertsch, Pesta, Wiscott and McDaniel searched PsycINFO from 1966 to 2005, asked primary researchers for file-drawer studies, and pooled 445 effect sizes from 86 studies covering 17,711 subjects. Weighted mean proportions were .471 read against .559 generated, an advantage of .088, which they standardised against the read condition to give .40, 95% CI [.38, .42].
Letting each experiment contribute a single mean moved that to .41, 95% CI [.39, .43]. Sampling error accounted for 11% of the observed variance, so the pooled figure hides real differences between conditions. McCurdy, Viechtbauer, Sklenar, Frankenstein and Leshikar covered 126 articles, 310 experiments and 1,653 estimates, and reported the raw difference rather than a standardised one: .102, 95% CI [.087, .117].
Their moderators track Bertsch's. Recognition gave .144, 95% CI [.121, .167], cued recall .132, 95% CI [.076, .189], and free recall .095, 95% CI [.042, .148]. Within-subjects designs gave .144, 95% CI [.126, .161], t(119) = 16.24, p < .001, and between-subjects designs a smaller but still reliable .073, 95% CI [.031, .114], t(119) = 7.58, p < .001.
Both meta-analyses test the same old objection. Slamecka and Katsaiti proposed that the effect is an artefact of learners abandoning rehearsal of read items when read and generated items share a list. Bertsch and colleagues note the theory predicts no effect in between-subjects free recall and report a 4% advantage there, d = .17. McCurdy and colleagues state the prediction for pure lists and found .103, 95% CI [.071, .149] against .144, 95% CI [.125, .162] for mixed lists. Correcting for small-study effects by extrapolating to infinite sample size cut the overall estimate to .080, 95% CI [.048, .111], t(122) = 5.00, p < .001.
The covered columnillustration
A learner covering the translation and producing it from the first letter remembers more of the list than one reading both columns. The words are identical. What differs is whether the word arrived from the page or came out of the learner's own memory at the moment of first study.
Working out the missing ingredientillustration
Someone who reads a method and works out which ingredient must be missing from the list remembers it later; someone who reads the finished list does not. This is not practice after learning, because both met the recipe once.
Filling in the drug nameillustration
A student given the class of a drug and its first two letters, and asked to supply the rest, recalls the name better than one who reads the chart. The gain lands on the word that was produced, not on the cue that prompted it.
Cases marked as illustrations describe nobody in particular. They are there to make the pattern recognisable, not to report that a particular thing happened.
Slamecka and Graf ran five experiments on University of Toronto undergraduates. Experiment 1 used 24 subjects and a 100-item list, 20 items under each of five rules: associate (lamp-light), category (ruby-diamond), opposite (long-short), synonym (sea-ocean) and rhyme (save-cave). Generation errors ran at a median 6%, and those items were dropped from scoring. Generated words were recognised more often, F(1, 20) = 9.68, MSe = .07, and were rated with higher confidence, F(1, 20) = 6.92, MSe = 1.04.
Pacing did nothing: timed and self-paced means of .75 and .79, F < 1. Rules mattered for overall level, F(4, 80) = 5.28, because rhyme responses sat lowest, but rules did not interact with generating, F(4, 80) = 2.09. Experiment 2 moved the contrast inside the subject, 12 people seeing both conditions, and the effect held at F(1, 10) = 27.17, MSe = .04.
Warning subjects about the test changed nothing, with informed and uninformed means of .72 and .77, F < 1. Asked to label each recognised word as generated or read, subjects were right 74% of the time against a chance value of 50%, t(11) = 6.00, SE = .04. Experiment 3 asked whether the benefit spreads to the cue.
Twenty-four subjects saw 66 rhyme pairs and were tested on either the stimulus or the response side, with the untested member always supplied. Recognition gave a main effect of generating, F(1, 20) = 26.19, MSe = 50.99, and an interaction with the tested side, F(1, 20) = 10.75, MSe = 50.99. A t test found the effect on responses, t(20) = 5.94, SE = 2.92, and none on stimuli, with no statistic printed.
The confidence ratings were analysed separately and gave t(20) = 4.82, SE = .10, on the response side against t < 1 on the stimulus side. Experiment 4 put 12 subjects through five cycles of study and free recall of 60 responses. Generated words won, F(1, 11) = 19.84, MSe = .05, and the two curves converged over the last two trials, F(4, 44) = 4.47.
Output position ruled out one artefact: generated items were not recalled earlier, with mean output positions of 13.45 against 12.74 across all trials. Experiment 5 ran 24 subjects through five cycles of cued recall on the 66 rhyme pairs. Response recall was .68 generated against .61 read, and stimulus recall was .53 against .51, F(1, 22) = 6.19, MSe = .02.
The response-side selectivity found in Experiment 3 did not reach significance here, F(1, 22) = 1.82. The generate and read condition means for the recognition experiments are printed in figures rather than in the text, so those proportions are not quoted here.
1 primary research5 replication and reanalysis2 commentary and critique
The paper is "The generation effect: Delineation of a phenomenon", and it names the effect. Slamecka and Graf described their own purpose as empirical rather than theoretical, writing that a search of the journals had uncovered no thoroughgoing treatment of the question and no cumulative body of literature. They raised four interpretations, listed three more in a sentence, wrote that the list was not exhaustive, and declined to press for any of them until the analytic studies were done.
The claim, and the replication this entry reads its figures from.
The effect has boundaries the pooled figures bury. In Bertsch and colleagues' table, nonword stimuli gave .05, 95% CI [.03, .07], against .41 for words and .87 for numbers. Anagrams reversed it, -.05, 95% CI [-.07, -.03], across 18 studies and more than 1,000 subjects. Lists longer than 50 items gave .09, 95% CI [.07, .11], against .60 for lists of 25 or fewer.
Two things about their intervals deserve caution. Nearly every confidence interval in that table is the estimate plus or minus .02, whether the row rests on 15 effect sizes or 445. One row prints an effect size of .37 beside an interval of .49 to .53, which does not contain it. Schmidt and Cherry reported five between-list experiments, starting with 222 undergraduates in the first, though their own abstract says four.
Generation hurt free recall of word pairs, cued recall of the second word from weak associates, pair integration and organisation in recall, and helped recognition of the second word and free recall of second words alone. Their reading is that generation buys individual-item processing at the cost of processing relations. The educational case is the weakest part, and the word-level pooled figures do not reach it.
Schindler, Richter and Mar write that Bertsch and colleagues included no studies using full text generation, and McCurdy and colleagues code stimuli only as words, nonwords or numbers. Schindler and Richter pooled 129 effect sizes from 20 studies, nested in 74 samples and 3,551 participants, for a text generation effect of g = 0.41, 95% CI [0.31, 0.52], with heavy heterogeneity, Q(128) = 528.26, p < .001.
Within that, texts of 900 words or more gave g = 0.05, 95% CI [-0.26, 0.37] on 12 effect sizes, and sentence unscrambling of expository texts gave g = 0.77, 95% CI [0.40, 1.13] on 11. The same authors then tried to reproduce the second figure. Across seven experiments varying intentionality, time limits, retention interval and design, with analysed samples from 40 to 152 and materials of 12 six-sentence texts from a psychology textbook chapter on Bandura, they found no positive generation effect on prompted recall anywhere.
Four experiments favoured reading: Experiment 3, OR = 0.69, 95% CI [0.53, 0.90]; Experiment 5, OR = 0.80, 95% CI [0.70, 0.91]; Experiment 7, OR = 0.76, 95% CI [0.65, 0.89]; and Experiment 4 marginally, OR = 0.80, 95% CI [0.64, 1.00], p = .048. On multiple-choice questions Experiment 5 again favoured reading, OR = 0.93, 95% CI [0.86, 1.00], while Experiment 6 went the other way, OR = 1.18, 95% CI [1.01, 1.38].
Their conclusion is that applications of text generation in education are rather limited. A registered direct replication by Schindler, Richter and Mar cuts the other way on the task itself. They re-ran McDaniel, Einstein, Dunay and Cobb's 1986 Experiment 2, recruiting 312 English-speaking and 300 German-speaking students and analysing 250 and 197 after dropping those who reported expecting the test.
The genre by task interaction that paper is known for did not appear. Recall was better for the narrative, F(1, 244) = 200.14, p < .001, and learning condition mattered, F(2, 244) = 7.90, p < .001. In a mixed model, unscrambling beat reading in both samples, English beta = 0.33, z = 2.29, p = .022, and German beta = 0.44, z = 3.65, p < .001, while letter completion fell below reading, English beta = -0.34, z = -2.75, p = .006, and German beta = -0.21, z = -2.36, p = .019.
Sources searched were Crossref, Europe PMC, Unpaywall, Semantic Scholar and general web search, together with the full texts of Slamecka and Graf, Bertsch and colleagues, McCurdy and colleagues, Schmidt and Cherry, Begg and colleagues, Schindler and Richter's meta-analysis and replication series, and Schindler, Richter and Mar's direct replication. No multi-site or many-labs replication of the word-level effect was located on 21 September 2026.
Each of these is named in this entry's own prose as something it gets taken for. Where the two came out differently, that is worth knowing before you quote either.
The original paper is read as showing that generating helps everything, and it does not. Slamecka and Graf had written that if generating also helped the cue, the name generation effect would be a misnomer, since cues are never generated. Experiment 3 found the benefit falling only on the produced word. Experiment 5 then failed to find that selectivity under cued recall, and they wrote that it might be prudent to defer a conclusion until a replication using more subjects was available.
A search on 21 September 2026 of Crossref, Europe PMC and the open web, including Slamecka's own later papers on the effect with McElroy in 1982, Fevreiski in 1983 and Katsaiti in 1987, located no such larger repeat. The second error is running the two numbers together. Bertsch's .40 is a standardised difference and McCurdy's .102 is a raw difference in proportion of items remembered, so they are one finding in two currencies rather than two findings.
The third error is treating effort as the explanation and reasoning that harder generation must help more. Bertsch and colleagues found incidental learning at .65 against intentional at .32. Their difficulty coding runs .41 easy, .37 moderate and .46 hard, so the ordering is not monotonic, and the hard row rests on 17 effect sizes and 685 subjects against 257 and 9,638 for easy.
Their verdict is that effort is too broad a construct to explain the effect. Generating a whole target did beat generating part of it, .55 against .32, which is the one result that fits effort. The fourth is confusing this with the testing effect. Retrieval practice compares testing against restudying material already seen; generation compares producing an item against reading it, at the moment of first encounter.
A design that restudies is not a design that generates. A fifth error is to assume learners underrate generating, as they underrate testing in entry 98. Begg, Vinski, Frankovich and Holgate had subjects rate each item for memorability on a 7-point scale. Pure generators gave 4.82 against 3.72 from pure readers, F(1, 46) = 19.4, MSe = 2.96, and within mixed lists generated items drew 4.31 against 3.70, F(1, 25) = 23.4, MSe = 0.41.
Recognition in the same subjects showed no advantage, .84 generated against .87 read between subjects, F(1, 46) = 1.58, and .85 against .83 within, F(1, 25) = 3.34, p < .08. The last error is the one with practical cost. The word-level effect survives 310 experiments, and moving it onto prose gives a mixed answer. Schindler, Richter and Mar's direct replication found sentence unscrambling beating reading across both genres, while the same team's seven conceptual replications on textbook material produced four reading advantages on prompted recall and none the other way.
Every claim on this page was held against these sources on 2026-09-21. Nothing here is written from memory. Written and checked by Krishna Chagti; how entries are written and corrected. Found a mistake? Report it, with the sentence and a source.