The benefit of generating errors during learning

Kornell et al. found that a test condition produced better final test performance than a read-only condition.

“Kornell et al. found that the test condition led to better final test performance than the read-only condition.”

• cites Kornell et al., 2009 sourceSupported

Generating responses can improve memory even when many errors occur, provided corrective feedback is given.

“On the other hand, there is also evidence that generating responses can be beneficial even when many errors are produced, as long as corrective feedback is given (e.g., Kornell, Hays & Bjork, 2009).”

• cites Kornell, Hays & Bjork, 2009 sourceSupported

Feedback can enhance the benefits of testing.

“However, there is evidence that feedback may enhance the benefits of tests (Butler & Roediger, 2008;Pashler, Cepeda, Wixted & Rohrer, 2005).”

• cites Butler & Roediger, 2008 sourceSupported
• cites Pashler, Cepeda, Wixted & Rohrer, 2005 sourceSupported

Kane and Anderson found that generating errors improved final test performance over reading even in an undetermined condition where participants almost always erred.

“Even in the undetermined condition, where participants nearly always produced an error, generating led to better final test performance than reading.”

• cites Kane & Anderson, 1978 sourceSupported

Tests that are harder and require more retrieval effort produce greater benefits for later memory.

“Moreover, it has been found that the harder the test, and the greater the effort required for retrieval, the greater the benefit to subsequent memory (e.g., Carpenter & DeLosh, 2006;Pyc & Rawson, 2009).”

• cites Carpenter & DeLosh, 2006 sourceSupported
• cites Pyc & Rawson, 2009 sourceNot checked — no full text of the source

It is difficult to explain Kane and Anderson's (1978) findings by Grimaldi and Karpicke's account, because in their undetermined condition the targets were unrelated to the most obvious incorrect completions, so retrieval along with them is highly unlikely.

“Furthermore, it is difficult to explain Kane and Anderson's (1978) findings by Grimaldi and Karpicke's account, since in their undetermined condition the targets were unrelated to the most obvious, incorrect, completions so it is highly unlikely that they would have been retrieved along with them.”

• cites Kane and Anderson's (1978) sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Greater attention is given to feedback that does not match expectations.

“There is also evidence that greater attention is given to feedback that does not match expectations (e.g., Butterfield & Metcalfe, 2001;Fazio & Marsh, 2009)”

• cites Butterfield & Metcalfe, 2001 sourceNot checked — no full text of the source
• cites Fazio & Marsh, 2009 sourceSupported

Kornell et al. found no advantage of generating over reading when total trial time was equated.

“Kornell et al. found no advantage of generating over reading when total trial time was equated.”

• cites Kornell et al., 2009 sourcePartly supported

  Experiment 2 held total trial time constant and found no significant difference in final recall for fictional items (Method and Results and Discussion, p. 4). But the claim presents that result as general: Experiments 4 and 6 also equated trial time and found significantly better recall after retrieval attempts than after reading (Results and Discussion, pp. 6–7). (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Ease of processing, or encoding fluency, influences JOLs.

“However, it has also been found that ease of processing, or encoding fluency, influences JOLs (e.g.Castel, McCabe, & Roediger, 2007;Koriat, 2008;Hertzog, Dunlosky, Robinson & Kidder, 2003;Schwartz, Benjamin & Bjork, 1997).”

• cites Castel, McCabe, & Roediger, 2007 sourceSupported
• cites Koriat, 2008 sourceSupported
• cites Hertzog, Dunlosky, Robinson & Kidder, 2003 sourceNot checked — no full text of the source

In Grimaldi and Karpicke's and Huelser and Metcalfe's studies, participants generated guesses that were highly related to the cues in both related and unrelated conditions.

“In both Grimaldi and Karpicke's (2012) and Huelser and Metcalfe's (2012) studies, participants generated guesses which were highly related to the cues in both the related and unrelated conditions (e.g., sleep for the cue pillow).”

• cites Grimaldi & Karpicke, 2012 sourcePartly supported

  The Experiment 1 procedure says participants guessed a target for both related and unrelated cue–target pairs. The Experiment 2 Materials section says the lure for each cue was its most frequently produced response, and Experiment 2 describes a lure as related to its cue. These passages support cue-related guesses in both pair types, but do not establish that participants’ guesses generally were highly related to the cues. (LLM, conf=0.81)

  Experimental — accuracy not yet measured
• cites Huelser & Metcalfe, 2012 sourceSupported

For unstudied items, generating correct responses has often, though not invariably, been shown to elicit higher JOLs than reading.

“For unstudied items, generating correct responses has often, though not invariably, been shown to elicit higher JOLs than reading, suggesting that participants are aware of the benefits of generation (e.g., Begg, Vinski, Frankovich, & Holgate, 1991).”

• cites Begg, Vinski, Frankovich, & Holgate, 1991 sourceSupported

Participants' predictions about learning reflected their subjective experience during learning.

“Predictions reflected participants' subjective experience during learning.”

In four experiments, on a final test, generating errors followed by feedback produced significantly better memory for the correct definition or translation than reading or making incorrect choices.

“In a final test of all words, generating errors followed by feedback led to significantly better memory for the correct definition/translation than either reading or making incorrect choices, suggesting that the benefits of generation are not restricted to correctly generated items.”

In Experiment 1, when the initial response was correct but the final test response was incorrect, participants picked the original lure at a rate no different from chance (33.3%; t(17)=.78, p=.447).

“when the initial response was correct but the final test response was incorrect, participants picked the original lure from the study phase at a rate no different from the chance rate of 33.3%, t(17) = .78,p = .447.”

In Experiment 2A, Read and Choice conditions did not differ, t(29)=.54, p=.596, d=.05.

“while the Read and Choice conditions did not differ (t (29) = .54,p = .596,d = .05).”

In Experiment 2B, there was no difference between reading and choosing, t(23)=1.68, p=.106, d=.12.

“There was no difference between reading and choosing, t(23) = 1.68, p = .106,d = .12.”

In Experiment 2A, Choice JOLs for definitions guessed correctly at study were much higher than for items answered incorrectly at study, t(29) = 7.02, p < .001, d = 1.11.

“Firstly, Choice JOLs for definitions guessed correctly at study were very substantially higher than for items incorrect at study (Fig. 3A), t(29) = 7.02, p < .001,d = 1.11.”

In Experiment 2A, for items incorrect at study, there was no difference between reading and incorrect choosing, t(29)=.90, p=.373, d=.10.

“There was no difference between reading and incorrect choosing, t(29) = .90,p = .373,d = .10.”

In Experiment 2B, for items incorrect at study, there was no difference between reading and incorrect choosing, t(23)=.14, p=.892, d=.011.

“with no difference between reading and incorrect choosing, t(23) = .14,p = .892,d = .011.”

In Experiment 1, choosing did not lead to poorer final test performance than reading, even when analysis was confined to items incorrect at study.

“Contrary to our expectations, choosing did not lead to poorer performance than reading, even when the analysis was confined to just those items which were incorrect at study.”

• Also stated in: § Discussion · plain

  “Our hypothesis that incorrect choosing might lead to poorer final test performance than reading was not supported”

In Experiment 2B, after making an error at study, participants tended to repeat that same error at test rather than choose another lure seen at study.

“in Experiment 2B, when participants made an error, they tended to persist with the same error they had made at study rather than select one of the other lures.”

In Experiment 1, for items answered incorrectly at study, the difference between Generate and Choice fell short of significance, t(23)=1.91, p=.069, d=.36.

“the difference between Generate and Choice fell short of significance, (t(23) = 1.91, p = .069,d = .36)”

• Also stated in: § Discussion · hedged

  “there was a marginal benefit when only items incorrect at study were considered”

In Experiment 2B, final test performance differed between study methods, F(2,46)=6.05, p=.005, partial eta^2=.21.

“final test performance differed between study methods, F(2,46) = 6.05, p = .005,p 2 = .21”

In Experiment 1, correct choices at study occurred at chance level, with a mean of 53.3%, SD of 10.5, t(23)=1.6, p=.133.

“correct choices were at chance (M = 53.3%,SD = 10.5, t(23) = 1.6, p = .133)”

In Experiment 1, final test performance differed by study method when all items were considered, F(2,46)=7.62, p=.001, partial eta^2=.25.

“Final test performance differed by study method, F(2,46) = 7.62, p = .001,p 2 = .25”

In Experiment 2B, the difference between generating and choosing was close to significant, t(23)=1.84, p=.079, d=.18.

“The difference between generating and choosing was close to significant, t (23) = 1.84, p = .079,d = .18.”

In Experiment 2A, when an incorrect choice had been selected at study and the final test response was also wrong, the same response was selected at test at a rate numerically but not significantly higher than chance (33%; M=44.0, SD=37.8; t(23)=1.38, p=.180).

“When an incorrect choice had been selected at study, and the final test response was also wrong, the same response was selected at test at a rate numerically but not significantly higher than chance (33%, because there are 3 incorrect lures at test), M = 44.0,SD = 37.8, t(23) = 1.38, p = .180”

In Experiment 2A, Choice item JOLs were higher than Generate item JOLs, t(29) = 6.37, p < .001, d = .43.

“and Generate (t(29) = 6.37, p < .001,d = .43)JOLs”

In Experiment 2A, generating with feedback was superior to reading, t(29)=4.27, p<.001, d=.40.

“Generating with feedback was superior to reading (t(29) = 4.27, p < .001,d = .40)”

In Experiment 2B, for items incorrect at study, the advantage of generating errors over choosing incorrectly was significant, t(23)=2.77, p=.011, d=.30.

“The advantage of generating errors over choosing incorrectly was also significant in this analysis, t(23) = 2.77, p = .011,d = .30”

In Experiment 2B, for items incorrect at study, there was a difference between study methods, F(2,46)=6.81, p=.003, partial eta^2=.23.

“There was a difference between study methods, F(2,46) = 6.81, p = .003,p 2 = .23”

In Experiment 1, for items answered incorrectly at study, final test performance differed by study method, F(2,46)=3.33, p=.044, partial eta^2=.13.

“This analysis revealed a similar pattern, F (2,46) = 3.33, p = .044,p 2 = .13.”

In Experiment 1, Read and Choice conditions did not differ in final test performance when all items were considered, t(23)=.77, p=.447, d=.09.

“The Read and Choice conditions did not differ (t(23) = .77,p = .447,d = .09).”

In Experiment 2B, correct responses to Choice items at study were above chance, M=31.3%, SD=14.5, t(23)=2.21, p=.046.

“Correct responses to Choice items (M = 31.3%,SD = 14.5) were again above chance 2 (t(23) = 2.21, p = .046).”

In Experiment 1, correct generation of definitions at study was rare, with a mean of 6.5% and SD of 12.3.

“At study correct generations were few (M = 6.5%,SD = 12.3)”

Experiments 2A and 2B replicated the advantage of errorful generation over reading found in Experiment 1.

“Experiments 2A and 2B therefore replicated the benefit of errorful generation over reading observed in Experiment 1”

In Experiment 2A, aggregate JOLs differed by study method, F(1.62, 46.99) = 7.78, p = .002, partial eta squared = .21.

“JOLs differed by study method (F(1.62,46.99) = 7.78, p = .002,p 2 = .21).”

In Experiment 2B, there was some evidence that lures selected at study can interfere with selecting the correct answer at test.

“As in Experiment 1, there was some evidence (in Experiment 2B) that lures selected at study can interfere with selection of the correct answer at test.”

In Experiment 2A, aggregate Read and Generate JOLs did not differ, t(29) = .85, p = .400, d = .10.

“but there was no difference between Read and Generate (t(29) = .85,p = .400,d = .10).”

In Experiment 1, when all items were considered, Generate items were better remembered than Choice items, t(23)=2.80, p=.010, d=.42.

“and Choice items (t(23) = 2.80, p = .010,d = .42)”

• Also stated in: § Discussion · plain

  “Generation was also more beneficial than choosing when all items were considered”

In Experiment 1, when an incorrect response was made at final test, it was more likely to be the original lure if that lure had also been picked at study (M=73.4%) than if a correct answer had been given at study (M=35.3%; t(15)=3.44, p=.004).

“this response was significantly more likely to be the original lure when that lure had also been picked at study (M = 73.4% of incorrect responses following initial selection of a lure, SD = 30.2) than when a correct answer had been given at study (M = 35.3% of incorrect responses following an initially correct choice, SD= 33.0), t(15) = 3.44, p =.004.”

In Experiment 2A, aggregate Choice JOLs were higher than aggregate Read JOLs, t(29) = 2.58, p = .015, d = .49.

“Choice JOLs were again higher than both Read (t(29) = 2.58, p = .015,d = .49)”

In Experiment 2A, generating with feedback was superior to choosing with feedback, t(29)=3.62, p=.001, d=.33.

“and to choosing with feedback (t(29) = 3.62, p = .001,d = .33)”

In Experiment 2B, more participants remembered more Generate than Read items (14) than the reverse (3).

“Fourteen participants remembered more Generate than Read items, while only 3 showed the opposite pattern.”

In Experiment 2A, more participants remembered more Generate than Read items (19) than the reverse (2).

“Nineteen participants remembered more Generate than Read items, while only two showed the opposite pattern.”

In Experiment 1, for items answered incorrectly at study, Read and Choice did not differ, t(23)=.01, p=.990, d=.002.

“Again there was no significant difference between the Read and Choice conditions (t(23) = .01,p = .990,d = .002).”

Participants' metacognitive judgments of learning consistently failed to predict the benefit of errorful generation.

“We also found that participants, in their metacognitive judgments of learning, consistently failed to predict this benefit.”

In Experiment 2A, correct responses to Choice items at study were above chance, M=30.3%, SD=10.6, t(29)=2.76, p=.010.

“Correct responses to Choice items (M = 30.3%,SD = 10.6) were above chance 2 (t(29) = 2.76, p = .010).”

In Experiment 2B, when a Choice item was answered incorrectly at study and the final test response was also wrong, participants chose the same incorrect response at test at a rate significantly above the 33% chance level (M = 59.7, SD = 41.9, t(16) = 2.59, p = .020).

“When an incorrect response was given at study for an item in the Choice condition, and the final test response was also wrong, participants selected the same incorrect response at test at a rate significantly higher than the chance level of 33% (M = 59.7,SD = 41.9),t(16) = 2.59, p = .020”

In Experiment 2A, aggregate Choice JOLs were higher than aggregate Generate JOLs, t(29) = 3.49, p = .002, d = .58.

“and Generate JOLs (t(29) = 3.49, p = .002,d = .58)”

In Experiment 1, for items answered incorrectly at study, Generate scores were higher than Read scores, t(23)=2.40, p=.025, d=.56.

“Generate scores were higher than Read scores, t(23) = 2.40, p = .025,d = .56”

In Experiment 2A, for items incorrect at study, Generate had an advantage over Read, t(29)=4.28, p<.001, d=.39.

“as did the advantage for Generate over Read items, t(29) = 4.28, p < .001,d = .39.”

In Experiment 2B, generating produced better final test performance than reading, t(23)=3.36, p=.003, d=.28.

“Generating produced better final test performance than reading, t(23) = 3.36, p = .003,d = .28.”

In Experiment 1, more participants remembered more Generate than Read items (15) than the reverse (3).

“Fifteen participants remembered more Generate than Read items and only three showed the reverse pattern.”

In Experiment 2A, Read item JOLs were higher than Generate item JOLs, t(29) = 2.59, p = .015, d = .17.

“and Read JOLs were higher than Generate JOLs (t(29) = 2.59, p = .015,d = .17).”

In Experiment 2A, only 0.3% of Generate responses at study were correct.

“At study only 0.3% of Generate responses were correct.”

For the English version of the task, item JOLs differed across the three study conditions, F(2,58) = 20.73, p < .001, partial eta squared = .42.

“For the English version of the task, item JOLs differed for the three study conditions, F(2,58) = 20.73,p < .001(Figure 3A), p 2 = .42.”

In Experiment 2A, for items incorrect at study, there was a clear benefit of generating over choosing incorrect definitions, t(29)=3.91, p=.001, d=.46.

“in Experiment 2A there was a clear benefit of generating over choosing incorrect definitions, t(29) = 3.91, p = .001,d = .46.”

In Experiment 2B, only one Generate response at study was correct across all participants, M=.2%, SD=1.0.

“At study only one response given in the Generate condition was correct across all participants (M = .2%,SD = 1.0).”

Participants erroneously judged errorful generation to be a less effective encoding method than reading or incorrect choosing.

“Indeed they erroneously judged errorful generation to be a less effective encoding method than reading or incorrect choosing.”

In Experiment 1, when both initial and final responses were incorrect, participants selected the same incorrect answer at test as at study at a rate considerably higher than chance, t(15)=5.31, p<.001.

“when both the initial and final responses were incorrect, participants picked the same incorrect answer at test as they had done at study at a rate considerably higher than chance, t(15) = 5.31, p < .001”

In Experiment 2A, Choice item JOLs were higher than Read item JOLs, t(29) = 3.71, p = .001, d = .27.

“Choice JOLs were higher than both Read (t(29) = 3.71, p = .001,d = .27)”

In Experiment 2A, for items incorrect at study, the difference between study methods remained, F(2,58)=10.01, p<.001, partial eta^2=.26.

“For items which were incorrect at study, the difference between study methods remained, F(2,58) = 10.01,p < .001,p 2 = .26”

In Experiment 1, when all items were considered, Generate items were better remembered than Read items, t(23)=3.65, p=.001, d=.47.

“Generate items were better remembered than Read (t(23) = 3.65, p = .001,d = .47)”

• Also stated in: § Discussion · plain

  “Our first experiment revealed a benefit of generating followed by feedback over reading during the learning of unusual English words, even though generation produced many errors at study.”

In Experiment 3, when all items were considered, there was a main effect of Study Method on test accuracy, F(2,76) = 9.17, p < .001, partial eta squared = .19.

“When all items were considered, whether correct or incorrect at study, there was a main effect of Study Method, F(2,76) = 9.17, p <.001, p 2 = .19”

In Experiment 3, when all items were considered, there was no interaction between Study Method and Group, F(2,76) = 1.57, p = .216.

“and no interaction, F(2,76) = 1.57, p = .216.”

In the EP group of Experiment 3, JOLs for Generate items were lower than Read JOLs, t(15) = 4.25, p = .001, d = .59.

“but were lower than Read JOLs, t(15) = 4.25, p = .001,d = .59”

In Experiment 2B, there was no difference between JOLs for Read and Choice items, t(23) = 1.38, p = .181, d = .12.

“but here there was no difference between JOLs for Read and Choice items, t(23) = 1.38, p = .181,d = .12.”

Participants' guesses were no more related to the correct definitions than the correct definitions were to each other.

“Participants' guesses were no more related to the correct definitions than the correct definitions were to each other.”

In Experiment 3, the percentage of items correct at study that remained correct at test was not significantly greater than the percentage of items incorrect at study that became correct at test.

“In fact, the percentage of items correct at study which persisted to be correct at test (M = 77.81,SD = 26.18)was not significantly greater than the percentage of items incorrect at study which were converted to being correct at test (M = 73.05,SD = 22.99), t(39) = 1.06, p = .296,d = .19.”

In Experiment 3, for items incorrect at study, there was no main effect of Group, F(1,38) = .19, p = .662.

“no main effect of Group, F(1,38) = .19,p = .662”

Participants' JOLs replicated the previously observed pattern: item JOLs were significantly lower in the Generate condition than in the Read or Choice conditions.

“Participants' JOLs also replicated the pattern observed previously, in that item JOLs were significantly lower in the Generate condition than in either the Read or Choice conditions.”

Experiment 3 replicated the benefit of generating errors over reading and incorrect choosing observed in previous experiments.

“Experiment 3 replicated the benefit of generating errors over both reading and incorrect choosing observed in our previous experiments”

In the EP group of Experiment 3, JOLs for Read items were higher than those for Choice items incorrect at study, t(15) = 2.66, p = .018, d = .40.

“and JOLs for Read items were higher than those for Choice items incorrect at study, t (15) = 2.66, p = .018,d = .40.”

In Experiment 3, the response a participant had generated at study was selected at test 13.2% of the time (SD = 21.5), significantly lower than the chance level of 25%, t(39) = 3.49, p < .001.

“The response a participant had generated at study was only selected at test 13.2% of the time (SD = 21.5), which was significantly lower than the chance level of 25% (since there were 4 incorrect options), t(39) = 3.49, p < .001”

In Experiment 3, JOLs for Choice items correct at study were significantly higher than JOLs for Read items.

“Replicating the findings of Experiment 2A and 2B, JOLs for Choice items correct at study were significantly higher than JOLs for Read items, t(39) = 7.83, p < .001.”

In Experiment 2B, JOLs for Choice items guessed correctly at study (M = 47.4, SD = 21.8) were much higher than for items guessed incorrectly (M = 31.6, SD = 15.7), t(23) = 5.52, p < .001, d = .83.

“First, JOLs for Choice items guessed correctly at study (M = 47.4,SD = 21.8) were much higher than for items guessed incorrectly (M = 31.6,SD = 15.7),t(23) = 5.52, p < .001,d = .83(Figure 3B).”

In the first three experiments, incorrect choosing did not lead to poorer performance than reading.

“but in fact our experiments revealed no detriment of incorrect choosing by comparison with reading.”

In Experiment 3, the percentage of correct answers for Choice items (M = 27.9, SD = 10.5) did not differ from the chance level of 25%, t(39) = 1.73, p = .091.

“and the percentage of correct answers given for Choice items (M = 27.9,SD = 10.5) did not differ from the chance level of 25% (t(39) = 1.73, p = .091).”

In Experiment 2B, there was a main effect of study method on aggregate JOLs, F(2,46) = 5.09, p = .010, partial eta squared = .18.

“There was a main effect of study method, F(2,46) = 5.09, p = .010,p 2 = .18.”

In Experiment 3, the benefit of generating errors persisted even though participants could select at test an incorrect guess they had generated at study.

“confirmed that this benefit persisted even though participants had the opportunity to select, at test, an incorrect guess they had generated at study”

In Experiment 2B, JOLs for Choice items guessed correctly at study were significantly higher than JOLs for Read items, t(23) = 4.36, p < .001.

“Second, JOLs for Choice items guessed correctly at study were significantly higher than JOLs for Read items, t(23) = 4.36, p < .001.”

In Experiment 3, an incorrectly chosen response was selected at test 24.0% of the time (SD = 22.6), which was no different from chance, t(39) = .27, p = .786.

“whereas an incorrectly chosen response was selected 24.0% of the time (SD = 22.6), which was no different from chance, t(39) = .27,p = .786.”

In Experiment 2B, participants gave lower JOLs to Generate items than to Choice items, t(23) = 4.31, p < .001, d = .50.

“or Choice (t(23) = 4.31, p < .001,d = .50)items”

In Experiment 3, when all items were considered, there was no effect of Group, F(1,38) = .27, p = .605.

“no effect of Group, F(1,38) = .27,p = .605”

In Experiment 2B, aggregate Choice JOLs were higher than aggregate Generate JOLs, t(23) = 3.21, p = .004, d = .67.

“Replicating the findings of the English version of the task, Choice JOLs were higher than Generate JOLs, t(23) = 3.21, p = .004,d = .67”

In Experiment 2B, aggregate JOLs followed a similar pattern to item JOLs.

“Aggregate JOLs followed a similar pattern to the item JOLs (Figure 3B).”

In Experiment 3, for items incorrect at study, generating produced better performance than reading, t(39) = 5.22, p < .001, d = .56.

“Generating produced better performance than reading (t(39) = 5.22, p < .001,d = .56)”

For the foreign language version, item JOLs differed across the three study conditions, F(2,46) = 12.60, p < .001, partial eta squared = .35.

“For the foreign language version of the task, item JOLs also differed for the three study conditions, F(2,46) = 12.60, p < .001,p 2 = .35(Figure 3B).”

In Experiment 2A, JOLs for Choice items incorrect at study were indistinguishable from JOLs for Read items, t(29) = 1.74, p = .093.

“and from JOLs for Read items, t(29) = 1.74, p = .093.”

In Experiment 2A, JOLs for Choice items correct at study were higher than Read JOLs, t(29) = 6.62, p < .001.

“Secondly, JOLs for Choice items correct at study were higher than Read JOLs, t(29) = 6.62, p < .001”

At study in Experiment 3, just one participant gave the correct response to just one Generate item.

“At study, just one participant gave the correct response to just one Generate item”

In Experiment 2B, participants gave lower JOLs to Generate items than to Read items, t(23) = 3.69, p = .001, d = .39.

“Once again, participants gave lower JOLs to Generate items than to either Read (t(23) = 3.69, p = .001,d = .39)”

In Experiment 3, there was an interaction between Study method and Group on JOLs, F(1.62, 61.54) = 3.78, p = .037, partial eta squared = .09.

“but again there was an interaction between Study method and Group, F(1.62, 61.54) = 3.78, p = .037,p 2 = .09.”

JOLs for Choice items incorrect at study were higher than JOLs for Generate items.

“JOLs for Choice items incorrect at study were higher than JOLs for Generate items.”

This is in line with the findings of Experiment 2A, where participants selected their own incorrect choices with no greater frequency than other incorrect options which had appeared as lures at study.

“This is in line with the findings of Experiment 2A, where participants selected their own incorrect choices with no greater frequency than other incorrect options which had appeared as lures at study.”

In Experiment 3, for items incorrect at study, incorrect choosing was no better than reading, t(39) = 1.36, p = .181, d = .17.

“but incorrect choosing was no better than reading, t(39) = 1.36, p = .181,d = .17”

As in Experiments 2A and 2B, participants gave much higher JOLs to Choice items they guessed correctly at study than to Choice items they guessed incorrectly.

“As in Experiments 2A and 2B, participants gave much higher JOLs to Choice items they guessed correctly at study than to Choice items they guessed incorrectly.”

In Experiment 2B, aggregate Read JOLs were higher than aggregate Generate JOLs, t(23) = 2.73, p = .012, d = .57.

“This time Read JOLs were also higher than Generate JOLs, t(23) = 2.73, p = .012,d = .57.”

Experiment 2 revealed a clear benefit of generating errors over incorrect choosing, both for rare English words and for foreign language words.

“In addition to the benefit of generating over reading, Experiment 2 also revealed a clear benefit of generating errors over incorrect choosing, both for rare English words and for foreign language words.”

In the SP group of Experiment 3, there was no difference between Read and Choice JOLs, t(23) = 1.34, p = .193, d = .24.

“with no difference between Read and Choice, t(23) = 1.34, p = .193,d = .24.”

In Experiment 3, choosing led to higher test scores than reading, t(39) = 2.49, p = .017, d = .28.

“Choosing also led to higher test scores than reading, t(39) = 2.49, p = .017d = .28”

The mean similarity of generated items to correct definitions was .096 (SD = .047).

“The mean similarity of the generated items to the correct definitions was .096(SD = .047).”

In Experiment 3, when calculated as a percentage of just the incorrect responses at test, the comparison between persisting with chosen versus generated definitions remained significant, t(39) = 2.21, p = .033.

“When we calculated the number of times a generated or chosen item was selected at test as a percentage of just the incorrect responses at test, the comparison remained significant, t(39) = 2.21, p = .033.”

In Experiment 3, there was no difference between choosing and generating, t(39) = 1.75, p = .089, d = .24.

“with no difference between choosing and generating, t(39) = 1.75, p = .089,d = .24.”

In Experiment 3 Choice condition, participants spent significantly longer studying items they got wrong at study than items they got right.

“In the Choice condition participants spent significantly longer studying items which they had got wrong at study (M = 5.75 sec, SD = 3.81) than items they got right (M = 4.44, SD = 4.11), t (23) = 3.64, p < .001,d = .33.”

The benefit of errorful generation over reading was observed even when participants regulated their own study time.

“This effect was observed even when participants regulated their own study time.”

In Experiment 3, aggregate JOLs showed a main effect of Study method.

“For the aggregate JOLs there was a main effect of Study method, F(2,76) = 4.10, p = .020,p 2 = .10(Figure 3C).”

In the EP group of Experiment 3, JOLs for Generate items were no different from JOLs for incorrect choices, t(15) = .95, p = .355, d = .17.

“In the EP group, JOLs for Generate items were no different from JOLs for incorrect choices, t(15) = .95,p = .355,d = .17”

In Experiment 3, for items incorrect at study, there was a main effect of Study Method, F(1.70, 64.44) = 7.35, p = .002, partial eta squared = .19.

“There was a main effect of Study Method, F(1.70, 64.44) = 7.35, p = .002,p 2 = .19”

In Experiment 2B, inspection of Choice JOLs in relation to study performance revealed the same pattern as in Experiment 2A.

“Again, inspection of Choice JOLs in relation to study performance is illuminating and reveals the same pattern of results as in Experiment 2A.”

In Experiment 3, generating produced better final test performance than reading, t(39) = 5.22, p < .001, d = .56.

“Generating produced better final test performance than reading, t(39) = 5.22, p < .001,d = .56.”

In the SP group of Experiment 3, Generate items were given lower JOLs than Read items, t(23) = 2.77, p = .011, d = .29.

“and Read JOLs, t(23) = 2.77, p = .011,d = .29”

At final test, participants were more likely to select the same incorrect choice they had made at study than an incorrect response they had generated.

“Indeed, at final test, participants were more likely to select the same incorrect choice they had made at study than an incorrect response they had generated.”

In Experiment 3, there was no effect of Group on JOLs, F(1,38) = .53, p = .469, partial eta squared = .01.

“There was no effect of Group, F(1,38) = .53,p = .469,p 2 = .01”

In Experiment 3 SP group, study time differed significantly across conditions.

“There was a significant difference between conditions in study time, F(2, 46) = 35.44,p < .001,p 2 = .60.”

Memory for Choice items incorrect at study was poorer than for Generate items.

“Memory for Choice items incorrect at study was, however, poorer than for Generate items.”

In the SP group of Experiment 3, Generate items were given lower JOLs than Choice items incorrect at study, t(23) = 3.04, p = .006, d = .49.

“In the SP group, Generate items were given lower JOLs than both Choice items incorrect at study, t(23) = 3.04, p = .006,d = .49”

In Experiment 3, participants spent longer studying Generate items incorrect at study than Choice items incorrect at study.

“Interestingly, although participants spent longer studying Choice items which were incorrect at study than they spent on correct items, they spent even longer on Generate items incorrect at study, t(23) = 3.13, p = .005.”

In Experiment 3, for items incorrect at study, there was no interaction, F(1.70, 64.44) = 1.53, p = .224.

“and no interaction, F(1.70, 64.44) = 1.53, p = .224.”

Generating errors followed by feedback was more beneficial than incorrect choice in all experiments except the first, where the difference fell just short of significance.

“Generating errors followed by feedback was also more beneficial than incorrect choice in all experiments except the first, in which the difference fell just short of significance.”

In Experiment 2B, there was no relationship between JOLs and test accuracy.

“In Experiment 2B there was no relationship between JOLs and test accuracy (Fig 4B).”

In Experiment 2A, JOLs for Choice items incorrect at study were indistinguishable from JOLs for Generate items incorrect at study (M = 31.3, SD = 16.7), t(29) = .77, p = .450.

“thirdly, JOLs for Choice items incorrect at study were indistinguishable from JOLs for Generate items incorrect at study (M = 31.3,SD = 16.7),t(29) = .77,p = .450”

In Experiment 3, participants were more inclined to persist with incorrectly chosen than with incorrectly generated definitions, t(39) = 2.49, p = .017.

“Participants were more inclined to persist with incorrectly chosen than with incorrectly generated definitions, t(39) = 2.49, p = .017.”

In Experiment 3, for items incorrect at study, incorrect choosing was less beneficial than generating, t(39) = 2.08, p = .044, d = .32.

“and was less beneficial than generating, t(39) = 2.08, p = .044,d = .32”

There was an interaction between Study Method and Group on JOLs.

“but there was an interaction between Study Method and Group, F(1.69, 64.22) = 3.76, p = .035,p 2 = .09.”

For items incorrect at study, Generate JOLs were lower than Read JOLs.

“These JOLs were lower than both Read JOLs, t(39) = 4.75, p < .001,d = .40”

The errorful generation benefit was observed whether or not the participant's initial guess appeared as one of the final test options.

“The errorful generation benefit was observed whether or not the participant's initial guess appeared as one of the final test options.”

Participants consistently gave lower JOLs to Generate items than to Read or Choice items, even when in Experiment 3 they could study correct answers as long as they liked.

“Participants consistently gave lower JOLs to Generate items than to Read or Choice items, even when, in Experiment 3, they had the opportunity to study correct answers for as long as they liked.”

JOLs for Choice items were heavily influenced by whether the correct definition had been chosen at study.

“Furthermore, JOLs for Choice items were heavily influenced by whether or not the correct definition had been chosen at study.”

Experiment 3 showed participants were more likely to reproduce at final test an incorrect guess made at study in the Choice condition than one made in the Generate condition.

“Experiment 3 showed that participants were more likely to reproduce, at final test, an incorrect guess made at study in the Choice condition than one made in the Generate condition.”

In Experiment 3, there was no interaction between Study method and Group on test accuracy.

“and no interaction, F(2,208) = 2.32, p = .101.”

For the SP group, Choice JOLs were higher than Read JOLs.

“and higher than Read JOLs, t(23) = 3.92, p = .001,d = .64”

In Experiment 3, there was no effect of Group on test accuracy.

“no effect of Group, F(1,104) = 2.72, p = .102”

In all experiments and within each condition, JOLs were numerically higher for items later answered correctly than for items answered incorrectly.

“Figure 4 shows that, in all experiments and within each condition, JOLs for items subsequently answered correctly at test were numerically higher than for items answered incorrectly.”

In Experiment 2B, the mean gamma correlation between JOLs and test accuracy did not differ significantly from zero in any condition.

“In Experiment 2B, the mean gamma correlation between JOLs and test accuracy did not differ significantly from zero in any condition”

In Experiment 3, generating led to better recall than reading.

“Generating (M = 75.38,SD =17.6) led to better recall than reading (M = 67.48,SD = 20.66),t(105) = 4.18, p < .001,d = .41.”

There was no main effect of Group on JOLs.

“There was no main effect of Group,F(1,38) .48,p = .491,p 2 = .013”

Choice JOLs were higher than Read JOLs.

“Choice JOLs were higher than both Generate, t(39) = 6.95, p < .001,d = .80,and Read JOLs”

For the EP group, Read JOLs were higher than Generate JOLs.

“For the EP group, Read JOLs were again higher than Generate JOLs, t(15) = 4.25, p = .001,d = .59”

The paper's findings demonstrate that generating responses followed by feedback is helpful to memory even when many errors are generated, compared with errorless studying without generation.

“Our findings demonstrate that generating responses followed by feedback is helpful to memory even when many errors are generated, compared with errorless studying without generation.”

For items incorrect at study, Generate items received the lowest JOLs.

“Again, the lowest JOLs were given to Generate items.”

In Experiment 2A, the mean gamma correlation between JOLs and final test accuracy was significantly above zero in all conditions.

“In Experiment 2A the mean gamma correlation was significantly higher than zero in all conditions (t(24) = 4.43, p < .001for Read; t(19) = 3.50, p = .002for Generate; and t(24) = 2.96, p = .007for Choice)”

Participants were very good at rejecting their own incorrect generations, choosing these at test at a rate significantly below chance.

“On the other hand they were very good at rejecting their own incorrect generations, choosing these at test at a rate significantly below chance.”

In Experiment 3, choosing was better than reading.

“Choosing (M = 72.79,SD = 21.31) was better than reading, t(105) = 2.45, p = .015,d = .25”

After removing four Experiment 2A participants and two Experiment 2B participants with especially high Choice study scores, Choice study performance was no longer significantly above chance in either experiment.

“Without these participants, Choice performance at study was no longer significantly greater than chance in either experiment.”

The accuracy of monitoring the relative recallability of items (resolution) did not differ between conditions.

“Resolution did not differ across conditions.”

Participants gave the lowest JOLs to Generate items even when they could choose how long to study correct definitions.

“Participants gave the lowest JOLs to Generate items even when they could choose how long to study correct definitions”

For the EP group, Choice JOLs were higher than Generate JOLs.

“and Choice JOLs were higher than Generate JOLs, t(15) = 2.72, p = .016,d = .59”

In Experiment 3, there was a main effect of Study method on test accuracy.

“There was a main effect of Study method, F(2, 208) = 7.36, p = .001,p 2 = .066”

A 3 (Study Method) x 2 (Group) ANOVA found a main effect of study method on item JOLs.

“A 3 (Study Method) * 2 (Group) ANOVA revealed a main effect of study method for the item JOLs.”

For items incorrect at study, there was a main effect of Study Method on JOLs.

“There was a main effect of Study Method, F(1.62, 61.54) = 6.84, p = .004,p 2 = .15(Huynh-Feldt correction).”

Participants' JOLs were heavily influenced by success or failure at study.

“In fact their JOLs were heavily influenced by success or failure at study.”

• AssumptionRests on an assumption: JOLs measure participants' beliefs or predictions about their future memory performance.

For the SP group, Choice JOLs were higher than Generate JOLs.

“For the SP group, Choice JOLs were higher than Generate JOLs, t(23) = 7.71, p < .001,d = .95”

Generating errors was beneficial to memory even when there was no pre-existing association that could be reinforced by corrective feedback.

“Generating errors, then, was beneficial to memory even when there was no preexisting association which could be reinforced by corrective feedback.”

• AssumptionRests on an assumption: Participants had no pre-existing associations with the novel words or foreign vocabulary before the experiment.

Read JOLs were higher than Generate JOLs.

“and Read JOLs were higher than Generate JOLs, t(39) = 4.75, p < .001,d = .40.”

For the EP group, Read and Choice JOLs did not differ.

“but there was no difference between Read and Choice JOLs, t(15) = .39,p = .701,d = .07.”

In Experiment 3, the mean gamma correlation between JOLs and test accuracy was significantly above zero in all conditions.

“In Experiment 3, the mean gamma correlation between JOLs and test accuracy was significantly higher than zero in all conditions”

In Experiment 3, choosing and generating did not differ.

“with no difference between choosing and generating, t(105) = 1.20, p = .234,d = .13.”

Resolution did not differ between conditions.

“There was no difference in resolution (i.e., accuracy at monitoring the relative recallability of items) between conditions, F (2,36) = .112,p = .894”

In all experiments, when participants made incorrect choices at study they were more likely to answer those items correctly at test than to make any type of error.

“However, in all experiments, when participants made incorrect choices at study they were more likely to answer those items correctly at test than to make any type of error.”

Confidence ratings did not differ between conditions.

“However, there was no difference between conditions”

Choice JOLs were higher than Generate JOLs.

“Choice JOLs were higher than both Generate, t(39) = 6.95, p < .001,d = .80,and Read JOLs”

For the SP group, Read JOLs were higher than Generate JOLs.

“which were higher than Generate JOLs, t(23) = 2.77, p = .011,d = .28”

For items incorrect at study, Generate JOLs were lower than Choice JOLs.

“and JOLs for Choice items incorrect at study, t(39) = 3.05, p =.004, d = .38”

The testing effect occurs even without corrective feedback.

“The testing effect is observed even when no corrective feedback is given (Allen et al., 1969;Carpenter & DeLosh, 2005, 2006;Kuo & Hirshman, 1996), so it cannot be attributed solely to more efficient processing of feedback in the test condition.”

• cites Allen et al., 1969 sourceNot checked — no full text of the source
• cites Carpenter & DeLosh, 2005, 2006Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Kuo & Hirshman, 1996 sourceNot checked — no full text of the source

The memory benefit of generation is not limited to items that were correctly generated.

“suggesting that the benefits of generation are not restricted to correctly generated items.”

Generating an item in response to a cue improves later memory for that item compared with reading it, the generation effect.

“Generating items in response to a cue (e.g.,”

• cites Slamecka & Graf, 1978 sourceNot checked — no full text of the source

Slamecka and Fevreiski found that generation led to better later memory than reading even when many errors were made during study.

“Slamecka and Fevreiski observed that, even when many errors were made at study, generation led to better memory on a later test than reading.”

• cites Slamecka & Fevreiski, 1983Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

A pre-existing semantic association between cue and target is not required for the errorful generation benefit to occur.

“A pre-existing semantic association between cue and target is not necessary for the benefit to occur.”

Retrieving items from memory improves later memory for those items compared with re-reading them, a phenomenon known as the testing effect.

“A robust and highly replicated finding from both laboratory and classroom studies is that the very act of retrieving items from memory enhances memory for the tested items, the”

• cites Roediger & Karpicke, 2006aNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Errorful generation may strongly promote encoding of corrective feedback even when the material to be learned is novel.

“Even when information to be learned is novel, errorful generation may play a powerful role in potentiating encoding of corrective feedback.”

Berlyne and Normore (1972) found that inducing and then satisfying curiosity by presenting a blurred picture before a clear picture of the same object led to better memory for the objects than presenting the clear picture for twice as long.

“Berlyne and Normore (1972) found that inducing, then satisfying, curiosity by presenting a blurred picture immediately before a clear picture of the same object, led to better memory for the objects than presenting the clear picture for twice as long.”

• cites Berlyne and Normore (1972) sourceNot checked — no full text of the source

In errorful generation with novel vocabulary, the incorrect guess is unlikely to be related to either the cue or the target.

“Neither the cue nor the task instructions constrain guessing, so the incorrectly generated item is unlikely to be related to either cue or target.”

In Experiment 1, items incorrectly chosen at study were strong enough to lead to perseverative errors at test when the alternatives were new lures and the correct target.

“items incorrectly chosen at study were strong enough to lead to perseverative errors at test when the alternatives were new lures (and the correct target)”

Errors made on the initial test can interfere with accurate retrieval at final test.

“suggesting that errors made on the initial test can interfere with accurate retrieval at final test.”

Hays, Kornell, and Bjork (2013) found evidence of interference from incorrect answers generated at study.

“Hays, Kornell, and Bjork (2013) also found evidence of interference from incorrect answers generated at study.”

• cites Hays, Kornell, and Bjork (2013) sourceNot checked — no full text of the source

Across the experiments, generating errors led to better memory than reading.

“we observed a benefit of generating errors over reading in all three experiments”

Across Experiments 2A, 2B, and 3, generating errors led to better memory than incorrect choosing.

“and a benefit of generating errors over incorrect choosing in Experiments 2A, 2B, and 3.”

Howard and Kahana showed that people tend to recall items semantically related to the item they just recalled.

“Howard and Kahana (2002) showed that participants tend to recall items which are semantically related to the item just recalled.”

• cites Howard & Kahana, 2002 sourceNot checked — no full text of the source

M. J. Kang et al. (2009) found that the higher participants rated their curiosity about a question answered incorrectly, the more likely they were to recall the correct answer on a surprise test two weeks later.

“M. J. Kang, Hsu, Krajbich, Loewenstein, McClure, Wang, and Camerer (2009) found that the higher participants rated their curiosity about a question answered incorrectly, the more likely they were to recall the correct answer on a surprise test two weeks later.”

• cites M. J. Kang, Hsu, Krajbich, Loewenstein, McClure, Wang, and Camerer (2009) sourceNot checked — no full text of the source

Discrepancies between what is expected and what occurs drive learning.

“much research shows that discrepancies between what is expected and what occurs drive learning (e.g., Rescorla & Wagner, 1972).”

• cites Rescorla & Wagner, 1972Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In Experiment 2A, items incorrectly chosen at study were not strong enough to lead to significantly more perseverative errors when the alternative test items were familiar lures and the correct target.

“but not strong enough to lead to such errors when the alternative test items were familiar lures (and the correct target).”

When material has been previously studied, taking an MCQ test followed by feedback yields comparable final test performance to restudying.

“When material has been previously studied, taking an MCQ test followed by feedback has been found to yield comparable final test performance to restudying (e.g.Butler & Roediger, 2007; S. H. K. Kang, McDermott, & Roediger, 2007;McDaniel, Anderson, Derbish, & Morisette, 2007).”

• cites Butler & Roediger, 2007 sourceNot checked — no full text of the source
• cites S. H. K. Kang, McDermott, & Roediger, 2007 sourceNot checked — no full text of the source
• cites McDaniel, Anderson, Derbish, & Morisette, 2007 sourceNot checked — no full text of the source

In Experiment 1, most definitions were unknown to participants before the experiment.

“confirming that most definitions were unknown to participants pre-experimentally”

The strength of incorrect choices at study to produce perseverative errors at test depends on whether the alternative test lures are new or familiar.

“Put differently, items incorrectly chosen at study were strong enough to lead to perseverative errors at test when the alternatives were new lures (and the correct target), but not strong enough to lead to such errors when the alternative test items were familiar lures (and the correct target).”

The higher JOLs given to Read than Generate items in both experiments are consistent with the hypothesis that Generate items are experienced as more difficult than Read items.

“The higher JOLs given to Read than Generate items in both these experiments are consistent with our hypothesis that Generate items are experienced as more difficult than Read items.”

In Experiment 2A, the pattern of JOLs suggests that Choice JOLs were largely driven by the fortuitous making of a correct choice at study.

“These findings suggest that Choice JOLs were largely driven by the fortuitous making of a correct choice at study.”

In Experiment 2A, together with higher Read than Generate JOLs, the data suggest that fluency of processing at study, influenced by the outcome of the preceding event, was one factor influencing JOLs.

“Together with the higher JOLs for Read than Generate items, this suggest that one factor influencing JOLs was fluency of processing at study, which was itself influenced by the outcome of the preceding event.”

This may lead to greater attention being paid to the corrective feedback for Generate items.

“This may lead to greater attention being paid to the corrective feedback for Generate items.”

The influence of guess accuracy on study time in the Choice condition supports the proposal that participants' perception of item difficulty is affected by the outcome of an event preceding study.

“The fact that study time, like JOLs, was influenced by the accuracy of the participant's guess at study in the Choice condition, adds further support to our proposal that participants' perception of the difficulty of remembering an item is affected by the outcome of an event preceding study.”

For Generate items, participants seemed able to recognise and reject their own incorrect study responses, selecting them less often than other incorrect options at test.

“For Generate items, on the other hand, participants seemed to be able to recognise and reject their own incorrect study responses, selecting them less often than the other incorrect options at test.”

This same fluency may mean that Choice items are processed less deeply, leading to poorer subsequent memory.

“However, this same fluency may also mean that these items are processed less deeply, leading to poorer subsequent memory.”

These findings are inconsistent with other versions of the semantic relatedness hypothesis discussed in the introduction (e.g., Huelser & Metcalfe, Hays et al., 2013).

“These findings are also inconsistent with the other versions of the semantic relatedness hypothesis that we discussed in the introduction (e.g., Huelser & Metcalfe, Hays et al., 2013)”

• cites Huelser & MetcalfeNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Hays et al., 2013 sourceNot checked — no full text of the source

• AssumptionRests on an assumption: Participants in Experiment 2B had no prior knowledge of Euskara.
• AssumptionRests on an assumption: In Experiment 2A, the obscure English words were likely to be largely unknown to participants before the experiment.

The illusion is exaggerated in the Choice condition when the participant happens to select the correct response.

“The illusion is exaggerated in the Choice condition when the participant happens to select the correct response.”

The amount of study time SP participants allocated to correct definitions reflected the same misconception.

“The amount of time participants in the SP group allocated to studying correct definitions also reflected this misconception.”

• AssumptionRests on an assumption: JOLs measure participants' beliefs or predictions about their future memory performance.
• AssumptionRests on an assumption: Study time reflects the amount of effort or attention participants devoted to encoding.

A pre-existing relationship between cue and target is not necessary for the errorful generation benefit to be observed.

“and show that a pre-existing relationship between cue and target is not necessary for the errorful generation benefit to be observed.”

• AssumptionRests on an assumption: Participants in Experiment 2B had no prior knowledge of Euskara.
• AssumptionRests on an assumption: In Experiment 2A, the obscure English words were likely to be largely unknown to participants before the experiment.

The low JOLs for Generate items in Experiment 2 did not stem from a perception that there was insufficient time to process correct feedback, because the pattern held when participants could study correct definitions as long as they liked.

“This was true even when participants were allowed to study correct definitions for as long as they liked, suggesting that the low JOLs for Generate items observed in Experiment 2 did not stem from a perception that there was insufficient time to process correct feedback.”

The pattern of JOLs is suggestive of participants using their own performance at study as a basis for predicting future memory performance.

“a pattern which is suggestive of participants using their own performance at study as a basis for predicting their future memory performance.”

In Experiment 3, despite the availability at test of definitions participants had generated themselves, generating errors followed by studying corrective feedback led to better final test performance than either reading or incorrect choosing.

“Despite the availability, at test, of definitions participants had generated themselves, generating errors followed by studying corrective feedback led to better final test performance than either reading or incorrect choosing.”

Participants' JOLs seemed to reflect their subjective experience at study.

“What we can conclude from both these experiments is that participants' JOLs seemed to reflect their subjective experience at study.”

High Choice JOLs were largely driven by high JOLs for items guessed correctly at study, suggesting participants used their own study performance as a basis for predicting future test performance.

“These high Choice JOLs were largely driven by high JOLs for items guessed correctly at study, suggesting that participants used their own performance at study as a basis for predicting their future test performance.”

• AssumptionRests on an assumption: JOLs measure participants' beliefs or predictions about their future memory performance.

Where a generated guess is related to the cue but not the target, strengthening of the interfering cue-guess association may cancel out the benefit generation confers on processing of feedback.

“Where the generated guess is related to the cue but not to the target, the strengthening of this interfering cue-guess association may cancel out the benefit generation confers on the processing of feedback.”

Participants did not use a different basis for JOLs in the Read and Generate conditions.

“suggesting that participants were not using a different basis for their JOLs in the Read and the Generate conditions.”

The fact that item JOLs were lower for Generate items than for other methods even when participants controlled study time suggests that the effort of generating a response followed by an error led to a perception that those items were more difficult to learn.

“The fact that these item JOLs were also lower for Generate items than for the other study methods even when participants could control how long they studied feedback, suggests that the effort involved in coming up with a response, followed by the making of an error, led to a perception that these items were more difficult to learn than items in the other conditions.”

Generation can benefit memory even when it yields many errors, by making the encoding of subsequent feedback more effective than when it is not preceded by a generation attempt.

“Our findings suggest that generation can benefit memory even when it yields many errors, by making the encoding of subsequent feedback more effective than when it is not preceded by a generation attempt.”

Item JOLs may give a measure of participants' perception of the ease or difficulty of learning for each individual item at the moment of study.

“item JOLs may give us a measure of participants' perception of the ease or difficulty of learning for each individual item at the moment of study”

• AssumptionRests on an assumption: JOLs measure participants' beliefs or predictions about their future memory performance.

Generating a response is likely to activate many concepts which, even if unrelated to the correct answer, could create a distinctive context for learning the answer and serve as retrieval cues on a later recall attempt.

“However, generating a response is likely to activate many concepts which could, even if unrelated to the correct answer, create a distinctive context for the learning of the answer and serve as retrieval cues on a subsequent recall attempt.”

By creating a situation where participants make both correct and incorrect guesses and collecting item-by-item JOLs, the paper showed that participants' perception of learning correct answers is strongly affected by whether they guessed correctly at study.

“By creating a situation in which participants make both correct and incorrect guesses, and by collecting item by item JOLs, we have been able to show that participants' perception of their learning of correct answers is strongly affected by whether or not they happened to guess correctly at study.”

• AssumptionRests on an assumption: JOLs measure participants' beliefs or predictions about their future memory performance.

In Experiment 2A, JOLs had some ability to predict test performance.

“showing some ability for JOLs to predict test performance.”

The low JOLs for Generate items did not arise from a perception that there was too little time to process corrective feedback in the Generate condition.

“suggesting that their low JOLs did not stem from a perception of having insufficient time to process the corrective feedback in the Generate condition.”

The act of generation per se, regardless of what is generated, makes the encoding of subsequent feedback more effective.

“Instead, the act of generation per se, regardless of what is generated, makes the encoding of subsequent feedback more effective.”

The benefit of errorful generation cannot be solely due to strengthening, through testing, of items related to the cue and target.

“The benefit of errorful generation cannot, therefore, be solely due to a strengthening, through testing, of items related to the cue and the target.”

Resolution was generally poor with the Experiment 2B version of the task.

“Resolution was generally poor with this version of the task.”

The paper's proposal is that participants focused more attention on correct feedback in the Generate condition than in the Read or Choice conditions.

“Our proposal is that participants focused more attention on correct feedback in the Generate condition than in the Read or Choice conditions.”

The errorful generation benefit does not depend on strengthening of pre-existing associations that enhance memory as a result of also being associated to the target.

“Thus we were able to show that the errorful generation benefit does not depend on a strengthening of pre-existing associations that enhance memory as a result of also being associated to the target.”

• AssumptionRests on an assumption: Participants had no pre-existing associations with the novel words or foreign vocabulary before the experiment.

The paper's design allowed observation of a benefit to encoding of feedback uncontaminated by an opposite interfering effect of generation on the generated item itself.

“Our design allowed us to observe a benefit to the encoding of feedback which was uncontaminated by an opposite, interfering effect of generation on the generated item itself.”

In the Choice condition, options available at study constitute the search set for the item, and these options interfered with correct recognition at test more than generated guesses did.

“Here, far from facilitating encoding of corrective feedback, options in the search set interfered with correct recognition at test to a greater extent than generated guesses did.”

The active process of generating a response, rather than merely selecting one, facilitated encoding of corrective feedback even when the generated response was incorrect.

“This suggests that there was something about the active process of generating a response, rather than merely selecting one, which facilitated encoding of corrective feedback, even when the generated response was incorrect.”

Participants were not more accurate in one condition than another.

“indicating that participants were no more accurate in one condition than in another.”