Exploring How Generating Metaphor Via Insight Versus Analysis Affects Metaphor Quality and Learning Outcomes

Abstract

Metaphor generation is both a creative act and a means of learning. When learning a new concept, people often create a metaphor to connect the new concept to existing knowledge. Does the manner in which people generate a metaphor, via sudden insight (Aha! moment) or deliberate analysis, influence the quality of generation and subsequent learning outcomes? According to some research, deliberate processing enhances knowledge retention; hence, generation via analysis likely leads to better concept learning. However, other research has shown that solutions generated via insight are better remembered. In the current study, participants were presented with science concepts and descriptions, then generated metaphors for the concepts. They also indicated how they generated each metaphor and rated their metaphor for novelty and aptness. We assessed participants’ learning outcomes with a memory test and evaluated the creative quality of the metaphors based on self- and crowd-sourced ratings. Consistent with the deliberate processing benefit, participants became more familiar with the target science concept if they previously generated a metaphor for the concept via analysis compared to via insight. We also found that metaphors generated via analysis did not differ from metaphors generated via insight in quality (aptness or novelty) nor in how well they were remembered. However, participants’ self-evaluations of metaphors generated via insight showed more agreement with independent raters, suggesting the role of insight in modulating the creative ideation process. These preliminary findings have implications for understanding the nature of insight during idea generation and its impact on learning.

1. Introduction

People use metaphors to comprehend and convey new ideas in almost all domains (Lakoff & Johnson, 1980). Educators routinely harness the power of metaphor to facilitate teaching new concepts (Amin, 2015; Cameron, 2002; Tiberius, 1986). However, much of the research in education has focused on how metaphor comprehension helps learning, with less research into production or generation. The act of generating metaphors for newly acquired concepts requires a genuine comprehension of the material and a cognitive leap from passive absorption to active construction of cross-domain mappings (Gentner et al., 1997; Holyoak & Thagard, 1997; Jamrozik, McQuire, Cardillo, & Chatterjee, 2016; Lakoff, 1993). While learning the concept of “cornea,” for example, students may think to themselves, “Cornea is the windshield.” Recently, metaphor generation has been increasingly studied as a creative process (Beaty et al., 2016; Benedek et al., 2014; Chen et al., 2023). The present study draws perspectives from both the fields of learning and creativity to explore the relationship between the manner in which people generate a metaphor (via sudden insight or deliberate analysis, see below), the learning outcome as assessed by a subsequent memory test, and the quality of the generated metaphor.

1.1. Does the manner of generation influence the learning outcome?

An important way for people to generate new ideas is via sudden insight, also known as an Aha! moment. Insight is often associated with a mixed feeling of surprise, pleasure, and certainty (Gick & Lockhart, 1995; Topolinski & Reber, 2010) when people suddenly become aware of a new idea. In contrast to Aha! moments when people cannot describe how the idea occurs to them (Schooler & Melcher, 1995), sometimes we reach an idea with deliberate and reportable step-by-step analysis (Bowden & Jung-Beeman, 2007; Danek, Wiley, & Öllinger, 2016; Webb, Little, & Cropper, 2016).

We are not aware of any studies that explicitly investigated the role of insight in generating ideas in open-ended tasks to the best of our knowledge. Insight has mostly been studied using convergent tasks in which the problems are heavily constrained to admit only one or very few solutions. Such problems include classical insight problems (Maier, 1930; Webb, Little, & Cropper, 2018), verbal puzzles (E. M. Bowden & Jung-Beeman, 2003), and object detection (Mooney, 1957). When people solve problems in a convergent task, they can recognize the correct solution with relatively low ambiguity. In comparison, a generative task, such as creating metaphors, admits many possible solutions. People could produce just one of many acceptable responses, which vary greatly in quality and cannot be judged on correctness, per se. Thus, behavioral outcomes associated with the insight when people generate new ideas may not perfectly align with what have been observed in prior literature adopting convergent tasks.

The main goal of the current study was to test whether the manner in which people generate metaphors (insight or analysis) affects their learning. Different hypotheses can be inferred from previous literature. One possibility is that generating metaphors via analysis would be more likely to enhance learning outcomes compared to generating via insight. A step-by-step analysis is effortful and deliberate. Deliberate processing, a type of learning that involves focused attention, active retrieval, and feedback, has been shown to enhance the retention of knowledge when compared to passive and superficial learning (Brabeck & Jeffrey, 2014; Brown, Roediger, & McDaniel, 2014; Ericsson, 2015). While deliberate processing does not exclude the potential for insights, the Aha! experience when an idea seemingly emerges from nowhere implies that the critical processes leading to insights occur outside of conscious awareness (Schooler & Melcher, 1995; Smith & Kounios, 1996). When we consider this in conjunction with the deliberate processing benefit, it suggests that generating a metaphor via analysis tends to leave a stronger memory trace compared to generating via insight.

However, another possibility is that generating with insight, rather than analysis, will enhance learning outcomes considering an insight memory advantage demonstrated in prior creativity literature with convergent tasks. Several studies have revealed that solutions generated with an Aha! experience are remembered better than those generated without (or weaker) Aha! (Ash, Jee, & Wiley, 2012; Becker, Sommer, & Cabeza, 2023; Danek, Fraps, von Müller, Grothe, & Öllinger, 2013; Kizilirmak, Thuerich, Folta-Schoofs, Schott, & Richardson-Klavehn, 2016; Ludmer, Dudai, & Rubin, 2011). People are also better at recognizing items (the problem cues) that induce an Aha! as compared to no Aha!, in a later memory test (Kizilirmak, Galvao Gomes da Silva, Imamoglu, & Richardson-Klavehn, 2016). Therefore, the insight memory advantage would imply that generating a metaphor via insight is associated with better learning compared to generating via analysis.

It should be noted that a “deliberate analysis benefit,” consistent with much past education research, and an “insight memory advantage,” as suggested by past creativity research, are not directly comparable nor in conflict. These ideas come from different theoretical constructs in different fields, and the evidence for each is based on different experimental designs and tasks. For instance, in prior research with convergent tasks, the problem and its solution are closely tied together in that the solution often presents a unique closure to the problem. However, within a learning context, the degree of fit between the learning material and a generated thought can vary significantly. Therefore, when investigating the relationship between the manner of generation and its impact on learning, it is crucial to differentiate the memory of the presented concept (the stimulus) and the memory of the generated metaphor for the concept (the response).

1.2. Does the manner of generation influence the quality of the creative output?

A secondary goal of the study is to examine whether the manner of generating metaphors affects the metaphor quality. Similar to other creative output, the quality of a metaphor can be measured by its aptness and novelty (Thibodeau & Durgin, 2011). Metaphor aptness measures how well the concept and its metaphor fit appropriately with each other (Pierce & Chiappe, 2009; Stamenković, Milenković, Ichien, & Holyoak, 2023; Tourangeau & Sternberg, 1981), and novelty measures how unconventional and unfamiliar a metaphor is (Blasko & Connine, 1993; Silvia & Beaty, 2012). Little, if any, metaphor production research has asked how the manner of generation is associated with the quality of the metaphor product.

Perhaps a relevant finding is the well-documented “insight-accuracy effect.” That is, solutions generated via insight are more likely to be correct than solutions generated via deliberate analysis in convergent tasks (Becker, Wiedemann, & Kühn, 2020; Danek & Salvi, 2018; Salvi, Bricolo, Kounios, Bowden, & Beeman, 2016; Webb et al., 2018). If the aptness of metaphors resembles correctness (highly fit) in generated solutions, one may speculate that metaphors generated via insight are more apt than those generated via analysis. However, findings from convergent tasks may not be readily extended to a generative task. For one, a solution to a convergent task can be judged as right or wrong with relatively little ambiguity. In comparison, a generative task, lacking an unambiguous answer, can yield different patterns in both the generation process and the output (more in Discussion). Thus, this study has the potential to add to the understanding of the role of insight in generating creative ideas.

Furthermore, idea evaluation is a critical component of a creative process (Ellamil, Dobson, Beeman, & Christoff, 2012; Lopez-Persem et al., 2024). In generative tasks, people actively engage in the evaluation and selection of ideas, before producing a response. Therefore, self-evaluation of potential ideas directly contributes to the response quality. To explore how self-evaluation is associated with the manner of generation in the current study, we assessed the effectiveness of self-evaluation by comparing them to ratings provided by independent raters, following previous research (Rominger et al., 2022; Schraw, 2009). Although self-evaluation has been found to track objective scores closely (Kenett, Gooz, & Ackerman, 2023; Sidi, Torgovitsky, Soibelman, Miron-Spektor, & Ackerman, 2020), there have been mixed findings regarding the bias in people’s self-evaluation, influenced by diverse task designs and operationalizations (Puente-Díaz, Cavazos-Arroyo, & Vargas-Barrera, 2021; Sidi et al., 2020). Furthermore, the accuracy of self-evaluation can vary under different conditions or experimental manipulations (Puente-Díaz & Cavazos-Arroyo, 2022; Sidi et al., 2020), as people often rely on metacognitive cues to rate their responses (Lebuda & Benedek, 2023; Puente-Díaz, 2023). In the current study, we compared the self-rated novelty and aptness ratings with those obtained from a group of independent raters; this comparison was then examined according to how people generated the metaphors, via insight versus analysis. The result can contribute new empirical evidence regarding the relationship between the manner of idea generation and the effectiveness of self-evaluations.

1.3. The current experiment

To put the current study in a learning context, participants listened to science concepts and their descriptions first, then they generated metaphors for the concepts. Immediately after generation, we asked participants to report whether they generated each metaphor via insight or analysis and rate the metaphor quality (self-ratings). Separately, we asked 15–20 independent raters who did not participate in the main experiment to produce crowd-ratings that evaluated each metaphor response (Hass, Rivera, & Silvia, 2018). Finally, we measured the learning outcome by conducting a follow-up survey to gauge the memory of the target science concepts and metaphor responses 2 days after the initial session.

We seek to answer two main research questions: (1) How does the manner in which people generate a metaphor affect their memory? We will compare our findings to the different predictions drawn from the fields of learning (“deliberate analysis benefit”) and creativity (“insight memory advantage”), respectively. (2) How is the manner of generation associated with the qualities of the metaphor output? While we cannot make a prediction due to the lack of comparable studies, we will discuss our results in relations to the findings in the insight literature.

2. Method

2.1. Participants

Initially, a total of 50 people participated to receive course credits. They completed pre-experiment surveys and metaphor-generation tasks. Data from one participant were excluded from the analysis for not following instructions. The remaining participants included in the analyses had a mean age of 18.86 (SD = 1.09, men = 16). Furthermore, 11 participants did not comply with the follow-up schedule to complete the online survey, thus their data were not included in the memory-related analyses. The study was approved by the Institutional Review Board of a University in the United States.

We did not conduct a power analysis to estimate sample size because no comparable studies have looked at the memory effect of metaphor generation in different manners. However, the current sample size is comparable to previous publications investigating the behavioral impact of insight in problem-solving (Danek & Wiley, 2020; Danek et al., 2013).

2.2. Material

Science concepts were chosen from the subjects of cell biology, anatomy, and plant biology. Most concepts appear in Miller & Levine’s Biology Student edition textbook, a commonly used Biology textbook in local high schools. A list of 60 concepts were normed within a separate university undergraduate sample of 113 undergraduate students (Mean age = 20.67, sd = 3.33; 82 female). Norming participants rated concept familiarity on a 1–5 Likert scale. Average concept familiarity was 3.14, sd = 1.46.

The current study selected 30 concepts that had low familiarity ratings (less than or equal to 3.5) out of the full list. This criterion was chosen to avoid the ceiling effect when measuring the post-experiment familiarity in the later memory test.

A two-sentence description was created for each concept with the help of a professor collaborator in the molecular and cellular biology department who conducted research and taught cell biology and related courses. All sentences used in the current study were approved by this biology expert to be correct descriptions of the term’s biological function. See the online repository (osf.io/kscy2/) for the full list of the concepts, descriptions, as well as participants’ responses.

For example, the description for the concept “cornea” is: “The cornea is the clear transparent section of the eye. It is the outermost tissue of the eye that protects it from dirt and other harmful particles.” All descriptions contained 19–31 words and had the same sentence structure across all concepts.

We generated an audio clip for each description recorded in a female voice with no specific intonation. The clips were automatically generated using https://wideo.co/text-to-speech/. The lengths of audio clips range from 8 to 12 s.

Additionally, participants’ preferences in using metaphors (Fetterman, Bair, Werth, Landkammer, & Robinson, 2016) may interact with the relationships we seek to investigate. Some people might like to use metaphors, whereas others might be very literal. Therefore, before the experiment, we administered the Metaphoric Triads Task (MTT) (Kogan, Connor, Gross, & Fava, 1980) to account for metaphor preference. MTT consists of a total of 29 triads (trials). A triad consisted of three images, for example, “violin,” “singing canary,” and “tree.” Participants were to determine “which two of the three make the best pair?” People who scored high on the MTT tend to select the pair forming a more metaphoric relationship (e.g., “violin” and “singing canary”) out of the three possible pairings (e.g., “singing canary” and “tree”). MTT has demonstrated internal validity and has been widely used in the developmental context (Glicksohn & Yafe, 1998; Seitz, 2010).

2.3. Procedure

See Fig. 1 for an overview of the procedure. During the first session, after participants gave informed consent, they completed pre-experiment surveys, which included a concept familiarity rating survey and the MTT on a device in the lab. Next, they completed the main metaphor generation task. During the second session, 48–72 h following the first session, participants completed a web-based post-experiment memory test, which consisted of the concept familiarity rating survey again, a concept recognition task, and a recall of the self-generated metaphors.

Fig. 1.

Experiment procedure.

Notes. (A) Pre-experiment Surveys: Participants completed a pre-experiment concept-familiarity survey (on a personal device) and the Metaphoric Triads Task (on a computer) before proceeding to the metaphor generation task. (B) Main Tasks: A trial includes an audio clip of a concept description, metaphor generation, a forced choice question of insight or analysis, self-rating of the metaphor’s novelty and aptness, and a brief explanation of how the metaphor fits with the concept. (C) Post-experiment Memory Test: In a follow-up session, participants completed an online survey to test how well they remembered the concepts they encountered during the previous session, as well as the metaphors they generated.

2.3.1. Pre-experiment concept familiarity

To obtain participants’ baseline familiarity with the science concept before the experiment, we asked them to rate their familiarity with each science concept on a 1–4 scale: unfamiliar, somewhat unfamiliar, somewhat familiar, familiar (Fig. 1 Panel A).

2.3.2. Main tasks

The main tasks (Fig. 1 Panel B) consisted of (a) metaphor generation task; (b) manner of generation question; (c) aptness and novelty rating; and (d) metaphor explanation.

• aIn the metaphor generation task, participants were instructed to come up with a metaphor for the concept based on its function described in the audio clips. A metaphor was described as “a figure of speech that describes something in a way that isn’t literally true but helps explain an idea or make a comparison.” They were provided a sample concept of “cornea” and an example metaphor for its function, “windshield.” Participants went through a practice trial with feedback to ensure they understood what a metaphor is.
• bParticipants were then explained different ways in which they might generate a metaphor, via insight or analysis:

Insight is like an Aha! moment. The idea comes to mind as a sudden surprise. It may be difficult to articulate how you got the idea. Analysis means that you put the answer together gradually. You might have searched with a deliberate strategy, and you are able to report the steps you took.

Participants were asked to indicate whether they generated via insight or analysis after each metaphor generation. They were instructed not to overthink and select the closest choice even if it felt like a mixture. Following previous insight research, we adopted the self-report measure to distinguish the manner of generation. Self-reports have been widely used to study neural substrates underlying insight (Danek, Williams, & Wiley, 2020; Laukkonen & Tangen, 2018). Leveraging self-report measures, extensive behavioral (E. Bowden, Jung-Beeman, Fleck, & Kounios, 2005; Yu, Salvi, Becker, & Beeman, 2023), physiological (Salvi, Simoncini, Grafman, & Beeman, 2020), and neuroimaging evidence (Becker, Sommer, & Kühn, 2020; Jung-Beeman et al., 2004; Santarnecchi et al., 2019; Sprugnoli et al., 2017; Yu, Oh, Kounios, & Beeman, 2022) has shown that solving with insight versus analysis recruits different cognitive and neural processes (for review, see Kounios & Beeman, 2014; Salvi, 2023.

• cParticipants were told that they would rate the novelty (how uncommon or creative) and aptness (how good and well-suited) of the metaphor.
• dParticipants were instructed that they would explain the metaphors they came up with.

After the abovementioned instructions, the session began. Each trial was initiated by the participant pressing the space bar. A concept word was first displayed at the center of the screen for 2 s. Then, an audio clip describing the scientific concept was played with the concept word on the screen. After the audio clip, the concept word remained on the screen while the participant was asked to come up with a metaphor for the concept. If they failed to generate a metaphor within 1 min, the experiment proceeded to the next trial. They were instructed to press the space bar as soon as they had an idea. The button press advanced the program to the next screen, where participants were asked to type the metaphor they generated.

Participants then made a binary choice about whether they generated the metaphor via insight or analysis, and rated the novelty and aptness of their metaphor on a 1–4 scale, where 1 meant not at all novel (apt) and 4 meant highly novel (apt). Finally, they entered a brief explanation for their metaphor on the screen before moving on to the next trial. The explanation was later on presented to independent raters (see below) before they evaluated the metaphor qualities.

2.3.3. Crowd-sourced ratings

To measure the creative quality of the metaphor responses from an independent source, we asked a separate group of undergraduate students to rate the novelty and aptness of each response on a 1–4 scale. Raters were provided with the same definition of novelty, aptness, and the rating scale as the participants. A total of 150 raters completed the online rating task for course credit. Each response received 15–20 ratings from different raters with an intraclass correlation coefficient (ICC; Shrout & Fleiss, 1979) of 74.9% for aptness and 45.6% for novelty. The ICC for novelty was on the lower end of what has been reported in creativity literature. However, compared to previous studies that mostly included 2–3 expert raters, the crowd-rating relying on laypeople may result in more noisy measurements (Hass et al., 2018).

In the text below, we used Crowd-Novelty/Aptness to differentiate crowd-sourced ratings from self-ratings, referred to as Self-Novelty/Aptness.

2.3.4. Post-experiment memory test

Two days after the metaphor generation session, participants were provided a link to complete a Qualtrics memory survey which was required to be completed within 24 h (Fig. 1 Panel C). During the survey, in each trial, they were presented with a science concept and asked to rate the familiarity of each science concept (post-experiment familiarity), whether the concept had appeared in the previous experiment (concept recognition), and if yes, what metaphor they generated (metaphor recall). Ten additional science concepts that did not appear in the experiment were included in the survey as lure items.

2.4. Analyses

To test how the manner of generation is associated with behavioral outcomes (learning outcome and metaphor qualities), we used mixed effect models to account for the correlation among variables, as each participant provided multiple sets of observations (dependent and independent variables). In all mixed-effect models (lme4 package in R), we included a random intercept and a random slope for the manner of the generation (insight or analysis) per participant. For fixed effects, we included the pre-experiment concept familiarity and the manner of generation (insight or analysis) as level-1 variables, because pre-experiment concept familiarity may directly contribute to the metaphor quality and could potentially confound the factors of interest.

To analyze the learning outcomes, we fitted three mixed-effect models corresponding to the three memory measures as dependent variables: the concept familiarity change, the recognition of the target science concept, and the recall of the generated metaphor. We also included the novelty and aptness of the metaphor responses because the quality of the generation outputs may explain variability during learning. Both recognition and recall are binary variables. Therefore, we used generalized mixed-effect models (glmmTMB package in R) in these two cases. In addition, to avoid a potential ceiling effect, we excluded trials in which the participant reported the highest score, level 4 (i.e., 23% of all trials) during the pre-experiment familiarity survey for that concept. We included these trials in a secondary analysis (Table S1) and observed similar results.

To investigate the creative quality of the generated metaphors, we used Crowd-Aptness and Crowd-Novelty ratings as dependent variables in two separate models.

In all mixed-effect models, we also included the MTT score as a level-2 variable to represent individual differences in metaphoric preference. Visual inspection revealed that the MTT scores (Mean = 8.5, SD = 6.3; range = [0,28]) were positively skewed (more extreme values representing strong metaphoric preference), failing the normality check (D’Agostino’s Kˆ2 test, p = .037). Therefore, we applied the log transformation to the MTT scores before entering it into the model so that results were less biased by a few large values. The transformed scores passed the normality check (p = .155, thus the null hypothesis of Gaussian distribution could not be rejected).

We did not analyze response time in the main analyses. The median response time (from the offset of the audio description to the bottom press indicating response) was 2.8 s. Twenty-three percent of the trials took less than 1 s. The overall short response time indicated that people started to think of a metaphor during the audio listening. Therefore, the response time did not reliably measure the actual duration it took to generate the metaphor.

3. Results

Participants generated metaphors in 96.3% (SD = 5.7%) of all trials. They reported that they generated with insight 57.9% (SD = 21.9%) of the time and with analysis for the remaining 42.1% (SD = 21.2%).

We measured metaphor quality with crowd-sourced novelty (Mean = 2.44, SD = 0.39) and aptness ratings (Mean = 2.46, SD = 0.49). The two measures were significantly correlated with each other (r = .566).

3.1. Metaphor-generation and learning outcomes

To answer our first research question, we associated the manner of generation (via insight or analysis) with learning outcomes measured through a post-experiment memory test that tested participants’ memory of the target science concepts and their generated metaphors.

We first turned to the memory of the target concepts. Participants showed an overall improvement in concept familiarity, from an average of 2.32 (SD = 1.20) pre-experiment to 2.75 (SD = 1.07) post-experiment, representing a statistically significant increase (z = 14.82, p < .001). The familiarity change was bigger (Mean = 0.65, SD = 0.99) if excluding trials where participants rated concept familiarity at level 4 on a 1–4 scale before the main experiment. Concept memory was also tested via a recognition test. Participants recognized 72.1% (SD = 18.5%) of the presented concepts. In comparison, participants said that 31.7% (SD = 21.8%; Cohen’s d = 2.02) of the lure concepts (not presented) appeared during the main task.

We used mixed-effect models to test the main research question: whether generation via insight or analysis is associated with learning outcomes. The model, fitted to 831 trials provided by 39 participants, indicated that generating metaphors via analysis led to more familiarity gains with the concept (p = .013, Table 1 and Fig. 2A), compared to generating via insight. Including trials that had level 4 pre-experiment concept familiarity yielded similar results (Table S1). Generation via analysis also slightly increased the likelihood of concept recognition but the effect was not significant (p = .097, Table 1 and Fig. 2B). The same results were observed when fitting a model with self-ratings instead of crowd-ratings (Table S2A).

Table 1.

Metaphor generation affects the memory of the target science concept

Concept familiarity change				Concept recognition
Predictor	Estimates	CI	p	Predictor	Estimates	CI	p
(Intercept)	1.12	0.53–1.71	<.001	(Intercept)	−1.19	−1.62 to 1.25	.798
Crowd-aptness	−0.14	−0.28 to 0.00	.049	Crowd-aptness	−0.59	−1.00 to −0.18	.005
Crowd-novelty	0.16	−0.01 to 0.34	.069	Crowd-novelty	0.38	−0.14 to 0.89	.149
i or a [insight]	−0.15	−0.26 to −0.03	.013	i or a [insight]	−0.29	−0.64 to 0.05	.097
Pre familiar	−0.45	−0.53 to −0.38	<.001	Pre familiar	0.73	0.57–0.90	<.001
MTT	0.22	−0.00 to 0.42	.045	MTT	0.20	−0.25 to 0.65	.389
Random effects				Random effects
σ 2	0.57			σ 2	3.29
τ00 Subject	0.21			τ00 Subject	0.99
τ11 Subject.insighti	0.00			τ11 Subject.insighti	0.01
ρ01 Subject	1.00			ρ01 Subject	−1.00
N Subject	39			N Subject	39
Observations	831			Observations	1083
Marginal R2/Conditional R2		0.234/0.409		Marginal R2/Conditional R2		0.168/0.342

Fig. 2.

Memory of the target science concept: familiarity and recognition.

Notes. The average change in (A) concept familiarity and (B) recognition rate for the target science concept after generating metaphors via insight and analysis.

Although not hypothesized, the above models also suggested that the quality of the generated metaphors affected participants’ learning outcomes. The Crowd-Aptness rating was negatively associated with the change in familiarity and recognition regarding the target concept (p = .049 and .005, respectively). In other words, generating metaphors that were deemed less apt by independent raters was associated with a better memory of the target concept. This effect only held for Crowd-Aptness, but not the Self-Aptness ratings (Table S2A).

Next, we turned to the memory of the generated metaphor. Out of the recognized concepts, participants recalled 30.2% (SD = 20.6%) of the metaphors they generated for the concepts. We used the binary (recall or not) variable as a dependent variable in a mixed-effect model and fit the model to a subset of the data where participants recognized the target concept (otherwise the recall question was not presented). The manner in which people generated metaphors did not affect their metaphor recall (Table 2). However, there was a significant and positive effect of Crowd-Aptness, suggesting that participants were more likely to recall their metaphor responses if those metaphors were rated as apt by independent raters. The model with self-ratings showed the same pattern (Table S2B).

Table 2.

The manner of generation does not affect the memory of the generated metaphor

Metaphor recall
Predictor	Estimates	CI	p
(Intercept)	−4.36	−6.14 to −2.57	<.001
Crowd-aptness	0.92	0.41–1.43	<.001
Crowd-novelty	−0.43	−1.03 to 0.18	.166
i or a [insight]	0.01	−0.43 to 0.44	.974
Pre familiar	0.66	0.47−0.85	<.001
MTT	0.12	−0.39 to 0.63	.641
Random effects
σ 2	3.29
τ00 Subject	1.13
τ11 Subject.insighti	0.05
ρ01 Subject	−1.00
N Subject	39
Observations	777
Marginal R2/Conditional R2		0.171/0.349

Taken together, the manner of metaphor generation had different effects on the memory of the target concepts and the memory of the generated metaphors. Participants’ familiarity with the concept increased more after generating metaphors via analysis compared to generating via insight. However, the manner of generation did not change the likelihood of recalling their metaphor responses. Another interesting finding was that the metaphor quality, measured by Crowd-Aptness, was negatively associated with the memory of the target concepts but positively associated with the memory of the generated metaphors (see Discussion for potential interpretations).

3.2. The creative quality of generated metaphors

Although not directly hypothesized, the above analyses suggested that the quality of the generation product (metaphors) was associated with the learning outcome. Thus, we turn to our second research question: whether and how the manner of metaphor generation, via insight or analysis, is associated with the metaphor quality.

The mixed-effect model, fitted to 1415 trials provided by 49 participants, indicated that metaphors generated via insight were neither better nor worse than metaphors generated through analysis, in terms of either Crowd-Novelty or Crowd-Aptness (Table 3).

Table 3.

Mixed-Effect Model for crowd-rated novelty (left) and aptness (right) of generated metaphors¹

Model for Crowd-Novelty				Model for Crowd-Aptness
Predictor	Estimates	CI	p	Predictor	Estimates	CI	p
(Intercept)	2.29	2.14–2.44	<.001	(Intercept)	2.25	2.08–2.43	<.001
i or a [insight]	0.00	−0.05 to 0.05	.895	i or a [insight]	−0.02	−0.09 to 0.05	.523
MTT	0.09	0.02–0.15	.013	MTT	0.11	0.03–0.19	.006
Pre familiar	−0.01	−0.03 to 0.01	.277	Pre familiar	0.00	−0.02 to 0.03	.962
Random effects				Random effects
σ 2	0.11			σ 2	0.19
τ00 Subject	0.03			τ00 Subject	0.03
τ11 Subject.insighti	0.01			τ11 Subject.insighti	0.03
ρ01 Subject	0.22			ρ01 Subject	0.11
N Subject	49			N Subject	49
Observations	1415			Observations	1415
Marginal R2/Conditional R2		0.028/0.274		Marginal R2/Conditional R2		0.031/0.246

The model also indicated that participant’s metaphor usage preference, measured by MTT (Metaphoric Triads Task), influenced the quality of metaphors they generated. The higher participants scored on MTT, the higher rated the metaphors they produced, in both Crowd-Novelty (p = .013) and Crowd-Aptness (p = .006). (Table 3). In other words, people who prefer metaphor usage tend to generate more apt and novel metaphors.

3.3. Insight and self-evaluation

Next, we asked whether and how the manner of generation was associated with the accuracy of participants’ self-evaluations. For metaphor responses generated via insight (811 trials), the correlation between Self-Aptness and Crowd-Aptness was .40, which was significantly higher than the correlation of .25 (604 trials) in those generated via analysis (Fisher’s z = 3.34, p < .001), suggesting that people rated the metaphor aptness more consistently with independent judges after generating via insight, compared to generating via analysis.

Similarly, the correlation between Self-Novelty and Crowd-Novelty ratings of metaphors generated via insight was .25, which was significantly higher than the correlation of .13 in those generated via analysis (z = 2.23, p = .018). In other words, people rated the novelty of their metaphors more consistently with independent judges after generating via insight, compared to generating via analysis.

Furthermore, the bias of self-ratings relative to crowd-ratings was modulated by the manner of generation. The average difference between Self-Aptness and Crowd-Aptness was 0.32 (SD = 0.99) for analysis responses, significantly higher than the difference of 0.13 (SD = 0.90) for insight responses (t = 3.87, p < .001; Cohen’s d = 0.22). This suggested that participants tended to rate the aptness (but not the novelty) of their own responses higher than independent raters although the overestimation was tempered for insight responses (Fig. 3).

Fig. 3.

Difference between self-ratings and crowd-ratings.

Notes. The difference between self-ratings and crowd-ratings grouped by the manner of generation. Error bars indicate 1 standard error.

Altogether, the current exploratory analysis suggests that people make self-evaluations more aligned with others after insight compared to after analysis, exhibiting higher correlation and less bias relative to independent judges.

4. Discussion

We investigated whether the manner of generation, via insight or analysis, was associated with the metaphor quality and learning outcomes. To this end, we implemented a metaphor-generation paradigm in a concept-learning context, measuring aptness and novelty for metaphor quality and measuring concept familiarity change, target concept recognition, and generated metaphor recall for learning outcomes.

4.1. Implications in learning and education

Participants showed more familiarity gain with the target science concepts if they had generated metaphors for them via analysis, compared to generating via sudden insight, in a previous session. While the implication of the current finding in learning should be considered with caution (see 4.3 Limitations), the result resonates with the emphasis on deliberate processing for enhanced learning effects in education literature (Brabeck & Jeffrey, 2014; Brown et al., 2014; Ericsson, 2015). Our finding suggests that generating metaphors via analysis required the learners to deliberately and explicitly consider features of the target concepts, which subsequently led to better familiarity and recognition.

Besides answering the main research question, we also reported a set of interesting, exploratory findings that can be informative to future research. The memory test also revealed an interesting set of relations between metaphor aptness and concept memory. People remembered a target concept better if they generated a less apt, compared to a more apt, metaphor for it. Perhaps participants produced a relatively low-apt metaphor because they had struggled to understand the target concept, compared to concepts for which they produced more apt metaphors. While they did not generate an apt metaphor, their prolonged efforts to understand the concept might help them encode it better, and make it feel more familiar later. Thus, the low-apt responses may indicate “desirable difficulties” during learning which have been shown to benefit long-term retention (Bjork & Bjork, 2020; de Bruin et al., 2023).

We also found that people were more likely to recall the metaphor they generated if their metaphor was apt. Tests for recalling the metaphors in the current study were conditioned on recognizing the target concept in the first place. Therefore, participants were more likely to remember their high-aptness metaphors, as these metaphors were more closely related to the target concepts, which serve as better retrieval cues than the concepts associated with low-aptness metaphors.

Finally, regarding the relationship between the manner of generation and self-evaluation, participants tended to overestimate the aptness of their metaphor responses relative to the crowd ratings. This finding appears to resonate with the “above-average effect” in education literature where people tend to be overconfident in their learning performance (Dunlosky & Rawson, 2012; Kruger & Dunning, 1999; Metcalfe, 1998). However, the overestimation was tempered when they were generated via insight (see further discussion on this below).

4.2. Implications in creative cognition

We found that generation via insight, compared to analysis, was negatively associated with the memory of the presented target concepts. While this result is consistent with the deliberate processing benefit as discussed above, it highlights the difference between the current study and previous creativity research that evidenced an insight memory advantage—solutions generated via insight were better remembered (Danek & Wiley, 2020; Danek et al., 2013; Kizilirmak et al., 2016; Ludmer et al., 2011). It is worth noting that prior insight memory research has mostly tested the memory of the solution, rather than the stimulus, in a convergent task context. Even in studies where a recognition test on the stimulus (problem cue or presented material) was administered, the stimulus and the solution were closely bound due to the nature of the task (Becker & Cabeza, 2023; Kizilirmak et al., 2016). For example, in a Mooney task where participants need to recognize an object from a degraded image, the solution—a perceived object—can serve as a mental cue for the previously presented image. Thus, the insight memory advantage pertaining to solutions may not map to improved memory of stimuli such as the presented target concept in the current study.

Further, why were generated metaphors equally well-remembered whether generated via insight or analysis in the current study? Unlike a convergent task where a solution represents a closure to the problem, the generative task in the current study yields responses that cannot be clearly judged as right or wrong and hence no closure. The difference in the task structure can lead to a different Aha! experience. While people may have a common understanding of an “Aha!” moment, the experience is multifaceted. Some researchers emphasized the suddenness and surprise (Jung-Beeman et al., 2004; Metcalfe, 1986; Schooler, Ohlsson, & Brooks, 1993), while others also included pleasure and certainty (Becker et al., 2020; Danek et al., 2013; Webb et al., 2016) as defining characteristics. In convergent tasks, different emotional components: surprise, certainty, and pleasure, appear to be highly correlated in Aha! moments (Becker et al., 2023; Danek & Wiley, 2017; Danek, Fraps, von Müller, Grothe, & Öllinger, 2014; Hedne, Norman, & Metcalfe, 2016). The insight memory advantage has been particularly attributed to the feeling of certainty and pleasure (Danek & Wiley, 2020). Because there is variability in people’s experience of solving under different task contexts (Chesebrough, Chrysikou, Holyoak, Zhang, & Kounios, 2023; Metcalfe & Wiebe, 1987; Webb et al., 2016), it is possible that our participants did not feel certainty and pleasure after generating metaphors via insight, at least not strongly enough to lead to a better memory of the generated metaphor responses. In other words, the feeling of suddenness in the current task may not overlap with other characteristics of insight as in prior studies. Therefore, insight-generated metaphor was not associated with better encoding and lasting memory, in contrast to insight-generated solutions.

Next, we compare the null relationship between self-reported insight and the metaphor quality (novelty and aptness) in the current study with the widely documented accuracy effect in insight literature—solutions generated with an Aha! were more likely to be correct compared to those generated via analysis (Danek et al., 2014; Hedne et al., 2016; Webb et al., 2016). The lack of the relationship in the current study is consistent with the multifaceted Aha! experience as discussed above. Different emotional components of an Aha! experience has varied relationships with solution accuracy, and not all dimensions equally predict solution accuracy. For example, confidence and pleasure are correlated with solution accuracy, while surprise is not (Webb et al., 2016). In convergent tasks where the feeling of confidence and pleasure co-occurs with suddenness, the accuracy effect can be robustly observed. For example, the accuracy effect was observed in verbal puzzle tasks when the same instruction for reporting insight as in the current study was used (Salvi et al., 2016; Yu et al., 2023). It is possible that, in a generative task, the feeling of suddenness (perhaps more akin to surprise) decouples from confidence and pleasure, and therefore, does not predict the quality of the response. This indeed seems to be the case in the current study because insight metaphors were not associated with self-rated novelty or aptness. Previous literature has associated the ability to generate creative metaphors with fluid intelligence (Beaty & Silvia, 2013; Primi, 2014), executive control (Menashe et al., 2020; Silvia & Beaty, 2012), and a more flexible semantic network (Li, Kenett, Hu, & Beaty, 2021). Rather than focusing on individual differences as in the above work, the current study provided information on the within-participants variabilities. The null relationship between the manner of generation and response quality suggested that creative metaphors can be achieved via different processes, including deliberate analysis and spontaneous insight.

Finally, our results indicate that people make evaluations of their metaphors more aligned with others when generated via insight compared to analysis. Idea evaluation has long been deemed an important component of creative cognition (Campbell, 1960; Kleinmintz, Ivancovsky, & Shamay-Tsoory, 2019), and there is an increasing interest in exploring this aspect empirically in a metacognitive framework (Kaufman & Beghetto, 2013; Lebuda & Benedek, 2023; Puente-Díaz, 2023). There have been mixed findings regarding the bias (over- or underestimation) and the accuracy of the self-evaluation under different task designs (Puente-Díaz et al., 2021; Sidi et al., 2020). One possible explanation for the current result is that the deliberate effort people engaged during analysis taxed cognitive resources, which then negatively affected the metacognitive performance. It is also possible that a metaphor generated via insight touches upon shared features between the metaphor and the concept that are salient to all (self and crowd), whereas a metaphor generated via analysis requires a few more turns in one’s own mind and is not readily comprehensible to other people. The current findings contribute to the growing literature on creative metacognition and suggest the potential role of insight during idea evaluations.

4.3. Limitations and future directions

The current study suggested that the manner of generating a metaphor affects the memory of the target concept, as measured by concept familiarity change and concept recognition. Both familiarity and recognition are widely used in memory research. Students may simply remember having seen and learned about the concept (familiarity) without remembering the details of the concept (recognition). To fully understand how the manner of generation modulates learning outcomes, future work should develop a more multifaceted test to assess concept knowledge retention.

One limitation of the current study is that participants reported insight (vs. analysis) via a binary choice that emphasized the feeling of a sudden surprise. Without additional multidimensional and emotional measures such as the feelings of pleasure and certainty, the insight studied here might differ from those insight experiences in a convergent task. Therefore, future work should measure and quantify suddenness, certainty, and pleasure on separate scales to improve our understanding of the nature of insight.

Second, while we asked participants to provide a brief explanation of their metaphor, we did not provide detailed instruction for the explanation. This led to highly heterogeneous explanations. They differ in the degree of detail and content (such as explaining shared features or describing the generation process). We did not analyze the explanations in the current study, although they may potentially provide rich information and be associated with learning outcomes.

4.4. Conclusion

In summary, we conducted exploratory analyses to investigate the relationship between the manner of generating metaphors and its impact on learning outcomes and metaphor quality, as a first study connecting these concepts. Our results indicate that generating via analysis, compared to insight, enhances memory retention of the presented learning material, aligning with the established benefit of deliberate processing. Additionally, we provided preliminary evidence suggesting that the characteristics of insight in an open-ended, generative task may differ from insight in convergent tasks, both in terms of phenomenology and response quality. While the findings reported here are preliminary, they can inform future research that connects the fields of learning and creative cognition.