Brainstorming with a Generative Language Model: Effect of Exposure to AI Ideas on Brainstorming Performance and Cognitive Load

For hypothesis set 1: human performance, we investigate how the human brainstorming performance, i.e., considering only ideas originating from the human, excluding GLM ideas, varies across the two conditions, uncovering how typical brainstorming performance dimensions are affected by GLM usage. By adding a GLM, individual human brainstorming is transformed into collaborative human–GLM brainstorming. Previous research on all-human brainstorming groups suggests that individual performance can be enhanced or harmed due to being part of a group (Pinsonneault et al. 1999). While these effects are well-documented in purely human groups, and while humans might socially respond to technical systems (Nass and Moon 2000), such effects in groups with non-human members (Stieglitz et al. 2022; Siemon and Wank 2021; Liu et al. 2023; Tao et al. 2023) are still being researched, i.e., it is uncertain whether humans themselves perform differently when working with or without a GLM (Tao et al. 2023; Specker et al. 2025). Previous research suggests group effects like cognitive stimulation, social loafing, or cognitive inertia to be relevant in our setting (Stieglitz et al. 2022; Tao et al. 2023; Memmert 2024; Memmert and Tavanapour 2023).

In brainstorming sessions, human fluency decreases after the first few minutes (Baruah and Paulus 2016; Kohn and Smith 2011). When running out of ideas, humans working with the GLM can request GLM ideas. These may act as stimuli, activating the human’s memory and triggering the human to themselves develop more own ideas, i.e., cognitive stimulation could occur (Nijstad et al. 2010; Dugosh et al. 2000), increasing the human’s fluency (Dugosh et al. 2000; Althuizen and Reichel 2016) as compared to solitary humans, who do not receive stimuli. However, a human working with a GLM might reduce their effort (i.e., social loafing could occur), particularly if they feel their effort is dispensable or feel less responsible for the results (Pinsonneault et al. 1999). Given that GLMs can outperform many humans in generating ideas (Haase and Hanel 2023), humans might feel dispensable and might, therefore, reduce their effort. Additionally, previous research suggests that humans might attribute some of the responsibility to the AI (Stieglitz et al. 2022; Memmert and Tavanapour 2023; Memmert 2024), potentially leading to a reduction in effort and fewer ideas contributed. While the above suggests both performance-enhancing and -reducing effects, recent studies of multi-human groups suggest that the flow of GLM ideas may make it easier to come up with more ideas (Specker et al. 2025; Zhang et al. 2025; Nomura et al. 2024) and may reduce loafing (Nomura et al. 2024); hence, we propose:

For flexibility or breadth of exploration, different effects could influence performance. Cognitive inertia refers to producing more ideas in the same category (i.e., higher within-category fluency) but exploring fewer categories (i.e., lower flexibility). It can occur when brainstorming alone (here, human-only condition) without seeing others’ ideas, thereby receiving external stimuli (Pinsonneault et al. 1999), e.g., from a GLM (here, human + GLM condition). However, the effect of GLM ideas depends on the degree of relatedness and the amount. Exposure to many closely related ideas might lead to higher conformity to those idea or idea categories, and, consequently, to reduced flexibility (Lamm and Trommsdorff 1973; Kohn and Smith 2011). On the other hand, if the stimuli like GLM ideas are diverse, covering different categories, inertia might be reduced (Briggs and Reinig 2010), aiding humans to explore new categories, i.e., increasing flexibility. Given recent literature suggesting that AI and GLMs, specifically, can expand the problem and solution space (Specker et al. 2025; Bouschery et al. 2023; Nomura et al. 2024), we suggest:

Stimuli can enhance idea novelty and value (Althuizen and Reichel 2016). However, exposure to typical or related ideas is associated with reduced idea novelty and increased idea value, and exposure to more ideas with increased fixation (Kohn and Smith 2011; Wang and Nickerson 2019). Prior literature comparing human and GLM ideas shows that GLMs can produce ideas of comparable or even higher creativity compared to many humans (Haase and Hanel 2023), which should result in humans also generating more novel ideas (Wang and Nickerson 2019). Results on the impact of GLMs on individual performance are mixed. Kumar et al. (2025) find that when humans work in a short brainstorming session (2 min) and receive a limited number of GLM ideas (7), they do not produce more novel ideas. However, Specker et al. (2025) report that when working with a GLM, participants report perceiving increased idea novelty at an individual level. We propose that when participants have more time and can request new GLM ideas without restriction, this high number of creative GLM ideas should induce cognitive stimulation (Leggett Dugosh and Paulus 2005) and should allow humans to more broadly and deeply explore the topic, which is associated with more novel and valuable ideas (Althuizen and Reichel 2016). Therefore, we suggest:

For hypothesis set 2: dyad performance, we compare the overall performance between the conditions, uncovering if superior performance by the human–GLM dyad compared to the individual – a key ambition for human–AI groups (Dellermann et al. 2019) – is achieved.

GLMs “are programmed to create vast amounts of text” (Haase and Hanel 2023, p. 10) and were shown to be capable of generating ideas (Haase and Hanel 2023; Dell’Acqua et al. 2023). Humans working with the GLM could repeatedly request and copy GLM ideas, steadily increasing group fluency directly. Thus, human–GLM dyads might outperform solitary humans with respect to fluency. In multi-human groups, GLM-supported groups produced more ideas (Specker et al. 2025). Therefore, we propose:

GLMs can produce many ideas, and a higher number of ideas is associated with higher flexibility (Nijstad et al. 2010). However, due to how machine-learning-based AI works, particularly as it is trained on historical data, there is a discussion on whether AI can be creative (Wu et al. 2021) or produces “more of the same”. Indeed, GLMs might reproduce bias and show limited output diversity (Bender et al. 2021; Lin et al. 2022). In a brainstorming context, producing “more of the same” might manifest in the GLM suggesting similar ideas, i.e., taking a narrow perspective on the topic, which aligns with the view of creativity being a comparative human strength (Dellermann et al. 2019). Additionally, research suggests humans working with GLMs to produce ideas with less variability, i.e., potentially lower flexibility (Dell’Acqua et al. 2023). However, recent work on group brainstorming indicates higher flexibility when working with a GLM (Specker et al. 2025). We thus propose:

Higher fluency and higher flexibility are associated with higher originality (Nijstad et al. 2010). Potentially, with the exploration of less common categories, less common (i.e., more original) ideas are explored (Nijstad et al. 2010). Additionally, a more in-depth exploration of categories is related to higher originality (Althuizen and Reichel 2016). Thus, the human and the GLM could jointly explore the problem more broadly or more in-depth, both of which is associated with higher originality (Nijstad et al. 2010). Previous research separately investigating humans and GLMs suggests that GLMs can produce ideas of comparable novelty and value to humans or even outperform many humans, though the most novel ideas still originate from humans (Haase and Hanel 2023; Boussioux et al. 2023). Thus, it was suggested that the human and AI could collectively outperform either entity alone, with humans focusing on producing novel and GLMs on producing valuable ideas (Boussioux et al. 2023). Furthermore, in a human–GLM brainstorming setting, many participants selected GLM ideas as their best ideas (Memmert and Tavanapour 2023). We thus propose:

For hypothesis 3: cognitive load, we consider how the cognitive load differs across the conditions. Social loafing describes the reduction of effort at the expense of others. Non-human group members may relieve cognitive burdens on humans and support them during task execution (Stieglitz et al. 2022; Boussioux et al. 2023). Therefore, the reliance of the human on AI might be desirable to maintain cognitive resources and improve work efficiency (“smart loafing”, Stieglitz et al. 2022), and previous research shows that humans might save cognitive resources when working with AI (Brachten et al. 2020). It is unclear whether this effect transfers onto divergent thinking tasks like brainstorming, i.e., whether humans brainstorming with GLMs could maintain cognitive resources and would, therefore, report lower levels of cognitive load upon task completion. However, having to pay attention to others’ ideas during brainstorming could result in a distraction effect (Pinsonneault et al. 1999) and potentially increase cognitive load (Coskun et al. 2000) due to the need to process the stimuli and additional knowledge activation (Santanem et al. 2004). Particularly, if the stimuli are of low relatedness, this could increase cognitive load if distant or even topic-irrelevant knowledge (Wang and Nickerson 2019) is activated. Moreover, prior research suggests that additional cognitive demand is placed on humans when working with GLMs, because they need to evaluate the outputs (Tankelevitch et al. 2024). Research for multi-human brainstorming groups showed that too many GLM hints can increase perceived cognitive load (Zhang et al. 2025) and that participants do not simply accept GLM ideas, but review and evaluate them (Gonzalez et al. 2024), adding to the cognitive load. We, therefore, propose:

We analyzed the responses to the open-ended survey questions to shine a light on potential cognitive stimulation or inertia when working with a GLM and receiving its dynamically generated ideas. In line with prior work (Houde et al. 2025; Memmert and Tavanapour 2023), several participants stated to have been inspired by the GLM ideas to start generating ideas, to generate additional ideas, or to get unstuck (sic!):

Cognitive stimulation might manifest in humans producing more, or more novel or valuable ideas (Althuizen and Reichel 2016; Dugosh et al. 2000). However, as described above (hypothesis set 1), humans in the human + GLM condition produced fewer ideas themselves (human_fluency) and did not show higher levels of flexibility, novelty, or value (human_flexibility, human_novelty, human_value), which seems contrary to these subjective reports of feeling inspired. We asked participants about their perceived cognitive stimulation, adapting a 6-item instrument by Gozzo et al. 2022 (α = 0.723). We did not find correlations between self-reported cognitive stimulation and human_fluency, human_flexibility, human_novelty, or human_value to be significant. Thus, potentially, participants did not actually experience cognitive stimulation but just felt to more efficiently and holistically cover the topic as a (human–GLM) dyad. Indeed, there were significant correlations of reported cognitive stimulation with dyad_fluency (Spearman’s ρ = 0.436, p = 0.007) and dyad_flexibility (r = 0.402, p = 0.014). From many of the participants’ answers, it was unclear if participants felt that the GLM ideas actually stimulated them or if these GLM ideas were “just” used as additional ideas for the long list (sic!):

Similarly, responses regarding GLM ideas’ flexibility were mixed. Some participants stressed that the GLM helped them to gain a new perspective, which could hint at the AI ideas successfully broadening the breadth of exploration and reducing cognitive inertia:

However, as reported above, while we observed increased flexibility on the dyad level (dyad_flexibility), we did not observe a significant difference on the individual level (human_flexibility). Indeed, several participants perceived GLM ideas as repetitive (e.g., P44, P45, P48), with some stating they felt “governed” or their focus was set in a certain direction, which some felt reduced (“cutting off”, P69) or “took away” (P42) their creativity or concentration:

This suggests that GLM ideas may risk narrowing ideation or reinforcing existing idea paths (cf. Zhan et al. 2024). Although fixation was not universal, this tension between stimulation and narrowing may help explain why individual performance effects were muted, even when some participants reported subjective feelings of stimulation and results surfaced dyad-level gains.

Reduced effort and disengagement: Regarding free riding, qualitative responses indicated that some participants reduced their effort or became disengaged, for example, stating:

Dispensability of effort: A contributing factor for free riding is felt dispensability of effort (Pinsonneault et al. 1999), i.e., the feeling that (additional) own effort would not (substantially) improve results. In line with recent literature finding that GLMs can produce creative ideas comparable to many humans (Haase and Hanel 2023), several participants expressed a sense of inferiority compared to the GLM when it came to generating ideas, e.g., describing themselves as “not that creative” (P60) or stating that GLM ideas were “overall better defined and closer to the actual topic” (P42). However, other participants described the GLM ideas as of “lower quality and more repetitive” (P44), “uncreative and boring” (P51), “too mundane and common” (P75), or as “generic answers already are established in a way” (P73) that would “create different problems” (P53), and stating that their “ideas are unique and make more sense to [them] than the ‘cheap’ AI suggestions” (P49). These mixed results may hint at underlying individual differences for working with GLMs in creative tasks (cf. Memmert et al. 2024b).

Diffused responsibility: An additional aspect related to free-riding is diffused responsibility (Pinsonneault et al. 1999; Latané et al. 1979). In the post-survey, we asked participants about the degree of responsibility they felt for the longlist and shortlist of ideas. We find that none of the participants claimed full responsibility for the long list, and only 6 participants (16.2%) for the shortlist of ideas. We found a strong correlation between the felt responsibility and the share of ideas originating from the human in the long list (r = 0.710, p < 0.001). The same holds true for the share of human ideas on and felt responsibility for the shortlist (r = 0.489, p = 0.002). This is also reflected in the qualitative responses, where many participants provided a rationale based on the contribution share for the level of felt responsibility, e.g., stating “I shared a low responsibility because I choose to include comparatively many ideas from the AI.” (P44).

Effort shift and cognitive load: Tying our findings back to theory, smart loafing refers to “the reduction of effort […] to maintain cognitive resources and enhance efficiency in work” (Stieglitz et al. 2022). While some participants reported having reduced their effort as described above, other participants offered more nuanced responses (sic!):

Rather than an effort reduction, these responses signal an effort shift from generating to curating ideas. This effort shift seemed to have been both a source of feelings of indispensability and non-diffused responsibility according to some participants, e.g., stating “[…] I think I wouldnt really relie on AI ranking” (sic!, P47) for the former, and “I evaluated all of the ideas on my own an[d] picked the what i thought were the best ideas” (P68) for the latter. When asked about how to achieve the best brainstorming performance with our app, most participants (n = 30) selected a mix between generating and curating ideas (see Fig. 10). As participants in the human + GLM condition spent time reviewing and selecting GLM ideas, this might explain their lower fluency. Indeed, the number of suggestion requests and human_fluency is negatively correlated (r = −0.419, p = 0.01).

Full size image

This effort shift may explain why we did not find a significant difference in cognitive load, contrary to expectations under the smart-loafing assumption. Though several participants reported to have felt working with the AI made the task “easier” (e.g., P41, P42), “less mentally straining” (P62), and “took the stress away” (P59), many reported to having to read and evaluate GLM ideas, explaining “I had to reject the ai suggestions and pay attention not to be influenced too much by it or to be bogged down in the ai's way of thinking” (P75). While not explicitly stated by many as additional cognitive load, this constant evaluation effort might have offset the load reduction due to the GLM contributing ideas (cf. Tankelevitch et al. 2024). Thus, while working with GLMs can offer performance improvements, there are also risks associated.

A key goal for human–AI collaboration is for the human–AI group to achieve superior performance compared to the human; that is, working with the AI system should yield better results than working alone (Dellermann et al. 2019). Similarly, Hemmer et al. (2021, p. 2) and Hemmer et al. (2025) describe the desired outcome for a human and an AI – which they label team members – as working together to achieve “complementary team performance (CTP),” defining it as “the team performance exceed[ing] the maximum performance of both individual entities.” AI was traditionally discussed as being helpful for (automating) repetitive tasks (Krogh 2018), whereas creative tasks were thought to be in the realm of humans (Redifer et al. 2021; Dellermann et al. 2019), particularly as AI systems’ output was thought to potentially lack diversity (Bender et al. 2021). After the creativity of GLMs and humans was investigated in isolation (Organisciak et al. 2023; Haase and Hanel 2023; Boussioux et al. 2023), we studied joint human–AI brainstorming, as has been called for in the literature (Di Fede et al. 2022; Liu et al. 2023; Rafner et al. 2023; Wu et al. 2021). In empirically demonstrating the superiority of the human–AI dyad in brainstorming, i.e., achievement of complementary performance (Hemmer et al. 2021, 2025) of the human–AI dyad compared to the individual human, we confirm what was hypothesized in the literature (Boussioux et al. 2023). Our results partially confirm earlier findings, such as those showing heightened fluency and flexibility when brainstorming with AI (Demir et al. 2024; Specker et al. 2025), but contradict other findings, e.g., those indicating reduced novelty (Demir et al. 2024; Kumar et al. 2025). Overall, the results align with the recent meta-analysis by Holzner et al. (2025), which suggests that human–AI dyads outperform solitary humans in creative performance, particularly when participants are non-domain experts and the AI is a capable GLM like GPT-3.5. However, our study offers a more nuanced analysis, distinguishing among four brainstorming metrics rather than using a single aggregated measure of creative performance. Holzner et al. (2025) report that AI reduces the diversity of ideas, whereas our study shows increased flexibility of human–AI dyads as compared to solitary humans. We find that human–AI dyads cover more categories of ideas compared to solitary humans. This may hint at the trade-off that although individuals may benefit from AI usage to increase the human–AI dyad’s performance, on an aggregated level – when comparing pooled ideas of solitary humans and pooled ideas of human–AI dyads – results may become more homogeneous (cp. Dell’Acqua et al. 2023). Future research should explore the differences in measurement approaches, for example, by comparing flexibility through assessing the number of categories covered for a topic (as in our study) with statistical approaches, such as embedding-based measures like cosine similarity (e.g., Dell’Acqua et al. 2023; Holzner et al. 2025).

Moreover, future research should extend performance benchmarking to include both human–human dyads and AI-only conditions to provide a more comprehensive comparison of creative performance across different collaboration types. Including a human–human condition would enable disentangling the effects of GLM support from the general effects of working with another idea-contributing partner. While we deliberately omitted a human–human group to isolate the cognitive and behavioral mechanisms of human–AI interaction, thereby avoiding human social dynamics, future work could systematically contrast these settings. Such comparisons would directly respond to calls for studying human–AI complementarity (Hemmer et al. 2025; Dellermann et al. 2019) and to recent findings that challenge the notion that human input necessarily has a positive performance impact (Lee and Chung 2024; Vaccaro et al. 2024).

Intuitively, the results may not come as a surprise – tool support leads to better performance. However, our analysis shows that when excluding GLM ideas, humans working with GLMs themselves do not perform better compared to solitary brainstormers; instead, they perform worse. This is contrary to expectations for traditional stimuli-based ICSS (Wang and Nickerson 2017). After all, stimulating the human to create more and more novel ideas is the goal of a stimuli-provider ICSS; i.e., ICSS-induced cognitive stimulation should manifest in higher individual brainstorming performance (Pinsonneault et al. 1999).

Siangliulue et al. (2015b, p. 88) demonstrated that stimuli-based ICSS are most effective when humans request stimuli themselves (i.e., pull; as in our app), and they showed that “participants primarily requested examples when they ran out of ideas.” While there is no agreed-upon duration for brainstorming, durations in experiments between 4, 10 (this study), and 15 min are common (Siangliulue et al. 2015a; Michinov 2012; Pinsonneault et al. 1999; Dennis et al. 2013), and past research has shown that fluency rapidly declines after the first few minutes (Kohn and Smith 2011; Baruah and Paulus 2016), with participants adopting more GLM ideas than producing ideas themselves after about 3–4 min (Memmert et al. 2024b). Given those empirical timings, participants in our study had sufficient time to both contribute their own ideas and interact with the ICSS when running out of ideas.

In our study, the ICSS provides stimuli in the form of full ideas rather than only cues, the latter of which is common for prior stimuli-provider ICSS. In that, working with our tool might be somewhat similar to brainstorming with another human, who also contributes full ideas (see Fig. 1). According to the literature, both incomplete ideas from ICSS and complete ideas from other humans can induce cognitive stimulation and consequently improve human brainstorming performance (Pinsonneault et al. 1999; Wang and Nickerson 2017). However, contrary to those expectations, here, external stimuli (i.e., GLM ideas) did not increase humans’ brainstorming performance, but rather resulted in the humans generating fewer ideas themselves. Thus, while both the traditional and the GLM-based ICSS resulted in better collective group-level performance, different mechanisms seem to drive these improvements.

What seems interesting is that despite participants objectively not producing more (novel) ideas when working with the GLM, they – in line with prior research (Specker et al. 2025) – subjectively reported feeling stimulated or inspired, which could be rooted in the new possibility that brainstorming with a GLM provides. Unlike with traditional ICSS, which require the human to develop a full idea from the stimuli, with our app providing complete, creative ideas – GLM ideas were rated as more novel than human ideas – these GLM ideas can simply be copied by the human without additional effort. While this is similar to brainstorming with another human, participants did not have to credit another person for ideas they had not developed themselves. Hence, humans potentially felt subjectively inspired and indeed perceived themselves as having covered the topic more broadly because they subconsciously attributed the GLM ideas to themselves. This perspective is compatible with the theory of the “extended self” (Mirbabaie et al. 2021) and with viewing GLM-based ICSS as merely tools.

Perhaps, this reduced the drive of humans to come up with more ideas themselves based on the GLM's ideas, which might mean that the dyad-level performance improvement is mainly driven by the GLM’s ideas, rather than by improved human performance. This raises the question of whether “collaborative complementarity potential” (Hemmer et al. 2025, p. 7) is realized – that is, not merely that the dyad outperforms either entity alone, but that their interaction yields performance exceeding the additive combination of their individual contributions – or, put simply, whether the whole exceeds its parts.

Another potential explanation for us not observing higher performance with increased cognitive stimulation could be that the performance-enhancing and -reducing effects canceled each other out. Liu et al. (2023) raised the question, “Will humans be Free-Riders?”. Indeed, we observe humans working with the GLM to contribute fewer ideas (hypothesis set 1), and some admitted to having become “lazy,” suggesting that free riding might have occurred. A cause for free riding is the phenomenon of diffused responsibility. Indeed, no participant felt full responsibility for the long list of ideas, and only a few for the shortlist. Such responsibility attribution to non-human group members has been observed before (Stieglitz et al. 2022). Thus, one might assume that humans reduced their effort, i.e., free rode (Latané et al. 1979). However, on a group level, human + GLM dyads performed better (hypothesis set 2), so task performance was not harmed. We thus concur with Stieglitz et al. (2022) that free riding or social loafing may need to be conceptualized differently in human–AI groups, as it may sometimes be desirable, particularly if it does not lead to a reduction in the quality of the result. Below, we reflect on their concept of smart loafing, i.e., humans' effort reduction to maintain cognitive resources while enhancing work efficiency (Stieglitz et al. 2022). We discuss the key parts of the definition below.

Though perhaps occasionally, overall, it is not apparent that participants reduced their effort when working with the GLM. Humans in the human-only condition focused exclusively on generating ideas, whereas humans in the human + GLM condition split their effort between generating and curating (requesting, evaluating, and adjusting) ideas, implying a difference in roles from pure creator to creator and curator (cf. Gonzalez et al. 2024; Tankelevitch et al. 2024). Perhaps there was not a reduction in effort but a shift in effort allocation.

Maintaining cognitive resources could mean conserving them when performing the task due to the AI. We would then expect lower cognitive load levels for humans working with the GLM. It could also mean improving performance while not increasing cognitive load, i.e., enhancing work efficiency. Work efficiency enhancement was included in the original definition (Stieglitz et al. 2022) and was observed in our study, i.e., with similar cognitive load, a higher collective task performance was achieved by the human–GLM dyad. We did not observe lower cognitive load, i.e., humans did not save cognitive resources when working with the GLM (hypothesis 3). A potential explanation is that there was a shift instead of a reduction in cognitive load. According to cognitive load theory, cognitive load comprises intrinsic, extraneous, and germane components (Sweller 1988; Paas et al. 2003). While the intrinsic cognitive load for generating own ideas might have decreased, an additional intrinsic cognitive load for curating ideas might have occurred (Fig. 11). Confirming earlier reports of such behavior (Gonzalez et al. 2024), many participants reported having reviewed the GLM ideas before adopting them. These findings align with prior research suggesting that GLMs may introduce additional cognitive demands, for example, for the evaluation of AI outputs (Tankelevitch et al. 2024). Moreover, additional extrinsic cognitive load due to the presentation of the GLM ideas may have occurred, and germane cognitive load – the mental effort required for constructing, refining, or integrating knowledge structures or schemas (Paas et al. 2003, p. 2) – may have been maintained. Together, this may explain why we did not observe a reduction in cognitive load.

Full size image

Future research might differentiate cognitive load types (Cheng et al. 2020; Javadi et al. 2013) and explore if GLM ideas can be controlled to balance the positive effects of offering ideas and the negative effects of inducing cognitive load, e.g., by controlling idea relatedness (Baruah and Paulus 2016; Wang and Nickerson 2019, 2017), e.g., through prompt design or filtering (Summers-Stay et al. 2023) of GLM ideas, or by delivery mechanisms (Siangliulue et al. 2015b).

Smart loafing implies a deliberate and intelligent approach to loafing. This may require reflection on one’s own and the GLM's comparative task-relevant capabilities, such as creative ability or topic knowledge, to ultimately achieve a symbiotic collaboration and a productive division of work (Benbya et al. 2024; Boussioux et al. 2023; Vaccaro et al. 2024). We found signs of this reflection in some qualitative responses. In the literature, various roles have been proposed for AI in human–AI work (Siemon 2022a). Perhaps similarly interesting is the human’s role expansion from creator to curator (Memmert et al. 2025). It might be fruitful to structure human roles and support individuals in taking a good role or effort split, i.e., striking a balance between creating and curating in open-ended settings, enabling them to loaf “smartly”.

A related aspect is the attribution of responsibility in human–AI settings. As in prior studies (Stieglitz et al. 2022), participants did not attribute full responsibility to themselves despite being the only human. The share of human ideas was strongly positively correlated with felt responsibility. While causation is unclear, this might hint at humans feeling less responsible the more the AI contributes, even when in control of result curation. With GLMs becoming more capable, this may introduce problems. Future research should explore the rationale for responsibility attribution, such as the perception of the AI on the spectrum from tool to colleague, to inform interventions for calibrating felt responsibility.

Smart loafing was suggested to be desirable for repetitive or overhead tasks (Stieglitz et al. 2022). We show that it might be desirable even for one-off tasks, such as brainstorming for a specific problem, and thus agree with the call for extending the focus of smart loafing beyond its initial, more narrow scope towards other, more cognitively challenging tasks.