How and For Whom Using Generative AI Affects Creativity: A Field Experiment

Abstract

We develop a theoretical perspective on how and for whom large language model (LLM) assistance influences creativity in the workplace. We propose that LLM assistance increases employees’ creativity by providing cognitive job resources. Furthermore, we hypothesize that employees with high levels of metacognitive strategies—who actively monitor and regulate their thinking to achieve goals and solve problems—are more likely to leverage LLM assistance effectively to acquire cognitive job resources, thereby increasing creativity. Our hypotheses were supported by a field experiment, in which we randomly assigned employees in a technology consulting firm to either receive LLM assistance or not. The results are robust across both supervisor and external evaluator ratings of employee creativity. Our findings indicate that LLM assistance enhances employees’ creativity by providing cognitive job resources, especially for employees with high (vs. low) levels of metacognitive strategies. Overall, our field experiment offers novel insights into the mediating and moderating mechanisms linking LLM assistance and employee creativity in the workplace.

Full text

How and For Whom Using Generative AI Affects Creativity: A Field Experiment
in press, Journal of Applied Psychology
Shuhua Sun
A.B. Freeman School of Business, Tulane University
Angelina Zhuyi Li
School of Business, Renmin University of China
Maw Der Foo
Nanyang Business School, Nanyang Technological University
Jing Zhou
Jesse H. Jones Graduate School of Business, Rice University
Jackson G. Lu
MIT Sloan School of Management, Massachusetts Institute of Technology
Author note: Correspondence concerning this article can be addressed to Shuhua Sun
(ssun7@tulane.edu), Angelina Zhuyi Li (lizhuyi@rmbs.ruc.edu.cn), and Jackson G. Lu
(lu18@mit.edu). The first two authors contributed equally.
Author contributions: Shuhua Sun played a lead role in conceptualization, formal analysis,
methodology, software, supervision, validation, visualization, resources, writing–original draft,
and writing–review and editing. Angelina Zhuyi Li played a lead role in data curation, funding
acquisition, investigation, methodology, project administration, and resources and a supporting
role in formal analysis and validation. Maw Der Foo played a lead role in methodology and a
supporting role in writing–review and editing. Jing Zhou played a lead role in methodology and a
supporting role in conceptualization and writing–review and editing. Jackson G. Lu played a lead
role in visualization and writing–review and editing and a supporting role in conceptualization,
methodology, resources, software, supervision, and validation.
Acknowledgments: We thank Huaizhong Chen, Jane Minyan Chen, Jiaqi Li, Sophia Yat-Mei
Suen, and Anna Manyi Zheng for their helpful feedback on earlier versions of the paper.

GENERATIVE AI USE AND CREATIVITY
How and For Whom Using Generative AI Affects Creativity: A Field Experiment
Abstract
We develop a theoretical perspective on how and for whom large language model (LLM)
assistance influences creativity in the workplace. We propose that LLM assistance increases
employees’ creativity by providing cognitive job resources. Furthermore, we hypothesize that
employees with high levels of metacognitive strategies—who actively monitor and regulate their
thinking to achieve goals and solve problems—are more likely to leverage LLM assistance
effectively to acquire cognitive job resources, thereby increasing creativity. Our hypotheses were
supported by a field experiment, in which we randomly assigned employees in a technology
consulting firm to either receive LLM assistance or not. The results are robust across both
supervisor and external evaluator ratings of employee creativity. Our findings indicate that LLM
assistance enhances employees’ creativity by providing cognitive job resources, especially for
employees with high (vs. low) levels of metacognitive strategies. Overall, our field experiment
offers novel insights into the mediating and moderating mechanisms linking LLM assistance and
employee creativity in the workplace.
Keywords: generative artificial intelligence, large language models, creativity, metacognitive
strategies, cognitive job resources.

1
GENERATIVE AI USE AND CREATIVITY
In November 2022, OpenAI launched ChatGPT, a large language model (LLM)-based
generative artificial intelligence (AI) that marked a new era in AI technology. LLMs are
considered general-purpose technologies akin to steam engines because of their potential to
transform how work is done (Eloundou et al., 2023). Corporate investment in LLM tools is
surging (Tiwari, 2024), with companies aiming to leverage them to boost employee creativity
(Ivcevic & Grandinetti, 2024), defined as the generation of novel and useful ideas (Amabile,
1988). However, a large-scale, nationally representative survey conducted by Gallup found that
only 26% of employees using LLM tools report improved creativity (Houter, 2024). This raises
an important question: Do LLMs increase employee creativity in the workplace?
This question demands rigorous research attention for three key reasons. First, employee
creativity is crucial for organizational success, as it drives innovation and enables companies to
adapt to changing market demands. By generating creative ideas, employees contribute to the
development of new products, services, and processes, ultimately strengthening an organization’s
competitive advantage (Anderson et al., 2014). Second, organizations are deploying LLMs to
boost their competitive edge, with an expectation that they will help enhance employee creativity
(Singla et al., 2024). It is, therefore, important to understand the actual impact of LLMs on
employee creativity (Ivcevic & Grandinetti, 2024). Third, a review of the existing literature
reveals little direct evidence about the impact of LLMs on employee creativity in real-world
organizational settings. Most studies on LLMs and creativity were conducted in online and lab
settings, and almost all of them focused on isolated, single tasks (e.g., Anderson et al., 2024;
Boussioux et al., 2024; Chen & Chan, 2024; Doshi & Hauser, 2023; Góes et al., 2023; Hitsuwari
et al., 2023). It remains unclear whether LLMs boost employee creativity in actual workplaces
where employees juggle multiple tasks and make complex decisions (Nathan, 2024).
Why might LLMs foster employee creativity? To answer this question, we draw on the
cognitive approach to creativity (Amabile, 1988; de Jonge et al., 2012; Zhou & Shalley, 2011).

2
GENERATIVE AI USE AND CREATIVITY
According to this perspective, creativity is inherently cognitive, requiring individuals to search
within and across knowledge domains to gather diverse information, integrate ideas from
multiple sources, and experiment with new ways of approaching tasks (Zhou & Shalley, 2011).
From this cognitive viewpoint, cognitive job resources—defined as resources essential for
addressing the cognitive demands of work—play a central role in fostering creativity (de Jonge
et al., 2012; Hargadon, 2002; Niks et al., 2017; Zhou & Shalley, 2011). Cognitive job resources
include information and knowledge as well as “the opportunity to determine a variety of task
aspects and to use problem-solving skills”—opportunities to adjust work methods and tasks such
as switching between complex and simple tasks and taking mental breaks (de Jonge et al., 2012,
p. 328). Information and knowledge are essential for creativity because creativity fundamentally
involves recombining and synthesizing information in novel and useful ways (Fleming et al.,
2007; Hargadon, 2002). Similarly, opportunities to switch between tasks and take mental breaks
are vital for creativity because they allow employees to break fixed mindsets and restore
cognitive capacity (Elsbach & Hargadon, 2006; Madjar & Shalley, 2008). Overall, cognitive job
resources provide employees with “room to think about existing problems and to develop new
and innovative ways of how to handle the cognitive job demands” (de Jonge et al., 2012, p. 326).
We thus propose that LLMs can enhance employee creativity by increasing cognitive job
resources, as LLMs are capable of providing information and knowledge and assisting with
various tasks, which would allow employees to switch between complex and simple tasks and
take mental breaks as needed (Acemoglu, 2024; Brynjolfsson et al., 2025; Zhao et al., 2023).
However, this proposition hinges on an assumption that employees can effectively
leverage LLMs to acquire these cognitive job resources. To deepen our theorizing, we draw on
the metacognition research to propose that the cognitive job resources mechanism is more likely
to hold for employees with high levels of metacognitive strategies. Metacognition research
emphasizes that successfully utilizing tools to acquire task-related resources depends on

3
GENERATIVE AI USE AND CREATIVITY
individuals’ metacognitive strategies, which involve “actively analyzing tasks and then planning,
self-monitoring, and revising strategies” (Chen et al., 2020, p. 14066; Flavell, 1979; Veenman et
al., 2004). Specifically, by continuously evaluating tasks and tracking the effectiveness of
strategies, individuals can better determine what information and knowledge they need, as well
as when to switch tasks or take mental breaks to restore cognitive capacity and break cognitive
fixation (Davidson & Sternberg, 1998; Sun, 2024; Winne & Nesbit, 2010). Consequently,
employees with higher (vs. lower) levels of metacognitive strategies are more likely to leverage
LLMs effectively to acquire cognitive job resources, thereby increasing creativity.
In sum, we hypothesize that for LLMs to enhance employee creativity, employees need
high levels of metacognitive strategies to effectively utilize LLMs to acquire the cognitive job
resources necessary for generating creative ideas. Figure 1 illustrates our conceptual model. To
test our hypotheses, we conducted a field experiment to compare employees’ creativity—
assessed with both supervisor and external evaluator ratings—across conditions with and without
LLM assistance.
This research offers important empirical, theoretical, and practical contributions. First, by
unpacking the mediating and moderating mechanisms that underlie the impact of LLMs on
employee creativity in a field experiment, we advance the theoretical and empirical
understanding of how and for whom LLMs can enhance creativity in organizational settings. Our
findings show that LLMs enhance employees’ creativity by providing cognitive job resources,
especially for employees who possess high levels of metacognitive strategies. This suggests that
to fully benefit from LLM use, employees must actively adopt metacognitive strategies—
analyzing tasks and their own thought processes, planning, self-monitoring, and revising
strategies (Chen et al., 2020; Flavell, 1979)—rather than being passive consumers of LLMs.
Second, our research contributes to the literature on cognitive job resources by
demonstrating how cognitive job resources can be enhanced through the effective use of LLMs.

4
GENERATIVE AI USE AND CREATIVITY
While research has highlighted the positive impact of cognitive job resources on creativity
(Amabile & Gryskiewicz, 1987; de Jonge et al., 2012; Elsbach & Hargadon, 2006; Lu et al.,
2017; Niks et al., 2017), little attention has been given to the antecedents of cognitive job
resources and how these resources can be enhanced through interventions in the workplace.
Third, our study extends the literature on metacognitive strategies by examining how they
enable employees to leverage LLMs to enhance cognitive job resources and, in turn, creativity at
work. Prior research has mainly examined metacognitive strategies in educational and non-work
contexts (for reviews, see Sun, 2024). By demonstrating their value in organizational settings,
our findings suggest that organizations should consider employees’ metacognitive abilities when
deploying LLMs. Even the most advanced LLMs may not boost creativity if employees lack the
metacognitive strategies needed to use them effectively. This insight is practically significant, as
metacognitive strategies—though often viewed as an individual difference—can be developed
through targeted training (Bell & Kozlowski, 2008; Sun, 2024).
Theory and Hypothesis Development
Effect of LLM Assistance on Creativity: The Mediating Role of Cognitive Job Resources
The cognitive approach to creativity underscores the cognitive nature of creativity and the
crucial role of cognitive job resources in fostering creativity (Amabile, 1988; de Jonge et al.,
2012; Zhou & Shalley, 2011). As Niks et al. (2017, p. 185) summarized, “Only if there are
sufficient cognitive resources (such as access to useful information), there is room for thinking
about problems and developing new ideas about how to deal with the job demands.” Therefore,
for LLMs to increase employee creativity, they must be able to provide the cognitive job
resources needed to generate creative ideas. Cognitive job resources involve information and
knowledge (as measured by items such as “I have access to useful information to help solve
complex tasks,” van den Tooren & de Jonge, 2010, p. 40) and “the opportunity to determine a
variety of task aspects and to use problem-solving skills” (de Jonge et al., 2012, p. 328; as

5
GENERATIVE AI USE AND CREATIVITY
measured by items such as “I have the opportunity to switch between simple and complex
tasks”). Below, we review the benefits of these cognitive job resources on employee creativity
and theorize how LLMs can increase each of these cognitive job resources.
LLMs can help employees access and process a wide range of information and
knowledge to generate creative ideas. Information and knowledge are the central ingredients for
creativity, as creativity is fundamentally about knowledge recombination (Hargadon, 2002).
Extensive research highlights the critical role of both a large number of and a diverse body of
knowledge bases in fostering creativity (Fleming, 2001; Mumford & Gustafson, 1988).
Individuals with access to a large number of knowledge components tend to produce novel ideas
because they can experiment with different combinations and recombinations of knowledge
(Oldham & Cummings, 1996; Reiter-Palmon & Arreola, 2015). Likewise, individuals with
access to diverse knowledge domains tend to produce creative ideas through the uncommon
recombination of distinct knowledge bases (Hargadon & Sutton, 1997; Leahey et al., 2017).
Acquiring information and knowledge, however, is a costly process because of what
Jones (2009) labeled the “knowledge burden,” or the underlying tension between the limited time
and cognitive abilities an individual has and the large amount of knowledge the individual needs
to acquire (Teodoridis et al., 2019). For instance, searching for information and knowledge
requires significant time and effort, as does processing and absorbing them (Leahey et al., 2017).
Additionally, to access knowledge outside one’s primary domain, an individual needs to invest
considerable time and resources in building social networks with people from different areas of
expertise who offer diverse perspectives and specialized knowledge that may not be available
within an individual’s immediate work environment (Perry-Smith, 2006).
LLMs can complement employees by helping them access a diverse array of knowledge
bases because LLMs are trained on large corpora of data (e.g., articles, books, and websites) and
can summarize and explain information and knowledge in accessible terms (Zhao et al., 2023).

6
GENERATIVE AI USE AND CREATIVITY
Importantly, LLMs can perform these functions—knowledge searching, summarization, and
elaboration—almost instantaneously (Bubeck et al., 2023). As such, considerable time and
resources are conserved, which can be used to experiment with and mull over ideas to solve
problems creatively. Moreover, LLMs’ efficient processing and elaboration of diverse
information reduces the chances of “combinatorial exhaustion”—a situation where novel
knowledge recombination within a set of knowledge bases is exhausted—that all humans face
due to finite knowledge and limited time (Fleming, 2001; Teodoridis et al., 2019).
In addition to information and knowledge, LLM assistance can provide employees with
greater opportunities to adjust work methods and tasks—such as switching between complex and
simple tasks and taking mental breaks, which can foster creativity (de Jonge et al., 2012; Elsbach
& Hargadon, 2006). As Elsbach and Hargadon (2006) highlighted in their model of workday
design, to enhance creativity, employees need to alternate between complex and simple tasks
throughout the workday. While creativity often requires challenging, complex tasks that
stimulate the creative process (Zhou & Shalley, 2003), constantly working on complex tasks can
be cognitively straining and ultimately hinder creativity (de Jonge & Dormann, 2006; Sonnentag
et al., 2010; Sun et al., 2020). The opportunity to alternate between complex and simple tasks
enables employees to focus on complex problem-solving while using simpler tasks and mental
breaks to restore cognitive capacity and shift away from fixed mindsets (Elsbach & Hargadon,
2006; Lu et al., 2017; Madjar & Shalley, 2008). As a result, creative insights are more likely to
“spring to mind” when individuals have the opportunity to switch tasks and take mental breaks
after periods of focused concentration (Beeftink et al., 2008; Smith, 1995).
Yet, in fast-paced professional environments, employees often have limited opportunities
to switch tasks or take breaks (Elsbach & Hargadon, 2006; Pfeffer, 2018). LLMs can expand
these opportunities by assisting with a wide range of tasks (Eloundou et al., 2023). For example,
employees can delegate routine, repetitive work to LLMs to free up resources for complex

7
GENERATIVE AI USE AND CREATIVITY
problem-solving (Davenport & Kirby, 2016). Common tasks such as summarizing text,
managing data, and drafting content are well within LLMs’ capabilities (Acemoglu, 2024).
Moreover, employees can use LLMs for support with complex, cognitively demanding tasks
while periodically shifting to simpler ones, allowing them to restore mental capacity and break
fixed mindsets. In this way, LLMs’ demonstrated capabilities in handling complex, knowledge-
intensive tasks (Bubeck et al., 2023; Zhao et al., 2023) make them valuable tools for reducing
cognitive overload and optimizing task management.
In sum, we propose that employees can enhance their creativity through LLM assistance
by acquiring cognitive job resources. Accordingly, we hypothesize:
Hypothesis 1: Employees with LLM assistance gain more cognitive job resources than
employees without LLM assistance.
Hypothesis 2: Employees with LLM assistance exhibit higher levels of creativity than
employees without LLM assistance.
Hypothesis 3: Cognitive job resources mediate the relationship between LLM assistance
and employees’ creativity.
Boundary Condition: The Moderating Role of Metacognitive Strategies
However, access to LLMs does not guarantee that employees can acquire cognitive job
resources from their use. Drawing on metacognition research, we propose that metacognitive
strategies moderate the relationship between LLM use and the acquisition of cognitive job
resources. Metacognitive strategies involve actively monitoring and regulating one’s thinking to
complete tasks and achieve goals (Chen et al., 2020; Flavell, 1979; Sun, 2024). Examples
include “thinking through the steps one needs to take to perform tasks,” “keeping track of how
effective one’s approach is,” and “reassessing one’s approach when noticing a lack of progress”
(Chen et al., 2020). Through ongoing evaluation of task demands and strategy effectiveness,
individuals become more aware of task difficulty, knowledge gaps, and mental states (Flavell,
1979; Sun, 2024). This awareness enables them to identify needed information and know when
to switch tasks or take breaks to restore cognitive capacity and break rigid thinking (Davidson &

8
GENERATIVE AI USE AND CREATIVITY
Sternberg, 1998; Winne & Nesbit, 2010). Consequently, employees with higher (vs. lower) levels
of metacognitive strategies are better equipped to leverage LLMs to acquire cognitive job
resources that enhance creativity. Below, we elaborate on the moderating role of metacognitive
strategies.
To begin with, employees with high levels of metacognitive strategies can effectively
utilize LLMs to acquire helpful information and knowledge that facilitate creative problem-
solving. These employees actively monitor and evaluate their tasks and cognitive processes,
allowing them to recognize their thinking and knowledge gaps (Davidson & Sternberg, 1998).
By recognizing these gaps, they can conduct targeted information searches (McCormick, 2003).
For instance, by assessing the effectiveness of their problem-solving, they can iteratively refine
LLM prompts to retrieve more relevant and precise information. This adaptive approach
enhances their ability to generate deeper insights and more creative solutions. In contrast, if
employees lack the metacognitive strategies to monitor tasks and thinking processes, they will
lack an awareness of their knowledge gaps (Flavell, 1979). As a result, they may fail to leverage
LLMs to acquire relevant information, limiting their creative problem-solving.
Besides acquiring information and knowledge, employees with strong metacognitive
strategies can effectively leverage LLMs to adjust work methods and tasks—such as switching
between complex and simple tasks and taking mental breaks—which foster creativity (Beeftink
et al., 2008; Elsbach & Hargadon, 2006). Metacognitive strategies involve analyzing tasks and
reflecting on one’s problem-solving strengths and weaknesses (Chen et al., 2020; Veenman et al.,
2004). Employees with high levels of metacognitive strategies keep track of which tasks are
better suited for them so that they can delegate others to LLMs (Davenport & Kirby, 2016),
freeing resources for in-depth problem-solving and idea generation. Furthermore, metacognitive
strategies help employees monitor their cognitive load during cognitively demanding tasks (Sun,
2024; Winne & Nesbit, 2010). This awareness enables them to strategically offload work to

9
GENERATIVE AI USE AND CREATIVITY
LLMs, creating opportunities for mental breaks or transitions to simpler tasks, which help restore
cognitive capacity and prevent mental fixation (Elsbach & Hargadon, 2006; Lu et al., 2017).
In sum, metacognitive strategies equip individuals with continuous monitoring and
assessment of task demands and problem-solving approaches, enabling them to engage with
LLMs effectively to obtain cognitive resources that enhance creativity. Hence, we hypothesize:
Hypothesis 4: The acquisition of cognitive job resources through LLM assistance
depends on an employee’s metacognitive strategies. Specifically, employees with higher
(vs. lower) levels of metacognitive strategies are more (vs. less) likely to obtain cognitive
job resources from LLM assistance.
Hypothesis 5: The mediated effect of LLM assistance on employee creativity via
cognitive job resources is moderated by employees’ metacognitive strategies.
Specifically, the mediated effect is expected to be stronger (vs. weaker) for employees
with higher (vs. lower) levels of metacognitive strategies.
Methods
Transparency and Openness
The experiment was approved by Renmin University of China (protocol #2023R19). We
describe our sampling plan, data exclusions (if any), manipulations, and measures, and adhere to
the Journal of Applied Psychology methodological checklist. This study is not preregistered.
Data were analyzed using Stata 16. While we are unable to publicly share the data because of a
confidentiality agreement with the firm, the data and materials are available upon request.
Empirical Setting
We conducted our field experiment in a technology consulting firm in China. The
company is attuned to technological innovations and, at the time of this research (August 2023),
had already established a research unit experimenting with OpenAI’s application programming
interface (API). It is worth noting that since its initial release, ChatGPT has not been directly
available for use in China. However, AI developers in China could access ChatGPT through
OpenAI’s API service until the service was suspended in 2024 (Reuters, 2024). At the time of our
research, the firm’s research team had developed an API-based interface and was preparing for
internal deployment. We thus exploited this opportunity to implement our randomized field

10
GENERATIVE AI USE AND CREATIVITY
experiment. This consulting firm was also an ideal context for studying the impact of LLMs on
employee creativity: Creativity is highly valued in consulting where employees need to generate
original ideas and develop customized solutions for diverse clients (Lu, 2024; Unsworth, 2001).
Participants and Procedures
All non-managerial employees, except those experimenting with ChatGPT API service,
were invited to participate, yielding a pool of 286 eligible employees across three departments:
technology, sales/consulting, and administration. The study proceeded in three main steps. First,
eligible employees were invited to attend information sessions. Participants were informed that
the study concerned work and work-related behaviors and that they would receive 100 RMB as a
token of appreciation for completing two surveys. After completing consent procedures, we
distributed the initial survey, covering demographics and job-related questions. Due to business
travel and illnesses, 36 employees did not participate, resulting in a final sample of 250
employees. Among them, 64.8% were male, with an average age of 29.59 years (SD = 4.37);
66.4% held a bachelor’s degree, 32.4% a master’s degree, and 1.2% a doctoral degree.
Next, on August 7, 2023, participants were randomly assigned to either the treatment or
control group using a number generator tool. Employees in the treatment group received
ChatGPT accounts with usage examples and were instructed that the accounts were for personal
use only and not to be shared or discussed with others. To alleviate potential job security
concerns, the company informed employees in the treatment group that ChatGPT was intended
to assist—not replace—their roles (Yam, Tang, et al., 2023).
Finally, on August 15, 2023, all participants were invited to answer a second survey that
measured the mediating and moderating variables as well as several attitudinal and motivational
control measures. Additionally, we invited the employees’ direct supervisors and two external
raters to evaluate each employee’s creativity. Both the supervisors and external raters were blind
to our research hypotheses and experimental design. Since our surveys used well-established

11
GENERATIVE AI USE AND CREATIVITY
measures originally in English, three bilingual researchers performed translation and back-
translation procedures, cross-checking for accuracy (Brislin, 1970).
Measures
Experimental condition and manipulation checks. To verify that employees in the
experimental group used ChatGPT and employees in the control group did not, we collected self-
reported measures of ChatGPT usage (0 = no, 1 = yes) and usage frequency (1 = never, 5 = very
often). All participants in the experimental group reported usage (Mfrequency = 4.14, SD = 0.81),
while none in the control group reported any usage (Mfrequency = 1, SD = 0). We also obtained
usage logs, tracking the number of times participants in the experimental group used their
ChatGPT accounts (M = 33.72, SD = 8.17), confirming that all employees in the experimental
group used ChatGPT.
Dependent variable: Creativity. To ensure the robustness of our findings, we measured
creativity using two complementary approaches. First, employees’ direct supervisors rated their
general creative performance over the week using Zhou and George’s (2001) creativity scale. A
sample item is “This employee came up with creative solutions to problems” (1 = strongly
disagree, 5 = strongly agree; α = .97). Supervisors were only approached at the end of the
experiment. They were unaware of the experiment and were blind to the study’s hypotheses.
Second, to supplement supervisor ratings, two external raters independently evaluated
employees’ responses to a question on privacy protection in the digital workplace. In the second
survey, employees were asked to respond to the following challenge: “In today’s era of
widespread digitalization, companies use numerous digital devices. What suggestions/opinions/
methods do you have for protecting employee privacy (e.g., preventing personal information
leakage and the possibility of company leadership monitoring every action) when using these
digital devices?” They were instructed to provide detailed responses of at least 70 Chinese
characters. Following the consensual assessment technique (Amabile, 1982), the raters

12
GENERATIVE AI USE AND CREATIVITY
independently rated how novel and useful each response was (1 = least, 5 = most; Lu et al.,
2017). Interrater reliability was good (ICCnovelty = .78; ICCusefulness = .69).1 Importantly,
supervisor ratings and external rater ratings were significantly correlated (novelty: r = .35, p
< .001; usefulness: r = .37, p < .001), reinforcing the validity of our creativity measures.
Mediator and moderator. Cognitive job resources. We used the cognitive job resources
scale from de Jonge and colleagues (2012). Starting with the item stem “Over the last week,” a
sample item is “I had access to useful information to help solve complex tasks” (1 = strongly
disagree, 5 = strongly agree; α = .91). Metacognitive strategies. We used the metacognitive
strategies scale from Chen and colleagues (2020). A sample item is “While working towards my
goal, I kept track of how effective my approach was” (1 = never, 5 = most of the time; α = .85).
Controls. Although control variables are not required to test the treatment effects in a
randomized experiment, we explored several motivational and attitudinal variables as potential
alternative mediators (Liu et al., 2016; Yam, Tang, et al., 2023): creative self-efficacy (Tierney &
Farmer, 2011; α = .86), intrinsic motivation (Grant, 2008; α = .82), and job insecurity (Feather &
Rauter, 2004; α = .90). To account for job differences, we also controlled for task characteristics,
specifically heuristic tasks, which are known to require creative problem-solving (Zhou, 2022).
We used established measures of heuristic tasks from George and Zhou (2001), including unclear
ends (α = .88) and unclear means (α = .84). Additionally, we included employee past job
performance from company records, given high and low performers may respond to AI
differently (Chen & Chan, 2024; Jia et al., 2024). We presented results with controls in the main
text and results without controls in the supplementary materials to illustrate the robustness of our
findings.
Results
Table 1 presents the descriptive statistics. Table 2 summarizes cognitive job resources and
creativity measures—supervisor-rated creativity (hereafter, creativity), and external-rater-rated

13
GENERATIVE AI USE AND CREATIVITY
novelty (novelty) and usefulness (usefulness)—across experimental conditions (see Figure 2 for
visual illustrations). We formally tested our hypothesis using multilevel analyses, given the
hierarchical structure of our data: 250 employee participants are nested within 30 supervisors,
who are nested within 3 departments (Bliese & Hanges, 2004).
H1 posits that LLM use increases cognitive job resources. As hypothesized, LLM
assistance increased cognitive job resources (Table 3 Model 1: γ = 0.66, SE = 0.10, p < .001).
H2 posits that LLM use increases employee creativity. In support of H2, LLM assistance
increased creativity (Table 3 Model 2: γ = 0.84, SE = 0.10, p < .001) and novelty (Table 3 Model
3: γ = 0.25, SE = 0.13, p = 0.049). Although the effect on usefulness was not significant with
controls (Table 3 Model 4: γ = 0.17, SE = 0.11, p = .136), it was significant without controls
(Table S5 Model 4: γ = 0.28, SE = 0.11, p = .009).2
H3 posits that cognitive job resources mediate the relationship between LLM use and
creativity. As depicted in Table 3, LLM assistance increased cognitive job resources (Model 1: γ
= 0.66, SE = 0.10, p < .001), which were positively related to creativity (Model 5: γ = 0.21, SE =
0.06, p = .001), novelty (Model 6: γ = 0.17, SE = 0.08, p = .021), and usefulness (Model 7: γ =
0.22, SE = 0.07, p = .001). We tested the indirect effects using parametric bootstrapping with
10,000 repetitions (Preacher et al., 2010). Results supported significant mediation effects: for
creativity (indirect effect = 0.14, 95% CI [0.052, 0.236]), novelty (indirect effect = 0.12, 95% CI
[0.018, 0.228]), and usefulness (indirect effect = 0.15, 95% CI [0.060, 0.252]). Thus, H3 was
supported.
H4 states that acquiring cognitive job resources through LLM use depends on employees’
metacognitive strategies. Supporting this, we found a significant interaction between LLM
assistance and metacognitive strategies on cognitive job resources (Table 3 Model 8: γ = 0.62, SE
= 0.16, p < .001). To interpret it, we used two methods. First, we follow Aiken and West (1991)
to plot simple slopes at one SD below and above the mean of metacognitive strategies. As shown

14
GENERATIVE AI USE AND CREATIVITY
in Figure 3, when metacognitive strategies were low, LLM use did not significantly increase
cognitive job resources (γ = 0.26, SE = 0.14, p = .067); however, when metacognitive strategies
were high, this relationship became significant (γ = 1.01, SE = 0.14, p < .001). Second, we used
the Johnson–Neyman technique to pinpoint the levels of metacognitive strategies at which the
simple slopes become significant (Preacher et al., 2006). As shown in Figure 4, the simple slopes
became significant when mean-centered metacognitive strategies were ≥ –0.58, indicating that
employees at or above this threshold benefited from LLM use. Thus, H4 was supported.
H5 posits that the indirect effect of LLM use on creativity via cognitive job resources is
moderated by metacognitive strategies. As hypothesized, metacognitive strategies moderated the
relationship between LLM assistance and cognitive job resources (Model 8: γ = 0.62, SE = 0.16,
p < .001), which were related to creativity (Model 9: γ = 0.23, SE = 0.06, p < .001), novelty
(Model 10: γ = 0.16, SE = 0.08, p = .041), and usefulness (Model 11: γ = 0.22, SE = 0.07, p
= .001). We used parametric bootstrapping to estimate conditional indirect effects at one SD
above and below the mean of metacognitive strategies. When metacognitive strategies were high,
the indirect effects were significant for creativity (0.23, 95% CI [0.099, 0.375]), novelty (0.16,
95% CI [0.008, 0.328]), and usefulness (0.22, 95% CI [0.083, 0.376]). When metacognitive
strategies were low, the effects were nonsignificant: creativity (0.06, 95% CI [–0.005, 0.142]),
novelty (0.04, 95% CI [–0.006, 0.118]), and usefulness (0.06, 95% CI [–0.005, 0.142]). In all
cases, the difference between the conditional indirect effects at high vs. low levels was
statistically significant: creativity (Δ = 0.17, 95% CI [0.057, 0.315]), novelty (Δ = 0.12, 95% CI
[0.005, 0.267]), and usefulness (Δ = 0.16, 95% CI [0.050, 0.314]). Thus, H5 was supported.
Exploratory analyses and robustness tests. Our findings show that metacognitive
strategies moderated the effect of LLM assistance on cognitive job resources, which in turn
affected employee creativity. Since the cognitive job resources scale contains multiple items
capturing different aspects of job resources, we conducted exploratory analyses at the item level

15
GENERATIVE AI USE AND CREATIVITY
(Tables S1 to S4). These analyses yielded results similar to those with the overall scale,
suggesting that the effects are robust across different aspects of the construct. We also ran a set of
robustness tests on the moderating role of metacognitive strategies and assessed the practical
significance of the moderated mediation effects, which were reported in Sections S6 to S8.
Discussion
Theoretical Contributions
First, our research contributes to the emerging literature on LLMs and creativity by
investigating the impact of LLMs on employee creativity in the workplace. Our field experiment
complements existing research that predominantly uses a single-task paradigm in online and lab
settings (e.g., Anderson et al., 2024; Boussioux et al., 2024; Chen & Chan, 2024; Doshi &
Hauser, 2023; Girotra et al., 2023; Góes et al., 2023). Importantly, we advance a theory that
explains how and for whom LLM use enhances creativity in the workplace. Regarding the how
question, our study highlights that LLM use enhances employee creativity by providing
cognitive job resources. This mediating role of cognitive job resources—such as opportunities to
switch between simple and complex tasks and to take mental breaks—cannot be detected in the
single-task paradigm used in prior research. Regarding the for whom question, we identify the
moderating role of metacognitive strategies, which enable employees to leverage LLMs to
acquire cognitive job resources that boost creativity. Because prior experiments often constrained
how participants interacted with LLMs, such as through predefined prompt engineering
techniques, their research designs prevented participants from using metacognitive strategies to
leverage LLMs for creative problem-solving. Overall, our field experiment advances a new
understanding of the mediating and moderating mechanisms that underlie the impact of LLMs on
employee creativity.
Second, an instructive finding of our study is the significant moderating role of
metacognitive strategies, indicating that LLM use does not automatically enhance creativity but

16
GENERATIVE AI USE AND CREATIVITY
depends on how employees engage with it. Specifically, employees with high levels of
metacognitive strategies—those who actively analyze tasks, monitor their thought processes, and
adjust their approaches (Chen et al., 2020; Flavell, 1979)—are better positioned to harness LLMs
in ways that foster creativity. By continuously evaluating tasks and assessing the effectiveness of
their approaches, these employees can more effectively identify the information they need and
determine when to switch tasks or take mental breaks to restore cognitive capacity (Davidson &
Sternberg, 1998; Sun, 2024). Our findings thus highlight the importance of employees actively
monitoring and regulating their thinking when using LLMs for creative work, offering a valuable
theoretical lens for future studies at the intersection of LLMs and creativity.
Third, our research contributes to the theory and research on cognitive job resources.
While research has highlighted the importance of cognitive job resources in fostering creativity,
little attention has been given to their antecedents. Our research identifies LLMs as a
technological source of cognitive job resources, thus extending this concept beyond conventional
job design factors (de Jonge & Dormann, 2006; Oldham & Fried, 2016). Further, although
traditional AI tools and human mentors or experts can also provide cognitive job resources,
LLMs are fundamentally distinct. Traditional AI tools—such as decision support systems or
customer service chatbots—are typically narrow in scope and designed for codifiable, repetitive
tasks (Acemoglu, 2024; Autor, 2014). By contrast, LLMs constitute a different class of
technology, offering broad, general-purpose capabilities (Eloundou et al., 2023).3 Additionally,
unlike human mentors, LLMs offer immediate access to a vast breadth of knowledge—far
exceeding what any single expert, or even a group of experts, can realistically provide (Luo et
al., 2024). LLMs also provide flexible, on-demand support, such as assistance with diverse tasks
that facilitate task-switching and mental breaks—support that would be difficult to request
repeatedly from human mentors. These distinctions highlight LLMs as a unique and scalable
source of cognitive job resources. We further contribute by demonstrating that metacognitive

17
GENERATIVE AI USE AND CREATIVITY
strategies interact with LLM use to shape cognitive job resources, providing a more nuanced
understanding of how cognitive job resources emerge through LLM use.
Practical Implications
Our findings offer practical guidance for organizations considering LLM deployment to
boost employee creativity. We show that LLMs enhance employee creativity by providing
cognitive job resources, suggesting that organizations should leverage LLMs to boost these
cognitive job resources and encourage employees to actively use LLMs to acquire these
resources for creativity. Critically, our findings highlight the enabling role of employees’
metacognitive strategies in leveraging LLMs to acquire these resources. Organizations should
therefore consider employees’ metacognitive abilities when implementing LLMs and invest in
developing them through training. Notably, metacognitive strategies, while often viewed as
individual differences, are teachable through interventions (for reviews, see Sun, 2024). These
interventions range from brief social-psychological exercises (e.g., Chen et al., 2017; Chen et al.,
2020) and training sessions (e.g., Bell & Kozlowski, 2008; Keith & Frese, 2005) to longer
programs spanning several days or weeks (e.g., Carpenter et al., 2019; Dierdorff & Ellington,
2012). For instance, Chen et al. (2020) developed a brief online exercise using anecdotes and
research findings to enhance metacognitive strategies, while Keith and Frese (2005) showed that
a 2.5-hour training combining metacognitive instruction and error management significantly
improved these strategies. Depending on budget and priorities, organizations may adopt brief
interventions or more extensive programs. Organizations may also combine training with
selective hiring, though the latter’s cost-effectiveness may vary with labor market conditions
(Weinstein, 2018).
Limitations and Future Directions
First, our reliance on self-reported metacognitive strategies is a limitation. However,
given metacognitive strategies involve personal awareness and regulation, self-reporting remains

18
GENERATIVE AI USE AND CREATIVITY
the most direct and practical method in large participant groups (Craig et al., 2020). Despite this
limitation, several features of our design mitigate biases associated with self-reported data (e.g.,
common method variance [CMV]). One mitigating factor is that the independent variable was
experimentally manipulated, while the dependent variables were assessed by supervisors and
external raters, thus reducing CMV (Podsakoff et al., 2003). Further, we focused on the
moderating effects of metacognitive strategies, and research shows that “interaction effects
cannot be artifacts of CMV” (Siemsen et al., 2010, p. 456).
Second, while our findings show that metacognitive strategies moderate whether
employees can use LLMs to acquire cognitive job resources necessary for creativity, future
research should explore additional individual-difference moderators. Complementing our
cognitive perspective, motivational factors may also shape how employees engage with LLMs to
enhance cognitive job resources and, in turn, creativity. For instance, learning goal orientation—
a person’s motivation to develop new skills and master new situations (Dweck, 1986)—may also
moderate the relationship between LLM assistance and cognitive job resources. Individuals high
in learning goal orientation are eager to acquire new knowledge (Vandewalle et al., 2019) and
may use LLMs more effectively to gather information, thereby boosting their creative potential.
Similarly, promotion focus—a motivation to pursue advancement and positive outcomes
(Higgins, 1997)—may encourage employees to proactively leverage LLMs to tackle complex
problems (Lanaj et al., 2012), thereby increasing creative potential. Moreover, given the
moderating role of metacognitive strategies observed in our experiment, future research could
also examine whether motivations influence the use of metacognitive strategies, thus shaping the
impact of LLMs on cognitive resources and creative outcomes (Muller et al., 2005).
Third, our field experiment was conducted within a single organization in China.
Although our theory is not tied to a specific organization or culture, the generalizability of our
findings across cultural contexts remains an open empirical question. Individuals from different

19
GENERATIVE AI USE AND CREATIVITY
cultures may differ in their attitudes toward AI (Yam, Tan, et al., 2023), and LLM outputs can
reflect cultural tendencies embedded in the training data or shaped by the language of prompts
(e.g., English vs. Chinese; Lu et al., 2025). This raises questions about the reinforcement of
cultural norms through LLMs, which is an important topic for future research.
Fourth, organizations are multilevel, with individuals nested within teams and broader
organizational systems. Because cognition and behavior are shaped by contexts (Johns, 2018),
future research should examine how team- and organizational-level factors influence the
cognition-based mechanisms in our model. For example, team and organizational environments
may affect employees’ use of metacognitive strategies when interacting with LLMs, as
environments requiring active thinking—such as those involving explorations, errors, and
challenge stressors—can promote the use of these strategies (Keith & Frese, 2005; Sun, 2024).
Additionally, organizational and team norms around LLM use may shape employees’ attitudes
toward adoption and usage (Kodapanakkal et al., 2020; Qin et al., 2025), ultimately impacting
their ability to access cognitive resources critical for creativity.
Finally, future research should explore long-term effects of extended LLM use. The
Matthew Effect (Rigney, 2010) suggests that individuals with initial advantages, such as strong
metacognitive strategies, may experience compounding benefits over time. However, prolonged
reliance on LLMs may also carry downsides. For instance, employees who enhance creativity
through LLM-assisted cognitive job resources may become increasingly dependent on these
tools, potentially reducing autonomy, learning, and networking—factors essential for sustaining
creativity over time. These contrasting possibilities underscore the need to identify conditions
under which such divergent outcomes may emerge. Longitudinal studies tracking cohorts of
employees over extended periods would enable researchers to examine not only changes in
creativity but also whether patterns of LLM use contribute to skill development or overreliance
that impairs independent thinking. These questions await future research.

20
GENERATIVE AI USE AND CREATIVITY
References
Acemoglu, D. (2024). The simple microeconomics of AI. National Bureau of Economic
Research. https://doi.org/10.3386/w32487
Amabile, T. M. (1982). Social psychology of creativity: A consensual assessment technique.
Journal of Personality and Social Psychology, 43(5), 997-1013.
Amabile, T. M. (1988). A model of creativity and innovation in organizations. Research in
Organizational Behavior, 10(1), 123-167.
Amabile, T. M., & Gryskiewicz, S. S. (1987). Creativity in the R&D laboratory. Technical
Report 30. Center for Creative Leadership.
Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homogenization Effects of Large
Language Models on Human Creative Ideation. arXiv preprint arXiv:2402.01536.
https://doi.org/https://doi.org/10.1145/3635636.3656204
Anderson, N., Potočnik, K., & Zhou, J. (2014). Innovation and creativity in organizations: A
state-of-the-science review, prospective commentary, and guiding framework. Journal of
Management, 40(5), 1297-1333.
Autor, D. H. (2014). Polanyi's Paradox and the Shape of Employment Growth (September 2014).
NBER Working Paper No. w20485, Available at SSRN:
https://ssrn.com/abstract=2496241.
Beeftink, F., Van Eerde, W., & Rutte, C. G. (2008). The effect of interruptions and breaks on
insight and impasses: Do you need a break right now? Creativity Research Journal,
20(4), 358-364.
Bell, B. S., & Kozlowski, S. W. (2008). Active learning: Effects of core training design elements
on self-regulatory processes, learning, and adaptability. Journal of Applied Psychology,
93(2), 296-316.
Bliese, P. D., & Hanges, P. J. (2004). Being both too liberal and too conservative: The perils of
treating grouped data as though they were independent. Organizational Research
Methods, 7(4), 400-417.
Boussioux, L., Lane, J. N., Zhang, M., Jacimovic, V., & Lakhani, K. R. (2024). The crowdless
future? Generative AI and creative problem-solving. Organization Science, 35(5), 1589-
1607.
Brislin, R. W. (1970). Back-translation for cross-cultural research. Journal of Cross-Cultural
Psychology, 1(3), 185–216.
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly
journal of economics, 140(2), 889-942. https://doi.org/https://doi.org/10.1093/qje/qjae044
Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., Lee, P., Lee, Y. T.,
Li, Y., & Lundberg, S. (2023). Sparks of artificial general intelligence: Early experiments
with gpt-4. arXiv preprint arXiv:2303.12712.
Carpenter, J., Sherman, M. T., Kievit, R. A., Seth, A. K., Lau, H., & Fleming, S. M. (2019).
Domain-general enhancements of metacognitive ability through adaptive training.
Journal of Experimental Psychology: General, 148(1), 51-64.
Chen, P., Chavez, O., Ong, D. C., & Gunderson, B. (2017). Strategic resource use for learning: A
self-administered intervention that guides self-reflection on effective resource use
enhances academic performance. Psychological Science, 28(6), 774-785.
Chen, P., Powers, J. T., Katragadda, K. R., Cohen, G. L., & Dweck, C. S. (2020). A strategic
mindset: An orientation toward strategic behavior during goal pursuit. Proceedings of the
National Academy of Sciences, 117(25), 14066-14072.

21
GENERATIVE AI USE AND CREATIVITY
Chen, Z., & Chan, J. (2024). Large language model in creative work: The role of collaboration
modality and user expertise. Management science, 1-17.
Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating normed and
standardized assessment instruments in psychology. Psychological assessment, 6(4), 284-
290.
Craig, K., Hale, D., Grainger, C., & Stewart, M. E. (2020). Evaluating metacognitive self-
reports: systematic reviews of the value of self-report in metacognitive research.
Metacognition and learning, 15, 155-213.
Davenport, T. H., & Kirby, J. (2016). Only humans need apply: Winners and losers in the age of
smart machines. Harper Business New York.
Davidson, J. E., & Sternberg, R. J. (1998). Smart problem solving: How metacognition helps. In
D. J. Hacker, J. Dunlosky, & A. C. Graesser (Eds.), Metacognition in educational theory
and practice (pp. 47-68). Lawrence Erlbaum Associates Publishers.
de Jonge, J., & Dormann, C. (2006). Stressors, resources, and strain at work: a longitudinal test
of the triple-match principle. Journal of Applied Psychology, 91(6), 1359–1374.
de Jonge, J., Spoor, E., Sonnentag, S., Dormann, C., & van den Tooren, M. (2012). “Take a
break?!” Off-job recovery, job demands, and job resources as predictors of health, active
learning, and creativity. European Journal of Work and Organizational Psychology,
21(3), 321-348.
Dierdorff, E. C., & Ellington, J. K. (2012). Members matter in team training: Multilevel and
longitudinal relationships between goal orientation, self-regulation, and team outcomes.
Personnel Psychology, 65(3), 661-703.
Doshi, A. R., & Hauser, O. (2023). Generative artificial intelligence enhances creativity.
Available at SSRN.
Dweck, C. S. (1986). Motivational processes affecting learning. American psychologist, 41(10),
1040-1048.
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). Gpts are gpts: An early look at the
labor market impact potential of large language models. arXiv preprint
arXiv:2303.10130.
Elsbach, K. D., & Hargadon, A. B. (2006). Enhancing creativity through “mindless” work: A
framework of workday design. Organization Science, 17(4), 470-483.
Feather, N. T., & Rauter, K. A. (2004). Organizational citizenship behaviours in relation to job
status, job insecurity, organizational commitment and identification, job satisfaction and
work values. Journal of Occupational and Organizational Psychology, 77(1), 81-94.
Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive–
developmental inquiry. American psychologist, 34(10), 906-911.
Fleming, L. (2001). Recombinant uncertainty in technological search. Management science,
47(1), 117-132.
Fleming, L., Mingo, S., & Chen, D. (2007). Collaborative brokerage, generative creativity, and
creative success. Administrative science quarterly, 52(3), 443-475.
George, J. M., & Zhou, J. (2001). When openness to experience and conscientiousness are
related to creative behavior: an interactional approach. Journal of Applied Psychology,
86(3), 513-524.
Girotra, K., Meincke, L., Terwiesch, C., & Ulrich, K. T. (2023). Ideas are dimes a dozen: Large
language models for idea generation in innovation. Available at SSRN 4526071.

22
GENERATIVE AI USE AND CREATIVITY
Góes, L. F., Volpe, M., Sawicki, P., Grses, M., & Watson, J. (2023). Pushing gpt’s creativity to its
limits: Alternative uses and torrance tests.
Grant, A. M. (2008). The significance of task significance: Job performance effects, relational
mechanisms, and boundary conditions. Journal of Applied Psychology, 93(1), 108-124.
Hallgren, K. A. (2012). Computing inter-rater reliability for observational data: an overview and
tutorial. Tutorials in quantitative methods for psychology, 8(1), 23-34.
Hargadon, A. (2002). Brokering knowledge: Linking learning and innovation. Research in
Organizational Behavior, 24, 41-85.
Hargadon, A., & Sutton, R. I. (1997). Technology brokering and innovation in a product
development firm. Administrative science quarterly, 716-749.
Higgins, E. T. (1997). Beyond pleasure and pain. American psychologist, 52(12), 1280.
Hitsuwari, J., Ueda, Y., Yun, W., & Nomura, M. (2023). Does human–AI collaboration lead to
more creative art? Aesthetic evaluation of human-made and AI-generated haiku poetry.
Computers in Human Behavior, 139, 107502.
Houter, S. (2024). AI in the workplace: Answering 3 big questions. Gallup.
https://www.gallup.com/workplace/651203/workplace-answering-big-questions.aspx
Ivcevic, Z., & Grandinetti, M. (2024). Artificial intelligence as a tool for creativity. Journal of
Creativity, 34(2), 100079.
Jia, N., Luo, X., Fang, Z., & Liao, C. (2024). When and how artificial intelligence augments
employee creativity. Academy of Management Journal, 67, 5-32.
Johns, G. (2018). Advances in the treatment of context in organizational research. Annual Review
of Organizational Psychology and Organizational Behavior, 5(1), 21-46.
Jones, B. F. (2009). The burden of knowledge and the “death of the renaissance man”: Is
innovation getting harder? The Review of Economic Studies, 76(1), 283-317.
Keith, N., & Frese, M. (2005). Self-regulation in error management training: emotion control and
metacognition as mediators of performance effects. Journal of Applied Psychology, 90(4),
677-691.
Kodapanakkal, R. I., Brandt, M. J., Kogler, C., & Van Beest, I. (2020). Self-interest and data
protection drive the adoption and moral acceptability of big data technologies: A conjoint
analysis approach. Computers in Human Behavior, 108, 106303.
Lanaj, K., Chang, C.-H., & Johnson, R. E. (2012). Regulatory focus and work-related outcomes:
a review and meta-analysis. Psychological bulletin, 138(5), 998–1034.
Leahey, E., Beckman, C. M., & Stanko, T. L. (2017). Prominent but less productive: The impact
of interdisciplinarity on scientists’ research. Administrative science quarterly, 62(1), 105-
139.
Liu, D., Jiang, K., Shalley, C. E., Keem, S., & Zhou, J. (2016). Motivational mechanisms of
employee creativity: A meta-analytic examination and theoretical extension of the
creativity literature. Organizational Behavior and Human Decision Processes, 137, 236-
263.
Lu, J. G. (2024). A creativity stereotype perspective on the Bamboo Ceiling: Low perceived
creativity explains the underrepresentation of East Asian leaders in the United States.
Journal of Applied Psychology, 109(2), 238-256.
Lu, J. G., Akinola, M., & Mason, M. F. (2017). “Switching On” creativity: Task switching can
increase creativity by reducing cognitive fixation. Organizational Behavior and Human
Decision Processes, 139, 63-75.

23
GENERATIVE AI USE AND CREATIVITY
Lu, J. G., Song, L. L., & Zhang, L. D. (2025). Cultural tendencies in generative AI. Nature
Human Behaviour
Luo, X., Rechardt, A., Sun, G., Nejad, K. K., Yáñez, F., Yilmaz, B., Lee, K., Cohen, A. O.,
Borghesani, V., & Pashkov, A. (2024). Large language models surpass human experts in
predicting neuroscience results. Nature Human Behaviour, 1-11.
Madjar, N., & Shalley, C. E. (2008). Multiple tasks' and multiple goals' effect on creativity:
Forced incubation or just a distraction? Journal of Management, 34(4), 786-805.
McCormick, C. B. (2003). Metacognition and learning. In I. B. Weiner (Ed.), Handbook of
psychology: Educational psychology (pp. 79-102). Wiley.
Muller, D., Judd, C. M., & Yzerbyt, V. Y. (2005). When moderation is mediated and mediation is
moderated. Journal of Personality and Social Psychology, 89(6), 852-863.
Mumford, M. D., & Gustafson, S. B. (1988). Creativity syndrome: Integration, application, and
innovation. Psychological bulletin, 103(1), 27–43.
Nathan, A. (2024). GEN AI: Too much spend, too little benefit? Goldman Sachs Top of Mind,
(129), 3.
Niks, I. M., de Jonge, J., Gevers, J. M., & Houtman, I. L. (2017). Divergent effects of
detachment from work: A day-level study on employee creativity. European Journal of
Work and Organizational Psychology, 26(2), 183-194.
Oldham, G. R., & Cummings, A. (1996). Employee creativity: Personal and contextual factors at
work. Academy of Management Journal, 39(3), 607-634.
Oldham, G. R., & Fried, Y. (2016). Job design research and theory: Past, present and future.
Organizational Behavior and Human Decision Processes, 136, 20-35.
Perry-Smith, J. E. (2006). Social yet creative: The role of social relationships in facilitating
individual creativity. Academy of Management Journal, 49(1), 85-101.
Pfeffer, J. (2018). Dying for a Paycheck: How Modern Management Harms Employee Health
and Company Performance and what We Can Do about it. HarperCollins Publishers.
Podsakoff, P. M., MacKenzie, S. B., Lee, J.-Y., & Podsakoff, N. P. (2003). Common method
biases in behavioral research: a critical review of the literature and recommended
remedies. Journal of Applied Psychology, 88(5), 879-903.
Preacher, K. J., Curran, P. J., & Bauer, D. J. (2006). Computational tools for probing interactions
in multiple linear regression, multilevel modeling, and latent curve analysis. Journal of
educational and behavioral statistics, 31(4), 437-448.
Preacher, K. J., Zyphur, M. J., & Zhang, Z. (2010). A general multilevel SEM framework for
assessing multilevel mediation. Psychological methods, 15(3), 209-233.
Qin, X., Zhou, X., Chen, C., Wu, D., Zhou, H., Dong, X., Cao, L., & Lu, J. G. (2025). AI
aversion or appreciation? A capability-personalization framework and a meta-analytic
review. Psychological bulletin.
Reiter-Palmon, R., & Arreola, N. J. (2015). Does generating multiple ideas lead to increased
creativity? A comparison of generating one idea vs. many. Creativity Research Journal,
27(4), 369-374.
Reuters. (2024). OpenAI to cut access to tools for developers in China, other regions, Chinese
state media says. https://www.reuters.com/technology/artificial-intelligence/openai-cut-
access-tools-developers-china-other-regions-chinese-state-media-says-2024-06-25/
Rigney, D. (2010). The Matthew effect: How advantage begets further advantage. Columbia
University Press.

24
GENERATIVE AI USE AND CREATIVITY
Siemsen, E., Roth, A., & Oliveira, P. (2010). Common method bias in regression models with
linear, quadratic, and interaction effects. Organizational Research Methods, 13(3), 456-
476. https://doi.org/https://doi.org/10.1177/1094428109351241
Singla, A., Sukharevsky, A., Yee, L., & Chui, M. (2024). The state of AI in early 2024: Gen AI
adoption spikes and starts to generate value. . McKinsey Global Institute.
Smith, S. M. (1995). Fixation, incubation, and insight in memory and creative thinking.
Sonnentag, S., Binnewies, C., & Mojza, E. J. (2010). Staying well and engaged when demands
are high: the role of psychological detachment. Journal of Applied Psychology, 95(5),
965.
Sun, S. (2024). A componential and functional framework for metacognition: Implications for
research in personnel and human resources management. In M. R. Buckley, Wheeler,
A.R., Baur, J.E. and Halbesleben, J.R.B. (Ed.), Research in Personnel and Human
Resources Management (Vol. 42, pp. 45–73). Emerald Publishing Limited.
Sun, S., Wang, N., Zhu, J., & Song, Z. (2020). Crafting job demands and employee creativity: A
diary study. Human Resource Management, 59(6), 569-583.
Teodoridis, F., Bikard, M., & Vakili, K. (2019). Creativity at the knowledge frontier: The impact
of specialization in fast-and slow-paced domains. Administrative science quarterly, 64(4),
894-927.
Tierney, P., & Farmer, S. M. (2011). Creative self-efficacy development and creative
performance over time. Journal of Applied Psychology, 96(2), 277-293.
Tiwari, A. (2024). Scaling AI: Unlocking new growth and tackling industry-wide challenges. The
Economic Times.
Unsworth, K. (2001). Unpacking creativity. Academy of management review, 26(2), 289-297.
van den Tooren, M., & de Jonge, J. (2010). The role of matching job resources in different
demanding situations at work: A vignette study. Journal of Occupational and
Organizational Psychology, 83(1), 39-54.
Veenman, M. V., Wilhelm, P., & Beishuizen, J. J. (2004). The relation between intellectual and
metacognitive skills from a developmental perspective. Learning and instruction, 14(1),
89-109.
Weinstein, R. (2018). Employer screening costs, recruiting strategies, and labor market
outcomes: An equilibrium analysis of on-campus recruiting. Labour Economics, 55, 282-
299.
Winne, P. H., & Nesbit, J. C. (2010). The psychology of academic achievement. Annual Review
of Psychology, 61, 653-678.
Yam, K. C., Tan, T., Jackson, J. C., Shariff, A., & Gray, K. (2023). Cultural differences in
people's reactions and applications of robots, algorithms, and artificial intelligence.
Management and organization review, 19(5), 859-875.
Yam, K. C., Tang, P. M., Jackson, J. C., Su, R., & Gray, K. (2023). The rise of robots increases
job insecurity and maladaptive workplace behaviors: Multimethod evidence. Journal of
Applied Psychology, 108(5), 850-870.
Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., &
Dong, Z. (2023). A survey of large language models. arXiv preprint arXiv:2303.18223,
1(2).
Zhou, J. (2022). Managing Creativity under Uncertainty. In G. Grote & M. Griffin (Eds.), Oxford
handbook on uncertainty management in work organizations. Oxford University Press.

25
GENERATIVE AI USE AND CREATIVITY
Zhou, J., & George, J. M. (2001). When job dissatisfaction leads to creativity: Encouraging the
expression of voice. Academy of Management Journal, 44(4), 682-696.
Zhou, J., & Shalley, C. E. (2003). Research on employee creativity: A critical review and
directions for future research. Research in personnel and human resource
management/Elsevier, 22, 165-217.
Zhou, J., & Shalley, C. E. (2011). Deepening our understanding of creativity in the workplace: A
review of different approaches to creativity research. In S. Zedeck (Ed.), APA Handbook
of Industrial and Organizational Psychology (Vol. 1. Building and developing the
organization.). American Psychological Association.

26
GENERATIVE AI USE AND CREATIVITY
Figure 1
Conceptual Model
LLM assistance
Cognitive job resources
Creativity
Metacognitive strategies

27
GENERATIVE AI USE AND CREATIVITY
Figure 2
Mean Levels of Cognitive Job Resources and Creativity (rated by supervisor and external raters) across Experimental Conditions
Note. Error bars indicate standard errors.

28
GENERATIVE AI USE AND CREATIVITY
Figure 3
The Moderating Patterns of Metacognitive Strategies: Simple Slopes

29
GENERATIVE AI USE AND CREATIVITY
Figure 4
The Moderating Patterns of Metacognitive Strategies: Regions of Significance Using the
Johnson–Neyman Technique
Note. Simple slopes for cognitive job resources between the lower bound (-2.23) and the upper
bound (-0.58) are not statistically significant, as the confidence bands contain zero within this
range. The plot indicates that simple slopes become significant when mean-centered
metacognitive strategies reach or exceed -0.58 (equivalent to a raw score of 3.49).

30
GENERATIVE AI USE AND CREATIVITY
1 Cicchetti (1994, p. 286) and Hallgren (2012, p. 32) provide the commonly cited ICC cutoffs, with agreement rated
poor (< .40), fair (.40–.59), good (.60–.74), and excellent (.75–1.0).
2 Both the independent-samples t-tests (Table 2; p values ranging from .010 to < .001) and the multilevel analyses
without control variables (Supplemental Materials S4; p values ranging from .006 to < .001) find a consistent
treatment effect across all three creativity measures (i.e., supervisor-rated creativity, external-rater-rated novelty, and
external-rater-rated usefulness). In the main text, we include models with control variables and additional mediators,
which lead to attenuated direct effects of LLM assistance on creativity outcomes.
3 Unlike traditional web search engines such as Google, LLMs generate responses based on context, enabling them
to synthesize and integrate disparate pieces of knowledge into coherent, accessible outputs, rather than merely
presenting a list of web pages (Lee & Chung, 2024; Zhao et al., 2023). Traditional search engines also lack the
capacity to support the wide range of cognitive tasks that LLMs, as general-purpose technologies, can facilitate
(Eloundou et al., 2023; Zhao et al., 2023).