Workers who relied on generative artificial intelligence to produce writing with little personal input reported lower confidence in their ability to work without AI, weaker psychological ownership of their output and lower perceived meaning than people who worked independently, according to a preregistered experiment.
The study, published in Scientific Reports, also found a different pattern when workers remained actively involved. Participants who drafted their own material before using AI to refine it generally reported psychological outcomes similar to those who completed the work without AI.
How the researchers tested AI use at work
Researchers recruited professionals from several occupational groups and gave them role-specific writing tasks designed to resemble real workplace activities. These included press releases, short reports, analysis plans and sensitive emails.
The primary experiment initially produced 408 completed participants. The preregistered main analysis included 269 participants who confirmed that they had followed their assigned AI-use instructions. Participants included consultants, data analysts, human resources professionals, managers and marketers.
They were randomly assigned to one of three conditions. One group completed the task without AI. A passive-use group used AI-generated material directly without modifying it. A collaborative group first wrote its own draft and then used AI to review and edit that draft.
After the first task, everyone completed a second writing task without AI. This allowed the researchers to test whether any psychological differences remained after participants returned to independent work. A separate survey of 270 working adults in the United States and United Kingdom examined whether similar relationships appeared in everyday professional AI use.
Passive AI use was linked to lower self-efficacy
Immediately after the first task, AI-independent self-efficacy differed significantly between conditions, F(2,266) = 3.54, p = 0.030. Participants who copied and pasted AI output reported a mean self-efficacy score of 5.16 on the seven-point scale, compared with 5.63 among those who used no AI.
The human-first collaborative group recorded a mean of 5.43. Its self-efficacy did not differ significantly from either the no-AI group or the passive AI group in the pairwise comparisons.
After participants subsequently completed a task without AI, differences in self-efficacy remained significant, F(2,266) = 6.25, p = 0.002. Those who had previously copied AI output recorded a mean of 4.95, compared with 5.66 among participants who had previously worked without AI. The collaborative group recorded 5.45.
A large difference emerged in psychological ownership
The clearest immediate difference concerned psychological ownership, or the feeling that the resulting work belonged to the participant. The overall effect was significant, F(2,266) = 27.12, p < 0.001, with a partial eta squared of 0.169.
The passive AI group reported a mean ownership score of 4.35, compared with 5.26 in the collaborative group and 5.34 in the no-AI group. Both differences were statistically significant. The collaborative and no-AI groups did not significantly differ from one another.
However, this ownership effect did not persist after participants completed the later manual task. At that point, psychological ownership no longer differed significantly between the groups. This suggests that independently producing a new piece of work may quickly restore a person’s immediate sense of authorship.
Work also felt less meaningful under passive AI use
Perceived meaningfulness differed significantly after the initial task, F(2,266) = 5.26, p = 0.006. Participants in the passive AI condition reported a mean score of 4.94, compared with 5.54 for no AI and 5.46 for human-first collaboration.
The passive group rated the task as significantly less meaningful than both comparison groups. The collaborative and no-AI groups were statistically similar.
Meaningfulness differences were still detectable after the later manual task, F(2,266) = 3.40, p = 0.035. Participants who had previously relied passively on AI recorded a mean of 5.07, while the collaborative and no-AI groups recorded means of 5.55 and 5.54 respectively.
Passive AI initially felt satisfying, but the advantage reversed
The results were not uniformly negative for passive AI use. Immediately after the first task, participants who copied AI output reported the highest satisfaction with their work. Their mean satisfaction score was 5.80, compared with 5.26 for the collaborative group and 4.50 for the no-AI group. The overall condition effect was significant, F(2,266) = 19.64, p < 0.001.
That advantage disappeared when participants returned to manual work. After the second task, the former passive AI group reported mean satisfaction of 4.12, significantly below the collaborative group’s 5.42 and the no-AI group’s 4.98.
Task enjoyment followed a similar reversal. After the manual task, participants who had previously copied AI output reported mean enjoyment of 4.67, compared with 5.56 for prior collaborators and 5.39 for those who had previously used no AI.
Real-world survey showed a similar pattern
The follow-up survey provided correlational evidence from 270 working adults. Greater passive reliance on AI was associated with lower self-efficacy, r = -0.45, p < 0.001, and lower psychological ownership, r = -0.22, p < 0.001. Passive reliance was also associated with lower outcome satisfaction, r = -0.16, p = 0.008.
Active collaboration showed the opposite relationships for several outcomes. It correlated positively with self-efficacy, r = 0.40, p < 0.001, psychological ownership, r = 0.30, p < 0.001, and outcome satisfaction, r = 0.14, p = 0.020.
These survey findings cannot establish that a particular AI-use style caused the psychological outcomes, but their direction was broadly consistent with the experiment.
Why the findings matter
The study suggests that workplace AI policies may need to consider more than whether employees use generative AI. How the technology is incorporated into a task may be equally important.
A workflow in which AI substitutes for most of the worker’s contribution may deliver immediate convenience and satisfaction while weakening feelings of competence, authorship and meaning. By contrast, asking workers to formulate ideas, make judgments and produce an initial contribution before using AI as an editing or refinement tool may preserve more of their psychological connection to the work.
The findings do not show that AI should be avoided. They instead indicate that human involvement may matter when organisations design AI-assisted workflows, particularly where learning, professional confidence, motivation and personal responsibility are important.
Important limitations
The experiment focused on occupation-specific writing tasks, so the results should not automatically be generalised to every form of AI-assisted work. Participants also accessed the AI tool outside the survey environment, preventing the researchers from recording detailed prompt-level behaviour and precise interaction patterns.
The active collaboration condition tested one specific workflow in which the human drafted first and AI edited afterward. Other forms of collaboration, such as AI-assisted outlining or repeated human-AI iteration, could produce different outcomes.
Only 269 of the 408 participants who completed the experiment met the preregistered manipulation-check criteria for the main analysis. The researchers reported that analyses using the full sample produced largely consistent results, but the substantial number of exclusions remains relevant when interpreting the study.
Finally, the follow-up workplace survey was correlational and relied on self-reported patterns of AI use. It strengthens the ecological relevance of the experimental results but cannot demonstrate causation.
Source Information
Study: Relying on AI at work reduces self-efficacy, ownership, and meaning while active collaboration mitigates the effects
Authors: Elena Hayoung Lee, Yidan Yin, Nan Jia and Cheryl J. Wakslak
Journal: Scientific Reports
Published: 15 March 2026
DOI: 10.1038/s41598-026-42312-6
Study design: Preregistered randomised online experiment with a main analytic sample of 269 participants, followed by a correlational survey of 270 working adults.
