AI Fatigue: When Relief Comes With a Cost

AI arrives in the workplace with a simple promise: take the repetitive work off people's plates so they can focus on the parts of the job that matter. For many employees, that promise is partly holding up. In a three-wave survey of 600 people who work with AI, confidence in using it was linked to higher engagement and job satisfaction, and generative AI was linked to higher productivity [1].

But the same body of research points to a second effect running alongside the relief. The early evidence suggests that AI does not always remove effort. Sometimes it moves it.

Using generative AI is not passive. It means writing prompts, reading outputs, judging whether an answer is accurate enough to trust, and deciding whether it needs revising. That cycle of checking and second-guessing draws on limited attention, and it can leave less capacity for deep thinking.

A survey of 1,488 full-time US workers by Boston Consulting Group and the University of California, Riverside gave this strain a name: "AI brain fry", meaning mental fatigue from using or overseeing AI tools beyond one's cognitive capacity. About 14% of respondents reported it. Those workers also reported more decision fatigue and more errors, and 34% said they intended to leave their jobs, compared with 25% of those who did not report brain fry [2].

Researchers have also begun to measure the problem. A 2026 study of 717 participants across four studies proposed a 15-item AI Fatigue Scale covering four dimensions: cognitive, emotional, physical and behavioural fatigue [3].

A note of caution before drawing big conclusions. These are self-reported surveys, so they show associations, not proof that AI causes the fatigue. Heavy AI users may simply have more demanding jobs. The AI Fatigue Scale is new and has so far been validated by its own authors. "AI fatigue" is best read as an emerging concept, not a settled diagnosis.

The picture is not uniformly negative. In the same three-wave study, "AI technostress" (the strain of coping with AI-related demands such as overload and uncertainty) was associated with more exhaustion, more work-family conflict and lower job satisfaction. Yet generative AI was associated with weaker negative effects of that technostress [1].

One plausible way to reconcile this with the verification burden described above: generative AI can ease some demands while adding others, and the net effect depends on the task, the tool and the workplace. The BCG authors reached a similar conclusion. In their view, the difference lies less in how much AI a person uses than in how workers, teams and leaders shape that use [2].

A different risk: being managed by an algorithm

A separate issue is algorithmic management, where software allocates tasks, tracks performance or sets the pace of work. A recent occupational-health review found that these systems are consistently associated with greater time pressure and fewer job resources such as autonomy and social support [4]. Most of this evidence comes from platform and logistics work, so it should not be assumed to apply to everyday use of an AI assistant. It does matter for organisations that build AI into how work is directed.

What organisations can do:

Assess before rollout. Build psychosocial risk assessment into the launch of a tool, not into the clean-up afterwards. Look beyond time saved: which tasks now involve verification, how much choice people have over when and how to use the tool, and whether there is room to recover. Instruments such as the AI Fatigue Scale could help, though they are still new [3].

Protect autonomy and support. The Job Demands-Resources model, a long-established framework in occupational psychology, treats autonomy and social support as resources that help people cope with demands [5]. Efficiency gains alone do not offset the loss of those resources.

Invest in leadership. A survey-based study found that AI adoption was associated with higher job stress and, through it, poorer physical health, but that coaching-style leadership (mentoring, guidance, treating disagreement as constructive) weakened that link [6]. The design of that study limits what it can say about cause, but the finding fits the broader pattern.

Be transparent about monitoring. A meta-analysis of electronic performance monitoring, covering more than 23,000 participants, found that monitoring tends to raise stress, that more transparent and less invasive monitoring is linked to better employee attitudes, and that monitoring without a clear purpose was not linked to better performance [7]. Whether AI-based monitoring can safely be used to detect fatigue early is, for now, an open question rather than an established benefit.

Build real AI literacy. Confidence with AI tools was associated with higher engagement in the study above [1]. Training that builds understanding, not just tool skills, is a reasonable bet against the uncertainty that fuels technostress, although this link has been tested less directly.

The useful question is not whether AI is good or bad for workers, but under what conditions it helps and under what conditions it wears people down. Organisations that measure only output will miss half the picture. Measuring the mental effort behind that output is a better place to start.

References

[1] Chuang, Y.-T., Chiang, H.-L., & Lin, A.-P. (2025). Insights from the Job Demands-Resources model: AI's dual impact on employees' work and life well-being. International Journal of Information Management, 83, 102887. https://doi.org/10.1016/j.ijinfomgt.2025.102887

[2] Bedard, J., Kropp, M., Hsu, M., Karaman, O. T., Hawes, J., & Kellerman, G. R. (2026, March 5). When using AI leads to "brain fry". Harvard Business Review. https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry

[3] Lau, G. R., Sandeeshwara Kasturiratna, K.T.A, Goh, A. Y. H., Tong, E. M. W., & Hartanto, A. (2026). AI fatigue in human–AI interaction: Conceptual framework, scale development and validation, and associations with AI engagement. Computers in Human Behavior Reports, 23, 101186. https://doi.org/10.1016/j.chbr.2026.101186

[4] Bowdler, M., Lahti, H., Jelenko, M., Buresti, G. & Valtonen, T. (2026). Algorithmic management and psychosocial risks at work: An emerging occupational safety and health challenge. Scandinavian Journal of Work, Environment & Health, 52(1), 1-5. https://doi.org/10.5271/sjweh.4270

[5] Bakker, A. B., & Demerouti, E. (2007). The Job Demands-Resources model: State of the art. Journal of Managerial Psychology, 22(3), 309-328. https://doi.org/10.1108/02683940710733115

[6] Jeong, J., Kim, B.-J., & Lee, J. (2024). Navigating AI transitions: How coaching leadership buffers against job stress and protects employee physical health. Frontiers in Public Health, 12, 1343932. https://doi.org/10.3389/fpubh.2024.1343932

[7] Ravid, D. M., White, J. C., Tomczak, D. L., Miles, A. F., & Behrend, T. S. (2023). A meta-analysis of the effects of electronic performance monitoring on work outcomes. Personnel Psychology, 76(1), 5-40. https://doi.org/10.1111/peps.12514