The Exhaustion Isn't From the AI Itself: A Three-Wave Finnish Study Points at Your Coworkers

For two years, the dominant narrative about AI at work has been a burnout story: keep up with the tools or be ground down by them. A new peer-reviewed longitudinal study from Finland pushes back on the simplest version of that story — and replaces it with something more uncomfortable. The exhaustion, the data suggests, is not coming from the AI. It is coming from the person at the next desk, and how you feel about them.

The paper, published in SSM – Population Health (Volume 35, article 101945) by Iina Savolainen, Teijo Osma, Roope Grönroos, Moona Heiskari, and Atte Oksanen of Tampere University’s “AI Disruption at Work” research project, tracked more than 2,100 employed Finns across three survey waves: fall 2024 (N = 2,109), spring 2025 (n = 1,680), and fall 2025 (n = 1,479). Follow-up response rates of 80% and 70% are unusually strong for panel research of this kind, and the sample closely mirrors the Finnish working population — mean age 42.6 versus the national 42.2, an even gender split, and comparable university-degree rates.

What the study actually measured

The researchers built their analysis around four instruments with solid psychometric track records:

• Work exhaustion — the five-item emotional exhaustion subscale of the Maslach Burnout Inventory, scored 0–6, with excellent internal consistency (McDonald’s ω between 0.92 and 0.93 across waves). Mean exhaustion scores actually drifted slightly down over the study period, from 14.89 to 14.54 on the 0–30 composite.
• Social comparison orientation (SCO) — the six-item Iowa-Netherlands Comparison Orientation Measure (INCOM), which captures the trait-level tendency to evaluate yourself against others (ω ≈ 0.85 at every wave).
• AI readiness — a five-item scale of perceived competence with AI tools: knowing how to use them, and believing you can apply them skillfully in your own work (ω ≈ 0.82–0.83).
• AI use at work — frequency from “never” to “several times a day,” dichotomized into daily versus less-than-daily use. Among respondents who reported any AI use at all, the share using it daily climbed from about 42% at baseline to roughly 54% by fall 2025 — a snapshot of how quickly daily AI use normalized in Finland during the study window.

Crucially, the team did not settle for simple correlations. They used a within-between (Mundlak) multilevel specification, fitted with REML in R’s lme4 package, which separates within-person change (you becoming more exhausted than your own baseline) from between-person differences (exhausted people versus non-exhausted people). That distinction is exactly what most cross-sectional burnout surveys can’t make — and it’s where the story gets interesting.

Three findings that reframe the debate

First: frequent AI use, by itself, predicted nothing. In the primary models, daily AI use at work showed no association with work exhaustion — neither within persons over time nor between persons. The tool is not, on this evidence, a direct stressor.

Second: social comparison was the robust predictor. Workers higher in comparison orientation reported significantly greater exhaustion at both levels of analysis — between-person B = 0.20 (95% CI 0.15–0.25, p < 0.001) and within-person β = 0.02 (95% CI 0.01–0.04, p = 0.001). In plain terms: people who habitually measure themselves against colleagues were more exhausted, and became more exhausted when their comparison tendency rose.

Third: perceived AI readiness behaved like a personal resource. Workers who felt competent understanding and applying AI reported lower exhaustion — between-person B = −1.41 (95% CI −2.17 to −0.64) in the first model, with protective within-person effects as well. Feeling ready for the AI era, whatever your actual skill level, correlated with resilience.

And one exploratory result deserves the asterisk it comes with: among employees high in social comparison tendency, frequent AI use was associated with elevated exhaustion (B = 0.05, 95% CI 0.00–0.09, p = 0.035). The interaction didn’t hold in the primary models, but the pattern is coherent — if you’re wired to compare, watching colleagues extract visible productivity from AI you haven’t mastered is a specific, corrosive stressor. The study frames this as “techno-comparison”: competence differences that were once private becoming increasingly visible as AI skills translate directly into output.

Why this matters now

This finding lands in the middle of a genuinely confused evidence base. A widely cited 2024 US survey reported 45% higher burnout among frequent AI users; Workday’s May 2026 global research found 62% of employees saying AI had reduced their burnout risk; a 2024 Nature Humanities & Social Sciences Communications study found AI adoption had no direct burnout effect but operated through job stress. The Finnish study offers a plausible reconciliation: the conflicting results may partly reflect who is being studied — comparison-prone adopters in competitive environments versus supported ones — rather than what the technology does.

The timing also matters. This was the top-circulating research story in AI coverage on September 27, 2026, precisely because the workplace-AI backlash has become a policy question, not just an HR one. If exhaustion tracks social comparison and perceived readiness more than tool usage itself, then the intervention levers change: training and transparent norms about AI use plausibly matter more than usage limits. Mandating “AI-free Fridays” targets a variable this data suggests is close to inert.

The honest caveats

The authors are careful, and so should we be. The design is associational — no causal claims about AI use, readiness, or comparison are possible. AI readiness measured perceived competence, not tested skills, and the scale’s factorial fit was only marginal (RMSEA 0.10, SRMR 0.09), which the authors flag themselves. Burnout was operationalized solely through the emotional-exhaustion dimension, leaving the cynicism and efficacy components of the full syndrome unmeasured. Everything is self-report, drawn from a Norstat online panel that skews toward pre-registered survey takers.

But the core result survived a demanding longitudinal design with strong retention, a representative sampling frame, and a statistical approach that separates person-level trait effects from within-person change. In a debate dominated by vendor surveys and cross-sectional snapshots, that is a meaningful upgrade in rigor.

The takeaway for anyone managing an AI-transitioning team: the burnout risk may live less in the rollout than in the rollout’s social texture — who looks fluent, who feels left behind, and how visible the gap between them becomes. The machine, on this evidence, is the smaller problem.