How does GenAI use affect developers' cognitive load during software work?

Perceived load, not productivity, is the outcome

The paper asks a question that adoption dashboards usually skip: when developers use generative AI tools during everyday work, does their experienced mental demand change? Industry monitoring tends to track active user rates, usage frequency, prompt or code completion counts, suggestion acceptance, and delivery metrics such as task completion time or pull-request cycle time. Those measures say little about the cognitive demands developers face while using GenAI.

The outcome here is perceived cognitive load, recorded by developers themselves on a seven-point Likert scale for each working task. That is a subjective rating, not a physiological index and not a productivity score. The study then tests whether GenAI use, task category, and task duration relate to that rating, and whether wristband-derived physiology explains variation the work context does not.

Twenty-one developers logged tasks for four days

In Q2 2025, professional developers at two SAP sites in California took part in a four-day study. The recruited cohort comprised 22 SAP employees: 12 software developers or engineers, 8 senior developers or engineers, one senior quality specialist, and one principal software architect. One participant was excluded from the current analyses for non-adherence to the study protocol, leaving 21. Participants reported an average of 5 to 10 years of experience, prior Java experience was required, and all reported regular Java use. No age or gender information was collected to protect privacy.

During the uncontrolled workplace phase, participants documented each task on a paper questionnaire kept in visual distance so it would remind them to log without interrupting work. For each task they recorded start and end times, perceived cognitive load, productivity, whether GenAI tools were used, and, if so, how helpful the tool was. They were encouraged to interact with GenAI tools for at least one continuous hour per day and could choose which tool to use. GitHub Copilot and ChatGPT were the most frequently used GenAI tools at work.

Throughout the workday participants wore an EmbracePlus wristband on their non-dominant hand, paired by Bluetooth to a smartphone. The device recorded blood volume pulse at 64 Hz, electrodermal activity at 4 Hz, skin temperature at 1 Hz, and three-axis wrist acceleration at 64 Hz. Acceleration was used only to detect motion artefacts, not as a statistical predictor. After the study, documented tasks were grouped into development-heavy, collaboration-heavy, or other activities.

GenAI use and task type both track cognitive load

GenAI use, task category, and task duration each significantly contributed to explaining perceived cognitive load. Tasks involving GenAI use were associated with an estimated 0.996-point higher cognitive load rating on the seven-point scale than tasks without GenAI use (χ²(1)=23.80, p<0.001, ΔR²=0.047). Perceived load also rose by 0.007 points for every additional minute of task duration (χ²(1)=15.21, p<0.001, ΔR²=0.038).

Task category added the largest explanatory value (χ²(2)=44.13, p<0.001, ΔR²=0.091). Development-heavy tasks were rated significantly higher than collaboration-heavy tasks (estimate=0.591, p=0.0153) and other activities (estimate=1.437, p<0.0001); collaboration-heavy tasks were also rated higher than other activities (estimate=0.845, p<0.0001). Together the three fixed effects accounted for 32.4% of the variance in perceived cognitive load, rising to 45.9% when participant-specific differences were included.

These are associations within a small field sample, not evidence that GenAI use causes extra mental demand. The direction is also worth stating plainly: in this cohort, GenAI-supported tasks carried higher perceived load, not lower. Because task category and duration were modelled alongside GenAI use, the GenAI estimate is adjusted for those factors, but unmeasured differences between tasks could still matter.

Wearable physiology adds little beyond work context

The study tested eleven physiological metrics one at a time, each added to a baseline model already containing GenAI use, task category, task duration, and participant-specific differences. After Benjamini-Hochberg correction across the eleven metrics, only the standard deviation of skin temperature significantly improved model fit: lower skin temperature variability was associated with higher perceived cognitive load (estimate=-0.224, pBH=0.041). The added explanatory contribution was small (ΔR²=0.014). Skin temperature range showed a similar negative association but did not remain significant after correction (pBH=0.105).

A multimodal model combined a cardiovascular principal component, an electrodermal principal component, mean skin temperature, and skin temperature standard deviation. Against the contextual baseline, the full physiological block was nearly significant (χ²(4)=9.42, p=0.052) with a small gain (ΔR²=0.019). Leave-one-modality-out comparisons located most of that information in skin temperature: removing the two temperature metrics significantly reduced fit (χ²(2)=8.87, pBH=0.035, ΔR²=0.018), whereas the cardiovascular component (pBH=0.844) and electrodermal component (pBH=0.379) did not. A follow-up showed the temperature result was driven by standard deviation (estimate=-0.251, pBH=0.006, ΔR²=0.018), not mean temperature.

Sensitivity analyses using cumulative link mixed models, which respect the ordinal nature of the seven-point ratings, broadly agreed. The full physiological model became significant (χ²(4)=10.89, p=0.028) and skin temperature standard deviation remained significant (estimate=-0.396, pBH=0.002). But when the seven-point scale was collapsed into low, medium, and high categories, skin temperature standard deviation no longer survived correction (pBH=0.102). That fragility is a reason to treat the physiological signal as marginal rather than settled.

Context explains more than the wristband does

The comparison that matters is between the two model families. Work-context factors reached a marginal R² of 0.324 and a conditional R² of 0.459, while the best individual physiological addition, skin temperature standard deviation, added ΔR²=0.014 and the full multimodal block added ΔR²=0.019. In this dataset, knowing whether a task used GenAI, what kind of task it was, and how long it lasted told you more about perceived load than the physiological recordings did.

The authors interpret this as a reason to evaluate perceived cognitive load in relation to the concrete work context, using wearable data as complementary rather than standalone information. That reading fits the numbers: the physiological associations point in a consistent direction across sensitivity analyses, but their explanatory share is small and, for the individual metric, sensitive to how the rating scale is grouped.

The study also cannot say why GenAI tasks were rated higher. It measured whether tools were used, not how they were used, which tool was used for which task, or how helpful the tool was perceived to be; the open-ended helpfulness question is documented but not analysed as a predictor in the results reported here. The higher ratings could reflect more demanding tasks attracting GenAI use, the effort of prompting and checking output, or other unmeasured factors. The paper's evidence supports the association, not a causal story.

A small industrial sample with a narrow scope

The findings come from 21 developers at two SAP sites in California over four days in Q2 2025, all with prior Java experience and regular Java use, and most reporting 6 to 12 months of GenAI experience. The controlled sessions with predefined tasks on the first and last study days are outside this paper's scope. The task categories were assigned after the study by the authors of a related work, based on existing categorisation schemes, so category boundaries are an analytical choice rather than something participants reported.

These scope conditions limit transfer. The estimates describe this cohort's self-reported load during real work, not developers generally, not other organisations, and not long-term effects of GenAI adoption. The physiological results are likewise specific to the EmbracePlus metrics and preprocessing pipeline used, which retained the most working tasks among comparable pipelines. Within those limits, the paper's contribution is a measured comparison: work context, including GenAI use, carried the explainable signal, and wristband physiology did not.

Judge GenAI workflows by task context first

For anyone evaluating an AI-supported workflow, the paper's priority is to record the work context alongside any load measure. GenAI use, task category, and task duration were each independently associated with perceived cognitive load, and task category contributed the most. A dashboard that reports only adoption or output metrics would miss the difference between a development-heavy task and a collaboration-heavy one, and would miss the roughly one-point higher rating attached to GenAI-involving tasks in this sample.

Wearable physiology should be treated as a secondary channel. The only metric that survived correction in the individual analysis, skin temperature standard deviation, explained about 1.4% additional variance, and it lost significance when the rating scale was collapsed into three categories. The practical threshold implied by this study is not a physiological cut-off but a design choice: pair subjective load ratings with task-level context, and use physiological data only to complement that record, not to replace it.

Sources used in this answer