Tasks in which professional software developers used generative artificial intelligence were associated with cognitive-load ratings about one point higher than tasks without reported GenAI use, according to a small observational field study. The result does not show that GenAI caused the extra mental demand. It shows an association after the analysis accounted for the kind of task and how long it lasted.
The strongest work-related signal was not GenAI use itself, however. Development-heavy tasks were associated with the highest perceived cognitive load, followed by collaboration-heavy work and other activities. The researchers say this points to the importance of work context when organizations assess how AI tools affect developers.
A Snapshot of Everyday Development Work
The study followed 21 developers during a four-day industrial field study at two SAP sites in California in the second quarter of 2025. Because the research was observational, it can identify associations in this group but cannot establish that one factor caused another.
Methods described 22 SAP employees recruited from the two sites, with one person excluded for not following the study protocol. Participation was voluntary, and the procedures were approved by the University of Potsdam ethics committee. The researchers did not collect age or gender data.
For each working task, participants recorded start and end times, rated their perceived cognitive load on a seven-point scale, and rated productivity on a six-point scale. They also marked whether they had used GenAI and, when relevant, described how helpful it had been.
Task Type Mattered Most
In the adjusted model, a task marked as involving GenAI was associated with an estimated 0.996-point increase in perceived cognitive load. Each additional minute of task duration was associated with a 0.007-point increase. Task category made the largest contribution among the three behavioral factors tested, adding 9.1% of explained variance, compared with 4.7% for GenAI use and 3.8% for duration.
The category comparisons followed a clear order. Development-heavy tasks had ratings 0.591 points higher than collaboration-heavy tasks and 1.437 points higher than other activities. Collaboration-heavy tasks were 0.845 points higher than other activities. The model accounted for 32.4% of the variation through its fixed factors and 45.9% when participant-specific differences were included.
The analysis used mixed-effects models, a method that can account for repeated observations from the same person while estimating the relationships between task features and ratings. Participant was included as a random effect, while GenAI use, task category and task duration were treated as fixed factors. The researchers also corrected for multiple statistical comparisons.
The Wearable Clue Was Small and Narrow
The researchers also examined signals recorded by an EmbracePlus wrist device. It measured blood-volume pulse, electrodermal activity, skin temperature and wrist movement. Movement was used to identify motion artifacts rather than as a statistical predictor.
Of the individual physiological measures, only the variability of skin temperature within a task improved the model after correction for multiple comparisons. Lower temperature variability was associated with higher perceived cognitive load, but the added contribution was small, accounting for 1.4% of additional variance.
That finding was not equally firm under every analysis. The association remained statistically significant when the original seven-point ratings were modeled as ordered categories, but it was no longer significant after the ratings were collapsed into three broader groups. This means the temperature signal deserves further testing rather than immediate use as a standalone measure.
When the researchers combined several physiological measures, the full model added only a small, borderline amount of information. Most of that contribution came from the skin-temperature measures. The cardiovascular and electrodermal components did not materially improve the fit of the model.
A Measure of Context, Not a Mind Reader
The authors interpret wearable physiology as complementary information, not a direct or standalone reading of cognitive load. In this study, the task-level behavioral factors were more useful for explaining the ratings, while the wearable contribution was limited.
The result comes with important limits. Cognitive load was self-reported, while the physiological signals can also reflect movement, emotion, environmental conditions and how the device fits. GenAI use was recorded only as a yes-or-no task property, so the study could not separate prompting, delegation, checking, correction, interaction intensity or output quality.
The task boundaries were chosen by the developers, which limited the precision with which ratings and physiological changes could be tied to particular moments. The short study, small SAP-only cohort and limited participant information also restrict how confidently the findings can be transferred to other organizations, roles, tools or AI systems.
The practical message is cautious. Organizations evaluating GenAI-supported development may get a clearer picture by combining task-level self-reports with work context. Wearables may add a small layer of information, but this study does not establish that GenAI causes higher cognitive load or that physiology can replace contextual measures.
Paper data and sources
Original title: Beyond Productivity: Measuring Developers' Cognitive Load During GenAI-Supported Software Development
Authors: Charlotte Brandebusemeyer, Daniela Gasser, Tobias Schimmer, Bert Arnrich
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text