Ethnographic research into developers’ modelling practices constitutes an emergent field of investigation. Methodological strategies to foreground this often-hidden labor become necessary to unveil how models take shape. To address this need, we suggest considering two moments as methodological entry points. First, transfer learning, that is, routine adaptation of general-purpose pre-trained models to specific requirements. By examining what developers (out of necessity) choose not to modify, alongside the minute technical features they do alter during such model tailoring, an infrastructural map of how things are tied together can be drawn. Second, what we define as the detection of shortcut learning, i.e., developers’ struggle to discern when a generative model is diverging from its expected path and to identify the specific patterns it has prioritized to avoid embedding these diversions in subsequent upgrades. As developers catch the model’s spurious correlations, the configuration of elements and the distribution of tasks through which synthetic data emerge become visible. Finally, these two methodological entry points suggest that ethnography in situ allows to observe this invisible work and the composite formation through which what is called universality is constructed and reconstructed.
Presented in panel P012 “The matter of method in researching AI: elusiveness, scale, opacity” (convenors: Claudia Aradau, Tobias Blanke, Annalisa Pelizza), session 3, Friday 11 September 2026, 9:00–10:30, room C-7 1.09.
The title in the EASST2026 programme is “Learning work: Toward an ethnography of AI”; the slides are titled “The invisible work behind universal models”. The abstract above is the paper long abstract as published in the programme.
hosting event
