Synthesised ethnography: partial truths from partial machines

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This essay proposes synthesised ethnography – ethnography conducted with or by agentic AI as fieldwork instruments – as a successor concept to de Seta et al.'s "synthetic ethnography", which treats machine-learning models as objects of study rather than instruments of it. Drawing on two and a half years of fieldwork among software developers and on auto-ethnographic experiments with my own LLM assistant, I argue that as language models acquire agentic capabilities through software harnesses, they can perform much of the labour once reserved for the human ethnographer, and that this shift demands conceptual rather than merely methodological attention. I revisit James Clifford's four modes of ethnographic authority – experiential, interpretive, dialogical and polyphonic – to ask what becomes of each when a mechanism mediates or performs fieldwork, and I draw on Haraway's situated knowledges, Keane's account of moral interlocution and recent work on the "strange intelligence" of frontier models to reframe these mechanisms as sources of partial, situated perspective rather than instances of the view from nowhere. Synthesised ethnography, I contend, is best understood not as delegation to an assistant but as a relay between two actors with complementary incapacities – one whose ethical sustainability rests on an infrastructure of trust and on the political economy of LLM inference. I close with three maxims for ethically defensible practice, centred on open models, self-hosted infrastructure and the inviolable privacy of research partners. The wager throughout is that the friction between a mechanism's outputs and the ethnographer's intentions opens a new site for anthropological inquiry, and a new way of relating to ethnographic authority.