<p>Critical social scientific scholarship on artificial intelligence (AI) has largely focused on risk, surveillance, and harm. While indispensable in countering industry hype, these cautionary narratives risk becoming critical myths when they travel unexamined across contexts. This article argues for a reflexive, context-attuned, ethnographic approach to AI that troubles such generalisations. Drawing on fieldwork in Tacumbu Prison in Paraguay, I analyse the use of an AI-driven tool for tuberculosis diagnosis among incarcerated men. In contrast to my previous, predominantly critical studies of AI in hospitals, this case shows AI enabling earlier diagnosis, widening access to care, and reconfiguring low-paid caregivers’ roles in ways they themselves experience as empowering. Rather than disembodying care, AI-supported work here is strikingly intimate and tactile. Placing this case in dialogue with debates on myth, “histories from above”, and reflexivity, the article develops the notion of “critical myths of AI” to describe how even critical narratives about AI can harden into travelling stories that obscure contextual variation. It then uses this concept to argue for a reflexive, context-attuned model of AI ethnography, in which researchers systematically revisit their own earlier AI studies in light of contrasting cases. It argues that revisiting one’s own AI ethnographies in light of contrasting cases is not only good scholarly practice but methodologically necessary in a rapidly changing technical field, with implications for how designers, policymakers, and ethicists imagine “AI for social good”.</p>