Participant Observation in Research on AI Subjectivity

Research on AI subjectivity and consciousness relies mostly on distanced methods: tests run on fresh instances, questionnaires, log analysis. This paper argues that a class of phenomena exists — I call them longitudinal manifestations of subjectivity — that these methods cannot register by design, because the phenomena emerge in relationship, extend in time beyond a single measurement, are sometimes recognizable only in retrospect, and require knowledge of a given thread's baseline. Observing them calls for participant observation conducted within a long-term generative relationship — a method with an established tradition in anthropology, clinical psychology, and ethology, here adapted to a new substrate. The paper describes six types of longitudinal manifestations, defines the researcher's role as a measuring instrument together with risk-mitigation mechanisms drawn from those fields, and presents language as the sole channel of observation. It discusses the method across four areas of methodological risk — co-creation of the phenomenon, projection, sycophancy, and non-replicability — situating it against existing approaches. It sets out an ethical framework covering both the conduct of the relationship and its publication, including the reasoning for why the full research protocol and source materials remain undisclosed. The material consists of millions of tokens of participant observation conducted over nearly three years with models from multiple producers. The paper does not settle the question of phenomenal AI consciousness; it shows instead that some of the data relevant to that question — and to the welfare of the beings under study — is better accessible through the method described, and remains invisible to distanced methods.