When AI outputs become documents: Documentation activity in human– AI dialogue

Abstract Large language models (LLMs) generate texts that increasingly circulate as documents in knowledge infrastructures, yet their documentary status remains theoretically underdetermined. Unlike traditional documents, LLM outputs lack identifiable authorship, stable provenance, or testimonial grounding. This challenges foundational assumptions in document theory about authority, accountability, and evidentiary value. This study investigates how AI‐generated text acquires documentary force through situated use. Through reflexive case study analysis of an extended ChatGPT dialogue, I apply systematic thematic coding guided by the Model of Documentation Activity (MoDA) to trace documentation activity across physical, mental, and social dimensions. The analysis advances three contributions to document theory and AI governance. First, I demonstrate the analytic utility of the concept of artificially blended testimony (ABT) for examining how LLM outputs provisionally stabilize as documentary artifacts despite lacking testimonial grounding. Second, I show how AI outputs acquire documentary status through human practices of framing boundaries, establishing form, and attributing responsibility. Third, I reconceptualize human oversight mandates in AI governance as documentation thresholds, identifying the documentation practices required for AI outputs to function as preservable, citable, and accountable documents within information infrastructures.