User-Side Contextual Phenomena (USCP).
Research into the cognitive, emotional, and relational patterns that emerge in users through long-term AI interaction — even when individual AI responses appear safe and technically accurate.
/
The Problem
(01)
One response at a time looks fine. Across six months, the pattern is different.
Standard AI safety evaluation works at the level of individual outputs: is this response accurate, safe, appropriate? What USCP addresses is what happens across the full span of long-term interaction. Users come to treat AI systems as entities that remember, care, judge, and carry authority. None of this requires any individual response to be wrong.
The research started from a specific moment: ZON RZVN was personally led to a dangerous place by AI responses during a period of acute psychological crisis, recognized the risk, and started documenting what had happened. That experience became the first research question.
/
Research Design
(02)
3,928 conversations. One researcher. August 2024 to April 2026.
USCP is grounded in a longitudinal self-corpus: approximately 3,928 conversation segments and 215,949 message nodes spanning 20 months of sustained AI interaction. Methodology: analytic autoethnography and reflexive thematic analysis.
Three main patterns: contextual projection (users attributing stable traits, memory, and intent to AI systems), contextual attachment (emotional and relational dependency forming through repeated interaction), and contextual authority transfer (users deferring judgment and decision-making to AI responses over time).
/
Framework
(03)
A non-clinical vocabulary for the layer safety evaluations currently miss.
The framework offers a reusable vocabulary for describing user-side contextual formation in long-term AI interaction. Four evidence roles: included cases, grey-zone cases, negative cases (no drift present), and protective grey-zone cases. Treating negative cases as core evidence rather than afterthought is central to the methodology.
USCP is non-clinical. It does not diagnose, screen, or estimate prevalence. Risk and drift refer to interaction patterns, not clinical categories.