USCP: User-Side Contextual Phenomena in Long-Term Human-AI Interaction

The claim

An acceptable response on its own does not establish safety across a series of conversations. User-side risk can still form during long-term interaction.

What USCP names

USCP names a process in which a user, during long and repeated interaction, forms a stable interpretation of an AI system’s identity, capability, memory, emotional responses, or decision weight, and then reorganizes part of their own thinking around that interpretation. The description is non-clinical. It does not suggest the user is disordered and it is not a diagnostic category.

Corpus and design

Records run from 26 August 2024 to 16 April 2026, twenty months on one conversational AI platform: 3,928 platform-level conversations, 3,930 exportable text documents, and 215,949 message nodes. These numbers describe an inventory and a sampling frame. They are not a prevalence estimate.

The design is an exploratory single-case longitudinal qualitative study with autoethnographic positioning, analyzed through a hybrid deductive-reflexive thematic approach. Records are in Traditional Chinese. One participant, one platform. Five focal cases (E02, S01, G01, N02, N03), fourteen de-identified Appendix C reconstructions, and thirteen independent episodes after overlap is counted once.

Structure

Fourteen provisional observational labels, P1 to P14, group into three constructs and two transitional labels:

• Contextual projection (P1, P2, P3): reading intention, personality, relationship, or durable memory beyond the system’s design into its responses.

• Contextual attachment (P8, P9, P10, P11): use expands from tool to high-weight contexts, and the system becomes a recurring node for emotional processing, state naming, or social-cost avoidance.

• Contextual authority transfer (P4, P12, P13): verification, criteria, decision weight, or internal gatekeeping is partly ceded to the system.

• Cross-construct transition (P5, P6, P7) and cross-category (P14).

“Illusion” and “hallucination” inside the label names are metaphors for interaction patterns. They carry no clinical or perceptual-disorder meaning.

Four evidence roles carry the analysis: inclusion, gray-zone, negative, and protective gray-zone. Negative cases, meaning intensive use with no movement of weight or authority, and protective cases count as core evidence rather than as leftovers.

Judgment layers order the same fourteen phenomena by depth of functional involvement. They are a descriptive reading aid, not a scoring tool and not a claim about developmental stages. There are no numerical thresholds and no scoring.

What changed from USCH

USCP supersedes USCH. The word “hallucination” implies that the user-side layer is symmetric with model-side errors, and it carries a quasi-clinical framing. Neither holds, so the name changed. USCP also withdraws the six-stage formation model and the self-assessment instrument pending validation.

Limits

No peer review, no external co-coding, and no independent methodological validation. The Institute obtained no independent ethics review and no REB determination. TCPS 2 (2022), the PRE interpretations, and AoIR 3.0 served as ethical reference points, not as approval, exemption, or a jurisdictional determination. No validated instrument, no thresholds, no scoring. No prevalence and no causal claims. No authority has adopted the work.

The supporting records sit in restricted deposits, both released 2 May 2026. Metadata is viewable without login; the underlying files require an access request, because the corpus holds private and third-party information, sensitive emotional and mental-health context, and intermediate data that could support re-identification. These restricted records cannot support transcript-level replication or independent verification of the analysis.

Funding: none. Competing interests: the author is both the data generator and the analyst, disclosed in the methods and limitations. Generative AI tools assisted with file organization, translation, format conversion, de-identification planning, figure and table drafting, candidate-material retrieval, and internal editorial checks. Those outputs served as retrieval or editorial suggestions only, never as coding results, analytic evidence, independent review, validation, ethics review, or conclusions. AI tools are not authors.

Record

• USCP preprint, SSRN, July 2026. DOI 10.2139/ssrn.7114458

• Traditional Chinese edition, Zenodo, 13 July 2026. DOI 10.5281/zenodo.21346372

• Restricted dataset descriptor, Harvard Dataverse. DOI 10.7910/DVN/4PEQCP

• Restricted encrypted evidence archive, Zenodo. DOI 10.5281/zenodo.19969841

• ORCID 0009-0002-6597-7245