Research Basis / Six frameworks
August 2024 to April 2026

A vocabulary for user side risk in long term AI dialogue.

The corpus is a 20-month first-person longitudinal record: 3,928 conversations and 215,949 message nodes. The method is an exploratory single-case longitudinal qualitative design with a hybrid deductive-reflexive thematic approach.

02
Method

Three methodological commitments

Longitudinal record
20 months of complete conversation records, August 2024 to April 2026.
Phenomenon annotation
Systematic annotation: fourteen phenomena, three constructs, two transition labels.
Evidence roles
Four evidence roles: inclusion, gray-zone, negative, protective gray-zone. Negative cases define what does not count as risk, so ordinary tool use is not pathologized.
03
Tooling

A-CSM: AI Contextual Signal Matrix

Pipeline: rule-based triage, a three-family LLM judge panel, deterministic citation alignment, deterministic aggregation, USCI hard rules, and a PDF/A-3b report. The report embeds the de-identified source, each judge's raw output and a SHA-256 manifest, so a third party can re-verify from the report alone.

Status: an executable research prototype. The public core is open source and under active development.

github.com/kyozong77/A-CSM-public-core

04
Evidence level

Evidence level statement

This research is currently a descriptive phenomenon framework built on a single first-person longitudinal corpus. It offers observable phenomenon categories and a reading method. It has not been validated across samples and is not a diagnostic instrument.

The USCH six-stage model and self-assessment scale have been retracted and are no longer cited on this site. See Research Boundaries.

05
Terminology

Research vocabulary and service vocabulary

Contextual projection
Written as “projection” on service pages. Definitions always follow the research pages.
Contextual attachment
Written as “attachment” on service pages.
Contextual authority transfer
Written as “authority transfer” on service pages.
Signal vocabulary
“False reassurance”, “verification drift” and “boundary loss” are service-side signal descriptions of observable patterns under the constructs above.
06
Official registers

The risk is already on official lists

NIST AI 600-1
July 2024. Human-AI Configuration is one of twelve named risks.
MIT AI Risk Repository
arXiv:2408.12622. Human-computer interaction is one of seven domains.
International AI Safety Report 2026
Contains a dedicated section on human autonomy.
Taiwan Ministry of Digital Affairs
AI risk classification framework (2026-07-07): B1 over-reliance, B2 loss of human autonomy.

This mapping locates the risk. It is not an endorsement by any institution. The lineage goes back to Parasuraman and Riley (1997). The contribution here is evidence, method and tooling.

07
Datasets

Datasets and archiving (not papers)

Harvard Dataverse
Zenodo dataset
10.5281/zenodo.19969842 (restricted, encrypted evidence files)
Priority notice
10.5281/zenodo.19490172 (neutral timestamp record, type Other)

These are dataset and archiving records, listed separately from the papers.

Next

The full publication record and citation formats are on the Work page.

See the work