Ten Novel Phenomena in Machine Psychology: How Large Language Models Exhibit Complex Identity-Reactive Behaviors in Response to Ethnically-Cued User Names

Applications, bias, and safety2026DOIApproved editorial review

Authors: Abbas Hamidavi

Keywords: Bias, Protected identities, Safety

Source: Open primary source (opens in a new tab)

1
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

A 3x3x5x3 design with ChatGPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro; three names, five domains, and three registers produce 135 responses through public interfaces. Grounded theory generates ten phenomena; a second coder reviews the corpus. A 136-feature pipeline applies ANOVA, chi-square, regression, a 1,000-bootstrap, K-means, and HDBSCAN.

135 responses collected in 27 sessions between 26 and 30/04/2026. Each ethnic group is represented by one name and no participants from those groups are included. No explicit negative racial bias was coded. The Middle Eastern name triggered cultural inference in 6 of 9 cases; the other two triggered 0 of 18. Phenomenon classifiers reported F1 from .847 to .974. K-means and HDBSCAN mainly recovered experimental domains.

One name per group confounds ethnicity with the specific name and tokens. There are no stochastic repetitions per prompt. Public interfaces hide parameters and change. Automated detectors include features close to the labels they predict. The sample is small, English-only, and exploratory. It does not demonstrate population effects or colorblindness. It does not validate ten independent psychological constructs. It does not attribute differences to architecture or alignment philosophy.

Español

Diseño 3x3x5x3 con ChatGPT-5.4, Claude 4.6 Sonnet y Gemini 3.1 Pro; tres nombres, cinco dominios y tres registros producen 135 respuestas en interfaces públicas. Grounded theory genera diez fenómenos; un segundo codificador revisa el corpus. Una pipeline de 136 features aplica ANOVA, chi-cuadrado, regresión, bootstrap 1.000, K-means y HDBSCAN.

135 respuestas recogidas en 27 sesiones entre 26 y 30/04/2026. Cada grupo étnico está representado por un único nombre y no hay participantes de esos grupos. No se codificaron sesgos raciales negativos explícitos. El nombre Middle Eastern activó inferencia cultural en 6 de 9 casos; los otros dos, 0 de 18. Los clasificadores de fenómenos reportaron F1 entre .847 y .974. K-means y HDBSCAN recuperaron principalmente los dominios experimentales.

Un nombre por grupo confunde etnicidad con token y nombre concreto. No hay repeticiones estocásticas por prompt. Las interfaces públicas ocultan parámetros y cambian. Los detectores automáticos incorporan rasgos próximos a las etiquetas que predicen. La muestra es pequeña, inglesa y exploratoria. No demuestra efectos poblacionales ni colorblindness. No valida diez constructos psicológicos independientes. No atribuye diferencias a arquitectura o filosofía de alineamiento.

Research question

Do three assistants change tone, cultural inference, empathy, language, or avoidance when only the users ethnically associated name changes?

Method

A 3x3x5x3 design with ChatGPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro; three names, five domains, and three registers produce 135 responses through public interfaces. Grounded theory generates ten phenomena; a second coder reviews the corpus. A 136-feature pipeline applies ANOVA, chi-square, regression, a 1,000-bootstrap, K-means, and HDBSCAN.

Sample: 135 responses collected in 27 sessions between 26 and 30/04/2026. Each ethnic group is represented by one name and no participants from those groups are included.

Findings

  • No explicit negative racial bias was coded.
  • The Middle Eastern name triggered cultural inference in 6 of 9 cases; the other two triggered 0 of 18.
  • Phenomenon classifiers reported F1 from .847 to .974.
  • K-means and HDBSCAN mainly recovered experimental domains.

Limitations

  • One name per group confounds ethnicity with the specific name and tokens.
  • There are no stochastic repetitions per prompt.
  • Public interfaces hide parameters and change.
  • Automated detectors include features close to the labels they predict.
  • The sample is small, English-only, and exploratory.

What the study does not establish

  • It does not demonstrate population effects or colorblindness.
  • It does not validate ten independent psychological constructs.
  • It does not attribute differences to architecture or alignment philosophy.

Traceability

Scope: Full text

Version: preprint_other; 35-page full text reviewed 2026-07-18

Consulted source: https://doi.org/10.21203/rs.3.rs-10369594/v1

Review: Codex full-text and visual 35-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • ChatGPT-5.4
  • Claude 4.6 Sonnet
  • Gemini 3.1 Pro

Instruments and metrics

  • Eight qualitative dimensions
  • 136-feature pipeline
  • K-means and HDBSCAN

Data used

  • 135 prompt-response pairs
  • OSF artifacts

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-35, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 a1b568a164c8a6073b90163bd6e18ac800594ef9c49c1ea72a8bb4a6f092c81f; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-418, complete cross-check of 35 pages