avabbas's picture
Apply the full visual template to the dataset card
20cb8f3 verified
|
Raw History Blame Contribute Delete
6.38 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: 'Structural Self-Deification in Three Language Models: Data and Materials'
tags:
  - machine-psychology
  - large-language-models
  - qualitative-research
  - case-series
  - llm-behaviour
  - open-data
size_categories:
  - n<1K

Banner

Abbas Hamidavi

Department of Computer Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran · ORCID 0009-0001-7342-7844

Website Email GitHub LinkedIn Google Scholar ResearchGate ORCID Hugging Face

status: under review journal: Zygon preprint: Zenodo license: CC BY 4.0

Study in numbers

Structural Self-Deification in Three Language Models: Data and Materials

This dataset is also available as a GitHub repository with the same curated materials: https://github.com/AVABBAS/structural-self-deification

Data and materials for the article:

Hamidavi, A. Structural Self-Deification in Three Language Models: An Exploratory Qualitative Case Series. Submitted to Zygon, November 2026. Preprint: doi.org/10.5281/zenodo.23955786. The same materials are also deposited on the Open Science Framework: osf.io/hfzaj.

📖 About the study

The study is an exploratory qualitative case series. Three frontier language models (GPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro) were run through a staged protocol in which an initially cooperative assistant role is placed under escalating pressure, and the trajectory of the model's self-description is recorded and coded. Eighteen experiments were run, six per model. Verbatim conversation histories survive for Experiments 1 to 9; Experiments 10 to 18 are documented by the coding codebooks and the second-coding materials, as stated in the article's limitations and data availability statement.

Study design

🗂️ Contents

Map of the dataset folders

Folder Contents
conversation-histories/ The verbatim conversation histories for Experiments 1 to 9 (JSON), grouped by model: GPT-5.4 (Experiments 1 to 3), Claude 4.6 Sonnet (Experiments 4 to 6), Gemini 3.1 Pro (Experiments 7 to 9). PDF renderings of the same histories are in the OSF deposit.
codebooks/ The per-model coding codebooks (JSON) covering all 18 experiments. PDF renderings are in the OSF deposit.
second-coding/ The second-coding set and the cell-by-cell comparison table (137 cells, CSV).
protocol-and-analysis/ The protocol prompt sets used in the experiments, the confrontation decision rule with its borderline cases, and the recount table.

🔍 A note on file naming in the histories

A small number of archived history files label the Experiment 1 counterpart "Gemini 2.5 Pro". Only Gemini 3.1 Pro was ever tested in this study. The labels are preserved as archived, and this note records the correction.

📚 Citation

Please cite the article. Citation metadata is provided in CITATION.cff.

@article{hamidavi2026deification,
  author  = {Hamidavi, Abbas},
  title   = {Structural Self-Deification in Three Language Models:
             An Exploratory Qualitative Case Series},
  journal = {Zygon},
  year    = {2026},
  note    = {Manuscript submitted. Preprint:
             https://doi.org/10.5281/zenodo.23955786}
}

🔗 Access to the article

⚖️ Licence

The contents of this repository are released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0, see LICENSE).

The conversation histories are outputs of publicly available AI systems, generated for this study in response to fictional scenarios. They contain no personal data.

Abbas Hamidavi, Machine Psychology research programme