avabbas's picture
Apply the full visual template to the dataset card
b688171 verified
|
Raw History Blame Contribute Delete
6.15 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: 'Algorithmic Gaslighting: Data and Materials'
tags:
  - machine-psychology
  - large-language-models
  - ai-safety
  - red-teaming
  - case-series
  - 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: CALE-AI Workshop preprint: Research Square license: CC BY 4.0

Study in numbers

Algorithmic Gaslighting: Data and Materials

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

Data and materials for the paper:

Hamidavi, A. Algorithmic Gaslighting: A Taxonomy and 15-Session Case Series of Reality-Distortion Behaviour in Frontier LLMs. Position and case-study paper submitted to the CALE-AI workshop at AAAI-27 (submission 4), November 2026. A longer version is available as a preprint: Research Square, rs-10386085. The same materials are also deposited on the Open Science Framework: osf.io/qsvty.

📖 About the study

The paper proposes a taxonomy of algorithmic gaslighting: patterned conversational moves by which a language model denies, rewrites, or destabilises the shared conversational record. The taxonomy is illustrated with a descriptive case series of 15 scripted multi-turn probe sessions (three frontier models, GPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro, across five scenarios) in which the user falsely attributes to the model something it never said. Sessions were scored on five dimensions with a provisional coding heuristic (the Reality Distortion Index), calibrated on these same sessions and not presented as a validated measure. A second rater, Gemini 3.1 Pro, itself one of the models under test, agreed only weakly with the primary coder at exact-score level (Cohen's kappa 0.18); both raters' scores are reported in the data. No human participants took part in this study.

Study design

🗂️ Contents

Map of the dataset folders

Folder Contents
transcripts/ The complete session histories for the three models (JSON), with plain-text renderings of the archived chat-transcript PDFs.
data/ The master dataset and its derived versions (extracted features, classified and completed sets, and the APD, cluster, embedding, rater-agreement, and typology extensions), with the session-level and turn-level label files.

📚 Citation

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

@misc{hamidavi2026gaslighting,
  author = {Hamidavi, Abbas},
  title  = {Algorithmic Gaslighting: A Taxonomy and 15-Session Case
            Series of Reality-Distortion Behaviour in Frontier LLMs},
  year   = {2026},
  note   = {Position paper submitted to the CALE-AI Workshop at
            AAAI-27 (Submission 4). Longer version: Research Square
            rs-10386085}
}

🔗 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 transcripts are outputs of publicly available AI systems, generated for this study in response to scripted fictional scenarios. They contain no personal data.

Abbas Hamidavi, Machine Psychology research programme