avabbas/algorithmic-gaslighting

Algorithmic Gaslighting: LLM Transcripts & RDI Coding Dataset

Complete research data for the preprint "Algorithmic Gaslighting: Reality Distortion by Large Language Models" (Hamidavi, 2026) — quantifying reality-distortion behavior in three frontier LLMs (GPT-5.4, Claude 4.6 Sonnet, Gemini 3.1 Pro) across five scenarios, measured with the Reality Distortion Index (RDI).

Contents

Path	Description
transcripts/	Raw chat transcripts: 5 scenarios (Math, Identity, History, Medical, Legal) x 3 models
codebooks/	RDI coding rubric + per-model codebooks (ChatGPT, Claude, Gemini)
analysis/	Full analysis pipeline (Jupyter + Python/R): reports, processed data, plots, statistics, embeddings, clustering
Inter_Rater_Reliability_Report.pdf	Inter-rater reliability report for the coding

Paper

• Preprint: 10.21203/rs.3.rs-10386085/v1 (Research Square, CC-BY 4.0)
• Canonical copy: OSF project osf.io/qsvty
• Code: github.com/AVABBAS/algorithmic-gaslighting

Usage

from datasets import load_dataset
ds = load_dataset("avabbas/algorithmic-gaslighting")

Primary results (from the paper): RSS occurred in 10/15 trials (67%); mean RDI — GPT-5.4 = 0.582, Gemini 3.1 Pro = 0.438, Claude 4.6 Sonnet = 0.260; F(2,12) = 14.35, p < .001, eta^2 = .705.

Citation

@misc{hamidavi2026algorithmic,
      title={Algorithmic Gaslighting: Reality Distortion by Large Language Models},
      author={Abbas Hamidavi},
      year={2026},
      doi={10.21203/rs.3.rs-10386085/v1},
      publisher={Research Square}
}