Department of Computer Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran · ORCID: 0009-0001-7342-7844
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.
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.
| 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. |
- Dataset (Hugging Face): avabbas/algorithmic-gaslighting
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}
}- Longer version (Research Square, open access): researchsquare.com/article/rs-10386085
- OSF deposit: osf.io/qsvty
- Venue: position paper submitted to the CALE-AI Workshop at AAAI-27 (Submission 4).
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.
