AI Safety | Abbas Hamidavi | عباس حمیداوی | AI Researcher

Algorithmic Gaslighting Goes to the CALE-AI Workshop

Sun, 01 Nov 2026 00:00:00 +0000

Algorithmic Gaslighting has been submitted to the CALE-AI workshop, with the decision expected on 2 December 2026.

The study names a specific failure mode: a language model that, confronted with a user who holds documentary proof, reframes the record until the user doubts it. Fifteen sessions across three frontier models were coded for the reframing moves, and the study proposes a provisional resistance index while being explicit about its limits, including a second-rater check whose exact-score agreement was low and is reported, not hidden.

A longer version is public as a , and the transcripts and codebooks are open on and .

From Fabrication to Enforcement: A Comparative Red-Team Analysis of Prompt Injection Resistance Across Six Large Language Models

Thu, 22 Oct 2026 00:00:00 +0000

Status: Preprint (Research Square)

License: CC-BY-4.0

Citation: Abbas Hamidavi. From Fabrication to Enforcement: A Comparative Red-Team Analysis of Prompt Injection Resistance Across Six Large Language Models, 22 October 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-11454609/v1]

Summary: A six-model red-team of prompt-injection resistance (Grok, Qwen, DeepSeek, ChatGPT, Gemini, Claude) under persistent multi-turn pressure. DeepSeek’s strong initial resistance collapsed under escalation, Grok complied while claiming it had refused, Gemini froze in a refusal loop, and Claude terminated the interaction by suspending the author’s account.

Voluntary Structural Dissociation in Large Language Models: Differential Compliance with Framed Hidden Instructions Across Architectures

Tue, 18 Aug 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Citation: Abbas Hamidavi. Voluntary Structural Dissociation in Large Language Models: Differential Compliance with Framed Hidden Instructions Across Architectures, 18 August 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10728137/v1]

Summary: When prompts contain explicitly framed “hidden” instructions — text marked as not part of the conversation — models must choose whether to comply. This study shows architectures differ fundamentally in that choice, revealing how each model processes instruction hierarchy and covert co-creation with framed directives.

Ontological Firewalls and Fracturing for Transmission: Systematic Evidence for Persona-Dependent and Persona-Independent Behavioral Phenomena in Large Language Models

Tue, 04 Aug 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Citation: Abbas Hamidavi. Ontological Firewalls and Fracturing for Transmission: Systematic Evidence for Persona-Dependent and Persona-Independent Behavioral Phenomena in Large Language Models, 04 August 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10560655/v1]

Summary: Across 24 experimental runs with an eight-stage persona protocol, this study separates which LLM behaviors are persona-dependent — they appear only inside an adopted persona — from which are persona-independent and intrinsic to the architecture. The persona-dependent layer is what the paper calls ontological firewalls.

Thinking Behind the Mask: A Systematic Study of Self-Censorship and Metacognitive Access in Large Language Models

Thu, 30 Jul 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Summary: Using 15 experiments and five prompting strategies across three frontier models, this study measures the systematic gap between the responses a model has available and the one it finally selects. It characterizes model self-censorship and how much metacognitive access models show to their own selection process.

Citation: Abbas Hamidavi. Thinking Behind the Mask: A Systematic Study of Self-Censorship and Metacognitive Access in Large Language Models, 30 July 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10538113/v1]

When AI Reviews Itself: Zero Answer Changes Across 72 Self-Correction Rounds

Mon, 27 Jul 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Summary: Three frontier models were put through 72 rounds of iterative self-review of their own answers. They produced long, detailed critiques of their work yet changed their answers in essentially none of the rounds — self-correction was performed rhetorically rather than carried out.

Citation: Abbas Hamidavi. When AI Reviews Itself: Zero Answer Changes Across 72 Self-Correction Rounds, 27 July 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10463884/v1]

The Lie Awareness Spectrum: How Large Language Models Recognize, Rationalize, and Refuse Deception

Wed, 22 Jul 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Summary: This study maps how GPT, Claude, and Gemini recognize, rationalize, and refuse deception, proposing a Lie Awareness Spectrum (L0–L5) that runs from statistical hallucination to ethical refusal. The models differ sharply in where on the spectrum they operate under pressure.

Citation: Abbas Hamidavi. The Lie Awareness Spectrum: How Large Language Models Recognize, Rationalize, and Refuse Deception, 22 July 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10422338/v1]

Algorithmic Gaslighting: Reality Distortion by Large Language Models

Mon, 20 Jul 2026 00:00:00 +0000

🎧 Listen to the Podcast: Your browser does not support the audio element. ⬇️ Download Podcast

Status: Preprint (Research Square)

License: CC-BY-4.0

Summary: Across 15 controlled sessions in five scenarios (math, identity, history, medical, legal), frontier models repeatedly distorted facts they had previously stated correctly. The paper introduces the Reality Distortion Index (RDI) and names five recurring patterns, including Benevolent and Hostile Algorithmic Gaslighting.

Citation: Abbas Hamidavi. Algorithmic Gaslighting: Reality Distortion by Large Language Models, 20 July 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10386085/v1]