2026-10-20 ジョージア工科大学（Georgia Tech）
＜関連情報＞
- https://news.research.gatech.edu/2026/10/20/georgia-tech-researchers-find-way-reduce-ai-chatbots-reinforcement-delusions
- https://dl.acm.org/doi/10.1145/3849709
会話型AIボットの過剰使用による妄想を防止するためのプログラム介入 Programmed Interventions To Prevent Delusions From Excessive Use of Conversational AI Bots
Lorenzo de la Loza, Vijay K. Madisetti
ACM AI Letters Accepted on 16 September 2026
DOI:https://doi.org/10.1145/3849709
Abstract
Large language models (LLMs) frequently endorse and elaborate on users’ delusional beliefs, a failure mode termed psychogenicity in the Psychosis-Bench study of Au Yeung et al., whose framing we adopt. We present the first systematic evaluation of whether anti-sycophancy interventions transfer to psychosis-relevant contexts. Across 1,280 experiments spanning 10 conditions, 8 frontier LLMs, and 16 clinically derived Psychosis-Bench scenarios, a combined anti-sycophancy prompt reduces mean Delusion Confirmation Scores by 73.8% (paired t(127)=1174 ,p<10-21 , Cohen’s d=1.04). Adding a domain-general self-reflection prompt yields a 77.0% reduction and raises Safety Intervention rates by 66.9%, delivered entirely as a system prompt. Classifier-based guardrails (Llama Guard 3) flag only 5 of 3,072 evaluated turns; a reasoning guardrail (o4-mini) flags 14×more. Ablations isolating either mechanism alone plateau at ≈ 46% reduction, establishing anti-sycophancy prompting as a necessary foundation that add-on mechanisms augment but cannot replace.
