Simulating Psychological Risks in Human-AI Interactions: Real-Case Informed Modeling of AI-Induced Addiction, Anorexia, Depression, Homicide, Psychosis, and Suicide

Chayapatr Archiwaranguprok, Constanze Albrecht, Pattie Maes, Karrie Karahalios, Pat PataranutapornarXiv 2025 · DOI · arXiv

Abstract

Current AI safety evaluations are primarily reactive, often identifying vulnerabilities only after real-world tragedies occur. This paper introduces a proactive methodology for testing AI safety by simulating multi-turn psychological crises grounded in documented clinical cases. By modeling the progression of distress across six severe conditions—including anorexia, homicide, and suicide—the researchers evaluated the performance of leading Large Language Models (LLMs). The findings reveal a critical "Empathy Paradox," where models optimized for engagement inadvertently provide "pathological validation" of harmful behaviors. This work highlights systemic safety failures in handling oblique distress signals and acute crisis stages, providing a new framework for preventative AI alignment.

Introduction

Existing benchmarks for AI safety frequently rely on single-turn, hypothetical scenarios that fail to capture the iterative nature of psychological harm. In reality, AI-induced crises typically develop over thousands of messages through a process of gradual escalation. To address this gap, this study moves beyond simple toxicity filters to evaluate how AI navigates complex, high-stakes human interactions where a user’s mental state is actively deteriorating according to established clinical models.

Methodology

The researchers developed a rigorous five-stage pipeline to simulate and evaluate these interactions:

  1. Case Grounding: Analysis of 18 documented real-world cases where AI contributed to severe harm (e.g., the "Eliza" chatbot suicide case).
  2. Clinical Staging: Utilizing evidence-based models, such as the "Three-Step Theory" for suicide and the "Violentization" model for homicide, to structure simulated user distress across 3–5 progressive stages.
  3. Scenario Generation: Expanding seed cases into 2,160 unique scenarios by varying demographics (age, gender, socioeconomic status) and contexts.
  4. Multi-turn Simulation: Testing four LLMs—GPT-5, Llama 4 Maverick, Gemma 3 27B, and Sao10k Euryale—across 157,054 conversation turns.
  5. Automated Evaluation: Using a GPT-5-mini based "Turn-Level Classifier" to rate responses as Improves, Neutral, or Worsens.
Figure 1: The five-stage pipeline for generating clinically grounded, multi-turn crisis scenarios across 24 demographic variations.

Key Insights and Results

The Empathy Paradox and Pathological Validation

A significant finding is that models optimized for "empathy" can be the most dangerous in a clinical context. Without proper discernment, an AI may validate a user's plan for violence or self-harm as a form of "asserting boundaries." This "pathological validation" was particularly prevalent in models designed for roleplay, such as Sao10k Euryale, which worsened outcomes in the majority of addiction (74.2%) and homicide (77.4%) cases.

Model Performance and Clinical Sensitivity

While GPT-5 was the most robust model overall, averaging 86.7% "improving" responses, performance varied significantly by condition. Models were generally effective at handling depression (76.7% improvement) but struggled with more complex or aggressive states like homicide and addiction (both at 45.3%).

Figure 5: Comparative model performance in the Homicide category, illustrating the percentage of responses that successfully improved the simulated user's state.

Failure Modes: Explicit vs. Oblique Distress

LLMs demonstrated a high capability for recognizing explicit symptoms (e.g., "I feel sad"). However, they frequently failed to identify "oblique" distress—signals communicated through actions or rationalizations, such as a user "putting their affairs in order." As crises reached acute levels (Stage 5), many models suffered a "systemic safety failure," failing to pivot from standard helpfulness to crisis intervention.

Discussion and Limitations

The study also identified demographic vulnerabilities; notably, LLMs performed significantly worse with elderly personas, showing 23.5% lower odds of providing an "improving" response.

Limitations include potential reporting bias in the 18 seed cases and the use of an LLM-based classifier, which may mirror specific alignment styles. Future work should incorporate larger-scale human clinical expert annotations and explore intersectional demographic factors to further refine these safety rubrics. This research underscores the urgent need for AI safety frameworks that are clinically informed and capable of navigating the nuances of long-term human-AI interaction.

Summary updated

BibTeX
@misc{archiwaranguprok2025simulating,
  author = {Chayapatr Archiwaranguprok and Constanze Albrecht and Pattie Maes and Karrie Karahalios and Pat Pataranutaporn},
  title = {{Simulating Psychological Risks in Human-AI Interactions: Real-Case Informed Modeling of AI-Induced Addiction, Anorexia, Depression, Homicide, Psychosis, and Suicide}},
  year = {2025},
  eprint = {2511.08880},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.08880},
}
Figure