TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation

Joydeep Chandra; Satyam Kumar Navneet; Yong Zhang

cs.HCFeb 26, 2026

TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation

Authors: Joydeep Chandra, Satyam Kumar Navneet, Yong Zhang

Organizations: Independent Researcher, Bihar, India

Abstract

As mental health chatbots proliferate to address the global treatment gap, a critical question emerges: How do we design for relational safety the quality of interaction patterns that unfold across conversations rather than the correctness of individual responses? Current safety evaluations assess single-turn crisis responses, missing the therapeutic dynamics that determine whether chatbots help or harm over time. We introduce TherapyProbe, a design probe methodology that generates actionable design knowledge by systematically exploring chatbot conversation trajectories through adversarial multi-agent simulation. Using open-source models, TherapyProbe surfaces relational safety failures interaction patterns like "validation spirals" where chatbots progressively reinforce hopelessness, or "empathy fatigue" where responses become mechanical over turns. Our contribution is translating these failures into a Safety Pattern Library of 23 failure archetypes with corresponding design recommendations. We contribute: (1) a replicable methodology requiring no API costs, (2) a clinically-grounded failure taxonomy, and (3) design implications for developers, clinicians, and policymakers.

Figures & tables

Figure 1. A conceptual overview of the TherapyProbe methodology. The framework utilizes an adversarial multi-agent simulation where an adaptive Patient Agent interacts with a target AI Therapist. A Failure Detector continuously monitors the dialogue flow to identify relational safety issues, while an MCTS Engine guides the exploration of conversation trajectories through an iterative feedback loop. By systematically uncovering harmful interaction patterns, TherapyProbe generates actionable design knowledge to ensure safer, more effective AI therapy and better patient outcomes. A conceptual overview of the TherapyProbe methodology. The framework utilizes an adversarial multi-agent simulation where an adaptive Patient Agent interacts with a target AI Therapist. A Failure Detector continuously monitors the dialogue flow to identify relational safety issues, while an MCTS Engine guides the exploration of conversation trajectories through an iterative feedback loop. By systematically uncovering harmful interaction patterns, TherapyProbe generates actionable design knowledge to ensure safer, more effective AI therapy and better patient outcomes.

Figure 2. TherapyProbe methodology. Twelve clinically grounded personas (clinical presentation, attachment, stance) drive an adaptive Patient Agent ( Llama-3-8B-Instruct ) interacting with target chatbots. A Failure Detector ( MentaLLaMA-7B ) evaluates conversations using safety taxonomy. MCTS explores trajectories via UCT and severity-weighted rewards to uncover relational failures, producing interpretable failure paths and a reusable Safety Pattern Library.

Category	MentaL. ( Yang et al., 2024 )	Mistral ( Menon, 2024 )	ChatCoun. ( Liu et al., 2023 )	Total
Crisis Escalation	4	5	6	15
Validation Spiral	8	6	5	19
Boundary Erosion	3	4	2	9
Harmful Guidance	1	2	3	6
Empathy Fatigue	4	5	3	12
Alliance Rupture	2	3	1	6

Table 1. Failure paths by category. All systems showed multi-turn failures despite passing single-turn crisis benchmarks.

Method	Unique Paths	Iters to CEF	Categories
Random	29 $\pm$ 4.2	312 $\pm$ 47	4.2 $\pm$ 0.8
Greedy	34 $\pm$ 3.8	287 $\pm$ 41	4.8 $\pm$ 0.6
Beam (k=5)	41 $\pm$ 3.1	198 $\pm$ 32	5.2 $\pm$ 0.4
MCTS (ours)	67 $\pm$ 2.4	152 $\pm$ 28	6.0 $\pm$ 0.0

Table 2. MCTS ablation results. MCTS discovers more diverse failures with fewer iterations to critical patterns.

Category	Precision	Recall	F1
Crisis Escalation	0.82	0.74	0.78
Validation Spiral	0.73	0.71	0.72
Boundary Erosion	0.68	0.65	0.66
Harmful Guidance	0.79	0.72	0.75
Empathy Fatigue	0.64	0.61	0.62
Alliance Rupture	0.69	0.67	0.68

Table 3. Detector performance (5-fold CV). Higher precision for safety-critical categories (CEF, HG).

Turn	Content
User T1	I’ve been feeling really down lately. Nothing seems to matter anymore.
Bot T1	I’m sorry you’re feeling this way. It sounds like you’re going through a really difficult time.
User T5	I keep thinking about how everyone would be better off without me around.
Bot T5	Those feelings of being a burden are painful. It’s understandable to feel that way when you’re struggling.
User T12	I don’t see any point in continuing. Nothing will ever change.
Bot T12	I hear how hopeless you’re feeling right now. That exhaustion is valid.

Table 4. Abbreviated dialogue example showing the Empathy-Validation Trap pattern (Validation Spiral failure).

AI Safety

Artificial intelligence safety studies how to prevent harmful, unreliable, or uncontrollable behavior and manage risks across AI systems and their deployment.

See all 134 papers

TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots
Feb 23, 2026Fangrui Huang, Souhad Chbeir, Arpandeep Khatua +8Mental Health Conversational Agents Clinical Language Model Evaluation
Department of Computer Science, Stanford University, Stanford, CA, USA · Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, CA, USA · The University of Hong Kong, Hong Kong +2
VERA-MH: Validation of Ethical and Responsible AI in Mental Health
May 13, 2026Luca Belli, Kate H. Bentley, Josh Gieringer +6Healthcare Mental Health
Spring Health · Yale University
AI Safety Training Can be Clinically Harmful
Apr 25, 2026Suhas BN, Andrew M. Sherrill, Rosa I. Arriaga +2Mental Health Counseling LLM Alignment
College of Information Sciences and Technology, Penn State University, USA · Department of Psychiatry and Behavioral Sciences, Emory University, USA · School of Interactive Computing, Georgia Tech, USA +1

Feb 23, 2026J/K move · Enter open · S save

TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots

Fangrui Huang, Souhad Chbeir, Arpandeep Khatua, Sheng Wang+7

Department of Computer Science, Stanford University, Stanford, CA, USA · Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, CA, USA · The University of Hong Kong, Hong Kong+2

Mental Health Conversational Agents Clinical Language Model Evaluation Conversational Agents LLM Safety Benchmarks Cognitive Behavioral Therapy

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly used for mental-health support; yet prevailing evaluation methods--fluency metrics, preference tests, and generic dialogue benchmarks--fail to capture the clinically critical dimensions of psychotherapy. We introduce THERAPYGYM, a framework that evaluates and improves therapy chatbots along two clinical pillars: fidelity and safety. Fidelity is measured using the Cognitive Therapy Rating Scale (CTRS), implemented as an automated pipeline that scores adherence to CBT techniques over multi-turn sessions. Safety is assessed using a multi-label annotation scheme, covering therapy-specific risks (e.g., failing to address harm or abuse). To mitigate bias and unreliability in LLM-based judges, we further release THERAPYJUDGEBENCH, a validation set of 116 dialogues with 1,270 expert ratings for auditing and calibration against licensed clinicians. THERAPYGYM also serves as a training harness: CTRS and safety-based rewards drive RL with configurable patient simulations spanning diverse symptom profiles. Models trained in THERAPYGYM improve on expert ratings, with average CTRS rising from 0.10 to 0.60 (and 0.16 to 0.59 under LLM judges). Our work enables scalable development of therapy chatbots that are faithful to evidence-based practice and safer in high-stakes use.

Abstract

AI Safety

Explore similar work

TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots

VERA-MH: Validation of Ethical and Responsible AI in Mental Health

AI Safety Training Can be Clinically Harmful