MindGuard: Guardrail Classifiers for Multi-Turn Mental Health Support

TL;DR

MindGuard introduces clinically grounded safety classifiers for multi-turn mental health support, improving safety and reducing false positives in AI conversations by collaborating with psychologists and leveraging real and synthetic data.

Contribution

The paper presents a novel risk taxonomy, a dataset of annotated conversations, and lightweight classifiers trained to enhance safety in mental health AI systems.

Findings

Classifiers reduce false positives at high recall

Lower attack success rates with clinician language models

Enhanced safety in multi-turn interactions

Abstract

Large language models are increasingly used for mental health support, yet their conversational coherence alone does not ensure clinical appropriateness. Existing general-purpose safeguards often fail to distinguish between therapeutic disclosures and genuine clinical crises, leading to safety failures. To address this gap, we introduce a clinically grounded risk taxonomy, developed in collaboration with PhD-level psychologists, that identifies actionable harm (e.g., self-harm and harm to others) while preserving space for safe, non-crisis therapeutic content. We release MindGuard-testset, a dataset of real-world multi-turn conversations annotated at the turn level by clinical experts. Using synthetic dialogues generated via a controlled two-agent setup, we train MindGuard, a family of lightweight safety classifiers (with 4B and 8B parameters). Our classifiers reduce false positives at high-recall operating points and, when paired with clinician language models, help achieve lower attack success and harmful engagement rates in adversarial multi-turn interactions compared to general-purpose safeguards. We release all models and human evaluation data.