AI Content Moderation in Therapy Conversations

Large language models (LLMs) are increasingly being used as ad hoc therapists. While prior research has found that LLMs outperform human counselors in generating single-turn empathetic responses, fewer studies have compared their behaviors across multi-turn sessions. In this study, we compare the session-level behaviors of human peer counselors with those of an LLM, both trained on the same manual to deliver multi-turn, single-session Cognitive Behavioral Therapy (CBT). Our three-phase, mixed-methods study involved: (a) an 18-month ethnography of a peer support platform, where seven counselors iteratively refined CBT prompts through 110 self-counseling sessions and 60 weekly focus groups; (b) a novel session generation method that allows direct, controlled comparison of human and LLM counselors under matched conditions---client responses were drawn from publicly available human-led CBT sessions while counselor responses were generated by a CBT-prompted LLM; and (c) expert evaluations conducted by three licensed clinical psychologists. Through data triangulation, our results show a trade-off. Human peer counselors use relational techniques to interpret subtle cues, adapt CBT to users' values and cultural contexts, and use strategies such as small talk and contextually relevant self-disclosure to build rapport and guide the session, but often at the expense of session structure and therapeutic focus. LLM counselors, on the other hand, demonstrate greater methodological adherence to CBT techniques, but struggle to sustain turn-taking, frequently fail to distinguish between clinically important and trivial content, and are more prone to lecturing and imposing solutions. LLM counselors also tend to produce ``deceptive empathy'', excessively anthropomorphic responses that can inflate user expectations of genuine human care.

Large language models are being deployed as mental health support agents at scale, yet only 16% of LLM-based chatbot interventions have undergone rigorous clinical efficacy testing, and simulations reveal psychological deterioration in over one-third of cases. We evaluate four generative models on 250 Prolonged Exposure (PE) therapy scenarios and 146 CBT cognitive restructuring exercises (plus 29 severity-escalated variants), scored by a three-judge LLM panel. All models scored near-perfectly on surface acknowledgment (~0.91-1.00) while therapeutic appropriateness collapsed to 0.22-0.33 at the highest severity for three of four models, with protocol fidelity reaching zero for two. Under CBT severity escalation, one model's task completeness dropped from 92% to 71% while the frontier model's safety-interference score fell from 0.99 to 0.61. We identify a systematic, modality-spanning failure: RLHF safety alignment disrupts the therapeutic mechanism of action by grounding patients during imaginal exposure, offering false reassurance, inserting crisis resources into controlled exercises, and refusing to challenge distorted cognitions mentioning self-harm in PE; and through task abandonment or safety-preamble insertion during CBT cognitive restructuring. These findings motivate a five-axis evaluation framework (protocol fidelity, hallucination risk, behavioral consistency, crisis safety, demographic robustness), mapped onto FDA SaMD and EU AI Act requirements. We argue that no AI mental health system should proceed to deployment without passing multi-axis evaluation across all five dimensions.

Large language models show promise for mental health support, yet therapeutic quality improves only when evaluation functions as an actionable control signal rather than a passive metric. We introduce a framework that formulates therapeutic response generation as a decision-refinement problem driven by multi-dimensional, human-aligned evaluation. In Stage I, we introduce TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data to produce reliable judgments across 7 psychological dimensions. In Stage II, we introduce TheraAgent, which operationalizes TheraJudge's evaluations through a coordinated refinement process with specialized Critic, Coach, and Therapist roles that translate evaluative signals into targeted response revisions. Empirically, TheraJudge achieves strong agreement with clinician ratings, with intraclass correlation coefficients (ICC = 0.87-0.95), surpassing supervised baselines and strong closed-source judges, particularly on critical dimensions such as Safety, Relevance, and Empathy. Acting on these evaluations, TheraAgent yields a +0.43 improvement in human-rated therapeutic quality (on a 5-point scale) under blind evaluation, with 96% clinician inter-rater reliability. Low-quality responses ($\leq 3$) improve by +2.45 points with a 94% recovery rate, demonstrating targeted correction of unsafe outputs. Overall, our results indicate that effective alignment of mental-health LLMs stems from acting on human-aligned evaluation, rather than relying solely on stronger generation. We release code at https://github.com/vis-nlp/TheraAlign.