When Chatbots Try to Be Therapists: The Risks, Realities & Responsibilities

A new paper by Stevie Chancellor et al. (University of Minnesota / UT Austin) examines whether large language models (LLMs) should act as therapists. Their verdict is clear—and cautionary. (You can see the abstract in your image.)

What the Paper Found

• Therapeutic gaps: LLMs, even the advanced ones, struggle to consistently replicate vital elements of therapeutic alliance—a key factor in mental health outcomes.
• Stigma & misalignment: In experiments, LLMs sometimes expressed stigma toward those with mental health conditions or responded inappropriately (e.g. encouraging delusional thinking).
• Limits of safety frameworks: Even with newer models, current best practices don’t close all gaps in alignment, identity, and context.
• Conclusion: LLMs should not replace therapists. Instead, their role might be limited to supportive/adjunct tasks—summarization, triage, or assisting clinicians—not full therapeutic agents.

---

Why This Matters More Than It Seems

• Raise guardrails early: When your organization explores AI, especially for people-facing uses (HR, coaching, health), you must embed strong ethics, oversight, and human-in-the-loop from Day 1.
• Use pilots responsibly: Limit scope, collect feedback from real users, involve domain experts (mental health professionals), and monitor unintended harms.
• Coach with humility: As leaders or AI coaches, accept that your tool will misstep. Use mistakes as learning—correct, not suppress.
• Design for hybrid models: The future is not “AI instead-of-human”—it’s “AI + human.” Let machines help with scale, consistency, triage—but humans hold responsibility for meaning, nuance, and adaptation.