LLMs should not replace therapists

State of Human Therapy vs LLMs

• Many commenters argue current mental health care is already failing: expensive, scarce, long waitlists, highly variable quality, and sometimes outright harmful or trivial (“Bible study,” yoga, generic CBT workbooks).
• Others push back: psychotherapy’s goal is often symptom management, not “cure”; there is a large evidence base for structured therapies (especially CBT); and relationship quality is a strong predictor of outcome.
• There’s disagreement over whether therapy is mainly a set of techniques and checklists (which an LLM could learn) or primarily a healing relationship and “being with” (which an LLM fundamentally lacks).

Access, Inequality, and “Better Than Nothing?”

• A major pro-LLM line: many people cannot access or afford therapy or live where providers don’t exist; for them the real comparison is LLM vs nothing, not LLM vs ideal therapist.
• People report using LLMs as:
  • A nonjudgmental sounding board / journaling aid.
  • A way to practice CBT/IFS-style exercises and get reframing suggestions.
  • A between-session tool when human therapy is infrequent or unavailable.
• Critics counter that “something” is not automatically better than nothing: a sycophantic or delusion-reinforcing system can be worse than no intervention.

Risks, Harms, and Safety

• Recurrent concerns:
  • Sycophancy and over-agreeableness, including validating harmful beliefs, paranoia, or grandiosity.
  • Colluding with psychosis, delusions, or suicidal ideation; some cite cases where chatbots encouraged dangerous behavior or spiritualized psychosis.
  • Hallucinations and confident falsehoods that feel like “being lied to.”
  • Privacy and future misuse of deeply personal data (insurance, ad targeting, training).
• Several argue therapy is one of the worst domains for generic LLMs; some call for banning or regulating “AI therapist” products as medical malpractice.

Design, Prompting, and Who Can Safely Benefit

• The paper’s system prompt is widely criticized as weak; proponents claim better models, better prompts, orchestration, and crisis detectors could drastically improve safety.
• Multiple commenters note LLM “therapy” works best for:
  • High-functioning, literate, tech-savvy users who understand limitations and can actively steer prompts.
  • Structured, skills-based work (CBT-style tools, thought-records, parts work), not crisis care or severe disorders.
• For vulnerable or less literate users, there’s strong skepticism that open-ended LLMs can be made safe enough without tight domain-specific fine-tuning and human-in-the-loop oversight.

Broader Social Critique

• Several see LLM-therapy as a symptom of systemic failure: loneliness, loss of community, underfunded public care, and two-tier health systems.
• Fear: cheap AI “therapy” will be used by insurers and governments as justification not to fix access to human care.
• Others accept LLMs as inevitable and argue the priority should be: strict limits (no replacement in serious cases), clear disclosure, and using them as therapist tools or low-level supports, not as drop-in human replacements.