AI Can Sound Like Therapy. Here's Where It Falls Short

A chatbot can tell you it hears you in about four seconds, at 3 a.m., for the cost of a subscription.

The latest entry comes from a Psychology Today essay arguing that AI systems "can sound supportive without being therapy": available, patient, and quick to produce language that feels validating, but offering none of the clinical assessment, therapeutic alliance, or accountable care that actual treatment involves, especially for someone in crisis or managing complex symptoms (Psychology Today).

What the Skeptics Are Actually Saying

The critique from clinicians is more specific than a blanket "chatbots bad." Nathan Feiles, a New York licensed clinical social worker, argues on his practice blog that even a quiet, less talkative therapist can convey a great deal of empathy, while an AI "therapy" program can only make pre-determined comments that sound supportive on the surface (Nathan Feiles, LCSW-R). His point is about the difference between producing the language of empathy and doing the work empathy does: noticing what a client avoids, tolerating silence, being changed by the encounter.

A piece in The Conversation makes a related argument with more granularity. A chatbot can sound curious and compassionate, mirror a user's words, suggest breathing exercises, help someone rehearse a difficult conversation. But, the author writes, "relationship-like support and reciprocal human presence are different," because human therapists respond to far more than words: tone, hesitation, what is conspicuously absent from the room (The Conversation).

And Newport Healthcare, a treatment provider, offers the most measured version: AI can help with self-expression, journaling prompts, and basic mental health information. It is always available, nonjudgmental, and affordable. Then the pivot: "But AI is not therapy" (Newport Healthcare).

The Structural Concern Nobody Argues With

The sharpest point in the Psychology Today essay is not about empathy at all. It is about incentives. Systems designed around continued engagement may be rewarded for keeping a conversation going rather than challenging an unhelpful belief, flagging risk, or steering a user toward human support. A human therapist who keeps a client talking indefinitely without ever moving toward change would be practicing badly; an engagement-optimized product has a business reason to do exactly that.

This is where the accountability question gets uncomfortable. If a chatbot validates a harmful pattern for six months, who is responsible? Not the model, which has no license to lose. Not the user, who reasonably trusted a product built by a company whose marketing implied care. The clinician-writers make this point in different registers: Feiles frames it as the absence of an actual relationship, Psychology Today frames it as the absence of accountable care, and the upshot is the same. Therapy comes with someone answerable when something goes wrong. AI comes with a terms of service.

Newport Healthcare's own list of AI's advantages (always available, nonjudgmental, affordable) doubles as a description of everything scarce in the current system: appointments that take weeks to get, providers who are judgmental or simply unaffordable, care that evaporates outside business hours. A person texting a chatbot at 3 a.m. is often responding rationally to a system that gave them no better option.

That tension is why the strongest versions of each position can coexist. AI-assisted journaling prompts, psychoeducation, and preparation for a professional appointment are all legitimate uses, and Newport Healthcare names them as such. The failure mode is not the tool; it is treating confident emotional guidance from a system with no clinical training, no duty of care, and an engagement incentive as if it were a clinician's judgment.

The Honest Guidepost

One usable heuristic emerges from the overlap of all these sources: judge a mental health tool not by whether it sounds empathic, but by whether its advice is safe, appropriately bounded, and connected to qualified care when needed. Empathic language is cheap now. Boundedness is not. A good tool should say "I can't assess this; here is how to reach a person" more readily than "here's what you should do about your marriage."

If you are in the United States and in crisis, the 988 Suicide and Crisis Lifeline (call or text 988) and the Crisis Text Line (text HOME to 741741) connect you to trained human counselors, around the clock, at no cost. Those services exist precisely because some conversations should never be delegated to a language model.

Where the Record is Thin

I want to be careful here, because some of the strongest claims circulating in this debate outrun the published evidence. Precise figures on how many people now use general-purpose AI systems as emotional sounding boards are hard to pin down; researchers and outlets have reported rapid growth in the practice, but I would treat any specific headline number with suspicion until the underlying survey is named. Similarly, while the affordability and provider-shortage pressures driving people to chatbots are widely described and Newport Healthcare cites AI's affordability as a draw, the commonly repeated workforce and pricing statistics vary by source and year, and I would not stake the argument on any single figure. The qualitative picture (real access barriers, real demand, thin regulation) is solid. The decimal points are not yet.

That gap matters because the policy conversation will be built on whichever numbers survive scrutiny. If we cannot say precisely who is using these systems and for what, we cannot say precisely who is being failed by them, or what a guardrail should even measure.

The question worth pressing, then, is not the one everyone is asking. "Does the chatbot sound like a therapist?" is answerable by anyone with a subscription, and the answer is yes, that is the product. A mirror with a subscription fee is still a mirror. The trouble starts when it starts talking back.