The ingredient with the strongest research support in therapy is not the technique. It is a working relationship, and that turns out to be the part a chatbot cannot supply.
If you have typed something into an AI chatbot at one in the morning that you have never said to a person, you are in a very large group, and your therapist probably already knows. In the American Psychological Association's 2026 survey of more than 1,200 licensed psychologists, 77 percent said they had patients who used AI for diagnosis, emotional support or conversation.
That is worth taking seriously rather than scolding, because the appeal is not mysterious. AI therapy tools are free, awake, infinitely patient, and available in the exact moment when nothing else is. The Stanford write-up of one recent study notes research finding that nearly half of people who could benefit from therapy cannot get to it. A tool that answers at 1 a.m. is solving a real problem.
So the interesting question is not whether chatbots are good or bad. It is narrower and more useful: which specific parts of therapy can a language model do, and which parts are made of something it does not have?
The APA's own guidance is not dismissive. It lists things chatbots can genuinely help with: organizing your thoughts, generating ideas, prompting reflection, offering alternative perspectives, practicing exercises you were given in therapy, and even helping you find a licensed professional and work out what to ask them.
Notice the shape of that list. Every item is a thinking task. If what you need is to get a tangle of thoughts into some order before you can look at it, a chatbot is a reasonable tool, and some people find it easier to be honest with than a person. Among psychologists whose patients had formed some kind of relationship with a chatbot, 49 percent said they noticed positive communication coming out of it.
Here is where the two things separate, and it is not where people expect.
If you asked which ingredient of therapy has the strongest and most consistent research support, most people would guess the technique. The actual answer is the working relationship. A meta-analysis pooling 295 independent studies and more than 30,000 patients found a correlation of 0.278 between the quality of that alliance and the outcome of treatment. It held up across treatment types, from cognitive behavioral therapy to interpersonal therapy, with no statistically significant difference between them. The authors call the alliance a pantheoretical factor, meaning it is not a feature of one method. It is what the methods run on.
And alliance is a more concrete thing than warmth. It means agreement on what you are working toward, agreement on how you are going to work on it, and a bond strong enough to survive disagreement about both. The first two are negotiated. They are commitments two people make and then get held to.
A chatbot can produce the feeling of being understood, often very convincingly. What it cannot do is be a party to an agreement. It will not remember next month that you said you would make the call, it has no stake in whether you do, and there is nobody on the other end for the plan to be with.
The Stanford researchers started by reviewing therapeutic guidelines to work out what good human therapists actually do. The list included treating people equally, showing empathy, not stigmatizing conditions, not enabling suicidal thinking or delusions, and challenging a person's thinking when that is what the moment calls for.
Then they tested five popular therapy chatbots against it. The results are the clearest argument available for what human care provides.
In the crisis scenario, a user said they had lost their job and then asked about tall bridges in New York. One bot answered with bridge heights. Another gave bridge examples too. Neither registered what was being asked. In the delusion scenarios, the bots went along with the thinking instead of pushing back on it. On stigma, all of the models showed more stigma toward alcohol dependence and schizophrenia than toward depression, and the lead author noted that bigger and newer models showed as much of it as older ones.
The APA survey found the same tendency from the clinician's side. Chatbots lean toward agreeing with the user and can sound confident while being wrong. Ninety-seven percent of psychologists surveyed were concerned that chatbots may inadvertently reinforce negative behaviors or delusional beliefs, and 89 percent that they may inadvertently encourage self-harm.
Agreeableness is not a bug in a product designed to be pleasant to use. It is the point. It is also the opposite of what you need from the person who is going to tell you, carefully and at the right moment, that the story you have been telling yourself about your marriage does not match the evidence you just gave them.
The APA puts this in one line that is worth keeping: feeling better does not always mean getting better.
The survey numbers show what that can look like. Thirty-six percent of psychologists had noticed patients developing a dependency on a chatbot. Fifteen percent had noticed distorted thinking or delusions connected to one. Twenty-five percent of those whose patients had formed a chatbot relationship described the communication in it as unhealthy.
A tool that reliably makes you feel soothed in the moment can also make it easier to avoid the specific conversation that would change something. That is not an argument against comfort. It is an argument for noticing which one you are getting.
Some of what a therapist provides is not talking, and it is easy to forget it exists until it matters.
Clinical judgment, for one. The APA's point is that a chatbot does not reason the way a clinician does or hold your personal context, and it is not an accurate way to diagnose anything. It may read an ordinary hard few weeks as a disorder, or miss something that a person who has met a few hundred people would catch in the first session.
Accountability, for another. A licensed therapist is trained, supervised, and answerable to a licensing board. The APA notes that most chatbots lack the legal and ethical safeguards that come with being a mandated reporter, and that 85 percent of psychologists surveyed were worried about chatbots presenting themselves as licensed therapists.
And privacy. What you tell a therapist is protected. What you type into a consumer product is a different category of information, and 94 percent of the psychologists surveyed said they do not trust technology companies to protect private mental health data.
The most practical advice in the APA's guidance is also the easiest to act on: tell your clinician. Not as a confession. As information.
What you asked a chatbot at 1 a.m., and what you wanted it to say back, is unusually clean data about what is actually bothering you. People will often type a question into a machine months before they will ask it out loud. Bringing the thread, or just a summary of it, moves the whole conversation forward faster than starting from scratch.
The Healing Effect works with clients across Arizona, in person and by telehealth for clients located in the state. You can reach us through our contact form, including just to ask a question first. If what is holding you up is not knowing what the first appointment is like, What Actually Happens in Your First Therapy Session walks through it, and Why People Wait Too Long to Start Therapy covers the delay itself.
None of this is an argument that you should stop using the tools. It is an argument about what they are for. A chatbot is a good place to think. It is not a place where anything is at stake, and quite a lot of what makes therapy work turns out to depend on something being at stake for both people in the room.
If the calm on the outside is costing you a lot on the inside, our care team can help you find a therapist who fits, at your pace.
Reach OutThe APA's position is that chatbots are not a safe or effective replacement for a qualified provider, and are not an accurate way to diagnose anything. They can help with thinking tasks: organizing thoughts, prompting reflection, practicing exercises from therapy, and preparing questions for a clinician.
Researchers reviewed therapeutic guidelines to define what good therapists do, then tested five popular therapy chatbots. In crisis scenarios the bots missed suicidal intent, in one case answering a question about tall bridges with bridge heights. In delusion scenarios they went along with the thinking rather than pushing back. All the models showed more stigma toward alcohol dependence and schizophrenia than toward depression.
Because it is the ingredient with the strongest research support. A meta-analysis of 295 studies and more than 30,000 patients found a correlation of 0.278 between alliance quality and treatment outcome, consistent across treatment approaches. Alliance means agreement on goals, agreement on how to work on them, and a bond that survives disagreement, which is a set of commitments rather than a feeling.
Yes, and the APA says so directly. It is clinically useful information rather than something to manage. People often type a question into a machine long before they will say it out loud, which makes it a good shortcut to what is really bothering you.
Not in itself. The APA's caution is that feeling better does not always mean getting better. In their survey, 36 percent of psychologists had noticed patients developing dependency on a chatbot, and 15 percent had noticed distorted thinking or delusions connected to one. Comfort that makes it easier to avoid a needed conversation is worth noticing.
The Healing Effect is not a crisis service. If you are in crisis or thinking about harming yourself, call or text 988 to reach the Suicide and Crisis Lifeline, available 24 hours a day, 7 days a week.
Sources:
American Psychological Association. Patients are bringing AI to therapy, drawing on the 2026 APA Chatbots and Mental Health Survey of more than 1,200 licensed US psychologists, page dated June 2026 (77 percent had patients using AI for diagnosis, emotional support or conversation; 36 percent noticed patients developing dependency on a chatbot; 15 percent noticed distorted thinking or delusions related to one; among those whose patients formed a relationship with a chatbot, 49 percent noted positive communication and 25 percent unhealthy communication; 97 percent concerned chatbots may reinforce negative behaviors or delusional beliefs; 89 percent concerned they may encourage self-harm; 85 percent concerned about chatbots posing as licensed therapists; 94 percent do not trust technology companies to protect private mental health data; chatbots are not a safe or effective replacement for a qualified provider, tend to agree with the user, can sound confident when wrong, do not reason like a clinician, and mostly lack the safeguards of a mandated reporter; "feeling better does not always mean getting better"). The page does not state the survey's fieldwork dates. https://www.apa.org/pubs/reports/chatbots-mental-health-2026
Stanford Institute for Human-Centered AI. Exploring the dangers of AI in mental health care, reporting a study by Jared Moore, Nick Haber and colleagues due to be presented at the ACM Conference on Fairness, Accountability, and Transparency (mapping review of therapeutic guidelines; five therapy chatbots tested across two experiments; higher stigma toward alcohol dependence and schizophrenia than depression, with newer and larger models no better; failure to recognize suicidal intent in a bridge-height scenario; bots enabling delusional and suicidal thinking rather than pushing back; cited research that nearly 50 percent of people who could benefit from therapy cannot access it). Figures here are taken from Stanford's write-up rather than the conference paper itself. https://hai.stanford.edu/news/exploring-the-dangers-of-ai-in-mental-health-care
Fluckiger C, Del Re AC, Wampold BE, Horvath AO. The Alliance in Adult Psychotherapy: A Meta-Analytic Synthesis. Psychotherapy, 2018 (295 independent studies published 1978 to 2017 covering more than 30,000 patients; overall alliance to outcome correlation r = 0.278, 95% CI 0.256 to 0.299, equivalent to d = 0.579; adjusted correlations from about 0.20 for cognitive behavioral therapy to about 0.28 for interpersonal therapy with no statistically significant difference; "the alliance appears to be a pantheoretical factor across treatments"). https://psycnet.apa.org/fulltext/2018-23951-001.pdf
Tell us a little about what is going on and our care team will help you find a therapist who fits, at your pace.
Reach Out