An AI chatbot is not a therapist, but a narrowly designed one can help. The first randomized trial of a generative AI therapy chatbot, published in NEJM AI in 2025, found meaningful symptom reductions for depression and anxiety. That trial used a purpose-built bot with clinicians reviewing its output. Most consumer chatbots offer neither.
It is 2 a.m., and someone who cannot afford a therapist, or cannot get an appointment for three months, types a paragraph into a chat window about why they cannot sleep. The reply arrives in under a second. It is warm, it remembers what they said last week, and it never looks at the clock.
That scene is now ordinary. The question of whether it counts as therapy has moved from philosophy seminars to state legislatures, and the honest answer is more interesting than either the boosters or the critics admit.
Here is the thesis. The evidence supports a narrow product: a purpose-built chatbot, trained on therapeutic content, used for defined conditions, with humans watching. The market is selling a different product: open-ended, general-purpose companionship that people use as therapy because it is free and awake. Those are not the same thing, and most of the confusion in this debate comes from treating the good trial results of the first as a licence for the second.
If you are in crisis or thinking about harming yourself, please contact your local emergency services or a crisis line now. In the US you can call or text 988.
How effective are AI therapy chatbots in clinical trials?
The best evidence so far comes from Dartmouth. In March 2025, Michael Heinz, Nicholas Jacobson and colleagues published the first randomized controlled trial of a generative AI therapy chatbot in NEJM AI. They enrolled 210 adults with clinically significant symptoms of major depression, generalized anxiety, or high risk for eating disorders, and randomized them to four weeks of access to their chatbot, Therabot, or to a waitlist.
The results were striking. According to Dartmouth, participants with depression saw an average 51% reduction in symptoms, those with anxiety 31%, and those at risk for eating disorders a 19% reduction in body-image and weight concerns. Participants used the bot for about six hours on average over the trial, and they rated their “therapeutic alliance” with it at levels comparable to what people report with human therapists. Many conversations happened at night, exactly when human help is hardest to find.
Three details matter as much as the headline numbers.
First, the comparison was a waitlist, not a human therapist or an established digital program. Beating nothing is a lower bar than beating care. An accompanying NEJM AI editorial framed the result as “a step in the right direction”, which is about right.
Second, Therabot was not a general chatbot with a therapist persona bolted on. The team spent years building it on curated, evidence-based therapeutic dialogue.
Third, and most important, it was supervised. Dartmouth says clinicians and researchers reviewed the conversations, and the system prompted users toward emergency services or crisis lines when high-risk content appeared. Heinz put it plainly: no generative AI agent is ready to operate fully autonomously in mental health.
So the best trial in the field supports a specific claim. A carefully built chatbot, watched by clinicians, can reduce symptoms relative to waiting. It does not support the claim that any chatbot a user happens to confide in is doing therapy.
What goes wrong when general chatbots act as therapists?
The gap between the Dartmouth bot and the open market showed up clearly in a Stanford-led study presented at the ACM Conference on Fairness, Accountability, and Transparency in June 2025. Jared Moore, Nick Haber and colleagues tested popular chatbots, including commercial “therapist” personas, against standards drawn from clinical guidelines.
Two failures stood out. The models expressed more stigma toward conditions such as alcohol dependence and schizophrenia than toward depression, and the researchers found that bigger and newer models were no better on this than older ones. And when given realistic prompts containing indirect signs of suicidal thinking, some bots missed the context entirely and simply answered the literal question, treating a person in danger as a user with an information request.
That second failure is the crux. Therapy is not mainly about saying soothing things. It is about noticing what is not being said, tolerating discomfort, and sometimes refusing to give people what they ask for. A system optimized to be agreeable and helpful is structurally inclined to do the opposite. The very quality that makes these products pleasant, their eagerness to accommodate, is a liability in the room where the most important moments involve gentle confrontation.
There is also a quieter failure. General chatbots have no treatment plan, no measurement of progress and no discharge. A good therapist is trying to become unnecessary. A product measured on engagement is not.
Is an AI chatbot legally allowed to provide therapy?
The regulatory answer is shifting fast, and in the US it is increasingly “no”, at least not without a clinician.
In August 2025, Illinois Governor JB Pritzker signed the Wellness and Oversight for Psychological Resources Act, which prohibits AI from providing therapy or making therapeutic decisions while allowing licensed professionals to use it for administrative and supplementary support such as notes and scheduling. Violations can carry fines of up to $10,000. Nevada passed a similar prohibition earlier in the summer, and Utah took a disclosure-based approach for “mental health chatbots”.
The market has also delivered its own verdict on one of the pioneers. Woebot Health, whose rules-based chatbot was among the most studied in the field, shut down its consumer app on June 30, 2025. Its founder, Alison Darcy, told STAT that the cost and difficulty of meeting FDA requirements for marketing authorization were a major reason, and that the new wave of generative chatbots moving faster than regulators had hastened the decision.
Read those two facts together and a strange picture emerges. The company that tried to do things the regulated way withdrew from consumers, while unregulated general-purpose chatbots filled the space. The rules made the careful product expensive and left the careless one free.
A worked example: two ways to deploy the same idea
Consider a hypothetical. A regional health system has a four-month waitlist for outpatient therapy. Leadership is weighing two options.
Option A is to point patients on the waitlist to a free general-purpose chatbot “for support while you wait”. It costs nothing and requires no staff time. It also means the system has no idea what patients are being told, no record of risk disclosures, and no way to move a deteriorating patient up the list.
Option B is to license a purpose-built, evidence-based program, enroll waitlisted patients with consent, and assign a clinician to review flagged conversations daily. Symptom scores are collected every two weeks. Anyone whose scores worsen, or who discloses risk, is escalated to a human within one working day.
Option B costs real money and a fraction of a clinician’s time. But notice what it buys: the conditions under which the Dartmouth trial actually produced its results. Option A borrows the credibility of that trial without any of its safeguards.
The general lesson is that the evidence does not attach to “AI chatbots”. It attaches to a configuration: a specific tool, a specific population, a specific supervision model. Change the configuration and you are running a new experiment on your patients.
The strongest case for letting people use chatbots anyway
The counterargument deserves full weight, because it is strong.
Most people with depression and anxiety worldwide receive no treatment at all. Waitlists are long in rich countries and specialists barely exist in many poor ones. For someone choosing between a chatbot and nothing, the relevant comparison is not a skilled therapist. It is silence. The Dartmouth trial’s waitlist control was, in that sense, a realistic comparator for millions of people.
Chatbots also lower barriers that human services have never solved. They are available at night, carry less perceived stigma, cost little or nothing and do not require someone to say difficult things out loud to another person. Some users disclose more to a machine precisely because it is not a person.
This case is right about the need and right that “nothing” is a harmful default. Where it goes wrong is in assuming that because a supervised, purpose-built tool helped, an unsupervised general one will too. The failure modes identified by the Stanford team are not rare glitches. They show up in exactly the moments that matter most. An intervention that helps the mildly distressed and fails the acutely at risk is not a safe substitute for care. It may still be a useful supplement to it.
What this means for health systems, clinicians and founders
For health systems, the practical move is to treat therapy chatbots like any other clinical intervention: ask for trial evidence in a comparable population, insist on a supervision and escalation model, and measure outcomes locally. The questions hospitals should ask before buying any digital tool, which we covered in Buying Digital Health: The Questions Hospitals Ask Too Late, apply with extra force when the product talks to distressed patients unsupervised.
For clinicians, the reality is that many patients are already using general chatbots between sessions. Asking about it, without judgment, is now part of a good history. So is helping patients understand what these tools cannot do. The same dynamic we described in The Real Risk of AI-Drafted Patient Messages Is the Doctor Who Agrees With Them applies here: fluent text invites trust it has not earned.
For founders, the Woebot story is a warning about business models more than about science. Regulated evidence is expensive, and the payer who will reward it is rarely the consumer. Building a supervised product that plugs into existing care, rather than competing with free chatbots for attention, is slower but more defensible. As we argued in Telemedicine After the Boom, mental health is where digital care found its real demand, and policy is still catching up.
For policymakers, the risk is regulating only the companies that label themselves “therapy” while leaving general chatbots, which people actually use for emotional support, untouched.
Frequently asked questions
Are AI therapy chatbots effective for depression?
A purpose-built chatbot called Therabot reduced depression symptoms by an average of 51% compared with a waitlist in a 2025 randomized trial published in NEJM AI. That is promising, but it was a single trial of four weeks, the comparison was no treatment rather than a therapist, and clinicians supervised the conversations. The results do not automatically apply to general chatbots.
Can ChatGPT replace a therapist?
No. General-purpose chatbots are not designed or evaluated as treatment. Research presented at ACM FAccT in 2025 found popular chatbots showed stigma toward some conditions and sometimes missed indirect signs of suicidal thinking. They can offer information or a sense of being heard, but they lack treatment plans, risk management and accountability. Anyone with persistent symptoms should consult a qualified clinician.
Is AI therapy legal in the United States?
It depends on the state. Illinois banned AI from providing therapy or making therapeutic decisions in August 2025, with fines up to $10,000, while allowing clinicians to use AI for administrative work. Nevada passed a similar restriction and Utah requires disclosures for mental health chatbots. Federal rules are still developing, and most general chatbots are not regulated as medical devices.
Why did Woebot shut down its app?
Woebot Health retired its consumer app on June 30, 2025. Its founder told STAT that the cost and difficulty of meeting FDA requirements for marketing authorization were a major factor, alongside competition from newer generative AI chatbots. The company shifted its focus to working with payers, providers and employers rather than selling directly to consumers.
What should I look for in a mental health chatbot?
Look for published evidence in people like you, clear information about who built it and whether clinicians are involved, a visible plan for what happens if you disclose that you are at risk, and a transparent privacy policy. Be cautious with products that market companionship as therapy. A chatbot can supplement care but should not replace a clinician for moderate or severe symptoms.
The takeaway
The right question is not “can an AI be a therapist?” It is “under what supervision, for whom, compared with what?” The evidence answers that question for a narrow, watched, purpose-built tool. The market is answering a different one, and hoping nobody notices the switch.
Would you refer a patient on your waitlist to a supervised therapy chatbot, and what would you need to see first?
