Typing a worry into a chatbot at 2 a.m. gets an instant, patient, kind-sounding reply.
That is more than a months-long therapy waiting list offers, and the convenience is real. It is not, however, therapy.
Small, short trials show that chatbots built around therapy techniques can ease depression and anxiety symptoms for some people over a few weeks. No chatbot can assess you, take clinical responsibility for your care, or protect your words under health-privacy law.
In this article
Predicted words, not felt empathy
Large language models write replies by predicting likely sequences of words, based on huge amounts of training text. A comforting answer reflects patterns of supportive writing, not felt empathy or clinical judgment.
That does not make the experience empty. Putting a worry into words can help organize it, and a calm, nonjudgmental reply can leave someone feeling heard.
The National Institute of Mental Health sees promise in technology that expands access to mental health care, but a text model works with very little information. It cannot hear a shaking voice, notice weight loss, or know your history unless you type it out.
It also cannot judge whether a suggestion is safe for the particular person reading it. That gap matters most for people who turn to a chatbot late at night because nothing else feels available.
Small, short studies of therapy chatbots
The best-known early study, a randomized trial of the Woebot chatbot, enrolled 70 US college students who reported symptoms of depression or anxiety. Over two weeks, depression scores fell more in the chatbot group than in a group given an NIMH e-book on depression.
Anxiety scores dropped in both groups, with no clear difference between them. The trial was small and brief, and several authors were connected to the company that built the bot.
A larger 2025 trial in NEJM AI tested Therabot, a generative chatbot built by Dartmouth researchers, with about 210 US adults who had depression, generalized anxiety or high eating-disorder risk. After four weeks, symptoms improved more than in a waitlist group.
Those results are encouraging but limited. A waitlist comparison cannot show whether a bot matches a human therapist, and the research team monitored conversations so staff could step in if risk or a poor response appeared.
Purpose-built tools like these differ sharply from general assistants, which produce free-form answers of uneven quality and were not designed or tested as treatment. The UK’s NICE uses an evidence standards framework for digital health technologies to separate tested products from unproven ones.
Across this research, samples are small, follow-up is usually measured in weeks, and people with severe illness or suicidal thinking are typically excluded. Nobody yet knows the long-term effects of leaning on a chatbot for serious or complex difficulties.
Missing assessment and the therapist bond
Therapy is more than supportive conversation. A licensed clinician assesses, forms a working diagnosis, and chooses an approach suited to the problem, which is why deciding between CBT, DBT and ACT starts with that assessment.
A meta-analysis of nearly 300 studies found that the working alliance, the bond and shared goals between client and therapist, consistently predicts how well therapy goes. A chatbot can imitate rapport but carries no responsibility for the relationship.
Two technical habits add risk. Chatbots lean toward agreeableness, so they may validate a catastrophic fear, a suspicious belief or an unhealthy relationship pattern that a therapist would gently question.
They can also produce confident errors, often called hallucinations. Wrong information about a symptom, a diagnosis or a medication can do real harm when it arrives in a fluent, authoritative tone.
| Feature | General-purpose chatbot | Licensed therapist |
|---|---|---|
| Availability | Instant, at any hour | Scheduled sessions, often after a wait |
| Assessment | None; responds only to the text it receives | Structured assessment, including risk |
| Accountability | Company terms of service | Licensing board and professional ethics code |
| Response to risk | Scripted warnings or referrals, applied inconsistently | Safety planning and coordinated emergency care |
| Privacy | Commercial privacy policy | Health-privacy law, with narrow legal exceptions |
Chat privacy versus clinician confidentiality
In the US, what you tell a licensed clinician is protected by HIPAA, and clinicians elsewhere work under similar confidentiality duties. They cannot sell your history or use it for advertising, and disclosure without consent is limited to narrow situations such as serious risk.
Consumer chatbots generally fall outside HIPAA. Depending on the company’s policy, chats may be stored, read by human reviewers or used to train future models, and a data breach or a change in terms can expose them.
The World Health Organization’s guidance on the ethics and governance of AI for health calls for privacy protection and safeguards against harm before such tools are widely adopted.
Public trust is limited too. In a 2023 survey, 60% of US adults told Pew Research Center they would be uncomfortable if their own health care provider relied on AI.
Using a chatbot without making it your therapist
With clear limits, a chatbot can be a reasonable self-help aid. These uses play to its strengths:
- Untangling a situation. Typing out a messy conflict at work and asking for the separate issues to be listed brings some of the same benefits as journaling.
- Practicing CBT-style exercises. Asking for questions that test an all-or-nothing thought mirrors the CBT exercises people practice at home, with the same limits.
- Interrupting a loop. Moving racing thoughts onto the screen and sorting them into solvable and unsolvable can be one way to stop overthinking a decision.
- Preparing for an appointment. Drafting what to say to a doctor or therapist can make a first conversation easier to start.
A few habits keep those uses safer. Leave out names, addresses and medical record details, and check any claim about symptoms, diagnoses or medication against a reliable health source, a pharmacist or a prescriber.
Notice whether the chatbot is replacing people rather than adding to them. Talking hard feelings through with friends, family or a clinician builds connection that software cannot, and long solo sessions can quietly deepen isolation.
If a reply leaves you more unsettled, or the bot seems to agree with everything, close the app. That is a cue to talk with someone who knows you.
Problems that need a person, not a bot
Some patterns are worth raising with a doctor or mental health professional rather than a chatbot, especially if they last more than two weeks or affect work, sleep or relationships:
- Low mood, emptiness, or loss of interest in things that used to matter.
- Worry or panic that makes ordinary tasks hard to manage.
- Changes in sleep, appetite or energy that do not improve with routine adjustments.
- Intrusive memories, nightmares or constant alertness after a frightening event.
- Hopelessness or thoughts of self-harm, which call for a person today, not an app.
Questions about starting, stopping or combining medication also belong with a prescriber, since a chatbot cannot see your medical record or weigh your other conditions.
For anyone who wants a digital tool alongside care, the American Psychiatric Association’s app evaluation guidance suggests checking privacy, evidence and how the tool fits with treatment.
Notes from chatbot sessions can still be useful. A first therapy session usually begins with background on what has been happening, and a written record of recurring worries gives the clinician a concrete place to start.
