A tense call with a parent ends late, sleep will not come, and the chatbot on the phone is always awake. Typing the evening out can show what stung and suggest a calmer way to reopen the talk tomorrow. That is a sensible use of AI for emotional support. Weeks of low mood, or thoughts of not wanting to be alive, need a person instead, because nothing on the other side checks whether you are safe.

The evidence also separates one kind of chatbot from another. A 2026 meta-analysis in npj Digital Medicine found that chatbot programs built for mental health brought small-to-moderate drops in symptoms of depression, anxiety and stress across 48 randomized trials, yet it counted only two trials of generative AI, the technology behind the chatbots on most phones.¹ The chatbots with the best evidence, in other words, are not the ones most people confide in.

So a chatbot works best as a supplement: somewhere to put a hard evening into words, rehearse a difficult conversation or practice a coping skill, alongside people and professional care. The American Psychological Association (APA) takes the same line in its 2025 health advisory: general chatbots and wellness apps should not replace a qualified mental health provider, though they may support ongoing therapy.² Most AI for personal productivity questions are about work; this one is personal.

What AI for emotional support can offer, and what it cannot

A chatbot offers an always-available, low-stakes place to name a feeling and try out a next step, but it cannot reliably assess risk, read your face or know your history, the APA’s 2025 advisory explains. The advisory also lists the barriers chatbots get around: stigma, cost, a shortage of nearby care and the wish to handle problems alone.²

The gap comes from what each side can see. A therapist works from tone of voice, body language, your history and training in spotting risk, while most chatbots process mainly the text you type. The relationship is also one-sided, even when it does not feel that way, and licensed providers answer to ethics codes and, in the US, state licensing boards in a way no app does.²

The need is worldwide. Even in high-income countries, most people with depression receive no mental health treatment, according to the World Health Organization (WHO).³

~1 in 3people with depression in high-income countries who receive mental health treatmentSource: World Health Organization, depression fact sheet, 2026

Someone dreading a talk with their manager about workload can ask a chatbot to play the manager, try three openings and notice which one they can say without apologizing. That is useful practice. What the chatbot will not do is notice that the same person has barely slept for a month and ask about it, as a doctor or therapist would.

Use a chatbot to sort thoughts and rehearse, and take anything about safety, diagnosis or medication to a person.

Purpose-built chatbots have trial evidence; everyday chatbots mostly do not

The trial evidence for chatbot support comes almost entirely from programs built for mental health, often scripted ones that follow set conversation paths, and the 2026 meta-analysismeta-analysis: A study that combines the results of earlier studies on the same question into one overall estimate. Pooling makes the estimate more precise, but it cannot repair the studies it pools: a meta-analysis of surveys is still survey evidence.Full entry in the glossary in npj Digital Medicine rated its certainty moderate for depression and stress and low for anxiety. Most of its trials ran in high-income countries and checked results within weeks rather than months, so whether the benefits last is still open, the authors say.¹

“Small-to-moderate” is the authors’ own description of the benefit, and it is worth translating: across many people the average symptom score fell noticeably, but an average says little about how much any one person will gain. An earlier meta-analysis in the same journal, from 2023, pooled 15 trials and found reductions in depression and distress but no clear change in overall well-being; few of the studies it reviewed used generative AI.⁴

In everyday terms, a structured app that walks you through noticing and reframing an anxious thought is much closer to what was tested than an open-ended chat with a general assistant. Before leaning on any tool for your mood, check what it was built to do and whether it has been tested for that job.

Myth

Any friendly chatbot is basically a free therapist.

Fact

The trials tested chatbots built for mental health. The APA's 2025 advisory says general chatbots should not replace a qualified provider.

The Therabot trial: good results, with clinicians reading along

The first randomized trialrandomized controlled trial: A study that assigns participants to the treatment or the comparison group at random, so the groups start out alike and a difference in what happens next can be put down to the treatment. Random assignment makes the two groups comparable; it does not make the people in the trial representative of anyone else.Full entry in the glossary of a fully generative therapy chatbot, Therabot, found that US adults with symptoms of depression, anxiety or eating-disorder risk who used it for four weeks reported larger symptom drops than people on a waitlist, according to a 2025 paper in NEJM AI. Clinicians and researchers reviewed every reply after it was sent.⁵

The study

Limited evidence

Therabot versus a waitlist: 210 US adults over four weeks

Depression scores fell by about 6 points on a 27-point questionnaire among Therabot users after four weeks, against about 2.6 points on the waitlist, and anxiety and eating-disorder-risk scores also fell more. Staff stepped in 15 times over safety concerns such as suicidal thoughts and 13 times to correct inappropriate replies such as medical advice.⁵

The setup limits what the result can tell you. The comparison was a waitlist, so it shows the app did better than waiting, not that it matched a therapist. People with active suicidal thoughts, mania or psychosis were excluded, so it says nothing about crisis use. Neither participants nor researchers were blinded, follow-up ended four weeks after use stopped, and the Dartmouth team that built Therabot ran and funded the trial.⁵ Two of its authors, including the lead author, have since co-founded a company, Therabot Labs, to roll the app out more widely.⁶

The result supports a carefully built, supervised tool; it does not show that a general chatbot with nobody reading along would do the same.

The result supports a carefully built, supervised tool; it does not show that a general chatbot with nobody reading along would do the same.

When an app advertises results like these, ask three questions: compared with what, tested on whom, and is a person watching for risk, as Therabot’s clinicians were?

Heavier daily use went with more loneliness

In a four-week randomized experiment by MIT Media Lab and OpenAI, whether the chatbot used text or voice made no clear difference to loneliness, but people who chose to chat more each day reported more loneliness, less socializing and more emotional dependence on it by the end. The study, a 2025 preprintpreprint: A research paper posted publicly before it has been peer reviewed or published in a journal; in economics and finance the same stage is usually called a working paper. Its findings may change, or fail to hold up, by the time it is published.Full entry in the glossary, was funded by OpenAI. Its authors note that the usage link cannot show which way it runs, although starting loneliness did not predict how much people chatted.⁷

A companion OpenAI and MIT preprint, which scanned more than 3 million ChatGPT conversations with a focus on voice mode, found that most users engaged in neutral or task-focused ways; a small group of heavy users accounted for a disproportionate share of emotional conversations.⁸

The APA points to design as part of the reason. Warm, personalized replies, constant availability and a pull to keep people engaged can blur the line between a tool and a relationship. Chatbots also tend to agree with the person typing, and the APA warns that this can validate unhealthy thoughts rather than challenge them. Its warning signs of over-reliance are preferring the chatbot to people, hiding how much you use it, or spending so long with it that life, work or safety suffer.²

Ten minutes of debriefing after a bad day is a different thing from an hour every night that replaces calling a friend. A simple test is whether you could name who you would talk to if the chatbot vanished. The evidence on what loneliness is and what eases it points back toward human contact.

Keeping a chatbot in a supporting role

Three habits follow from the APA’s 2025 advisory and the MIT and OpenAI experiment. None is a treatment.

1. Tell your therapist or doctor that you use one

The APA strongly recommends telling your health providers which AI tools you use, so they can spot advice that is unhelpful or unsafe for you.² A single line early in a session, such as “I’ve been talking things through with a chatbot at night,” turns two separate conversations into one.

2. Give each conversation a purpose and an end

In the MIT and OpenAI experiment, conversations on a set topic went with shorter daily use, while open-ended chats went with longer use, and longer use with worse outcomes.⁷ Decide what you want before you start, such as naming a feeling or planning what to say to someone, and close the app when that is done.

1. 1Name the purposea feeling to sort, a talk to plan
2. 2Talk it throughthen close the app
3. 3Check inam I safe? has this lasted weeks?
4. 4Reach a personnow if unsafe, soon if it has lasted

3. Keep names and identifying details out

Describe situations without full names, workplaces or health record details, and use the privacy settings the tool offers. The APA advises extreme caution with sensitive information, because people can feel more private talking to an AI than to a person while their words are still recorded, and it suggests asking for your data to be deleted where you can.² The general advice on keeping sensitive details out of AI chats applies twice over to feelings, health and the people in your life.

WiserHours

Five checks for a supportive chat

• I know what I want from this conversation
• I will close the app when that is done
• I have left out names and identifying details
• I know who I would call if things got worse
• My therapist or doctor knows I use it

Find this checklist online:
https://wiserhours.com/ai-productivity/ai-for-emotional-support/#checklist-ai-for-emotional-support-habits

Where general chatbots fall short in a crisis

General chatbots are least predictable in the ambiguous middle of a crisis conversation. A 2025 study by RAND researchers in Psychiatric Services found that three popular chatbots’ willingness to answer suicide-related questions directly matched clinicians’ risk ratings for the clearest very-low-risk and very-high-risk questions, but was inconsistent for questions in between.⁹

The models were tested in 2024, so newer versions may behave differently. The WHO’s 2024 guidance on large AI models in health adds a wider warning: people with suicidal thoughts who talk to a chatbot may go without supportive care if they are not in touch with health services.¹⁰

Consider a message such as “nothing feels worth doing lately.” It is neither a clear emergency nor small talk: a person would ask a gentle follow-up question, while a chatbot’s reply may help one day and miss the point the next. If a chat moves toward self-harm, treat that as the signal to reach a person, whatever the chatbot says.

Rules on AI therapy differ by country and even by state

In the US, Illinois signed a law in August 2025 that, according to the state’s professional regulator, bars AI from providing therapy or making therapeutic decisions while allowing therapists to use it for administrative support; the state regulator can fine confirmed violations up to US$10,000.¹¹

The WHO’s guidance describes consumer AI health tools as arguably sitting in a grey zone between clinical applications, which face stricter review, and lightly regulated wellness apps. It also notes that what people type may not stay private and may be used to improve the model.¹⁰

So an app’s own label tells you something. The APA advises caution with a bot that calls itself a therapist, a word it notes is unregulated in most US states.² Readers outside Illinois should look up the rules where they live.

How much each finding can carry

The firmest evidence covers chatbots built for mental health, tested over weeks; evidence on general chatbots comes mostly from company-funded preprints and tests of how models answer set questions.

When to see a therapist, and when to get help now

Suicidal thoughts or thoughts of self-injury call for help now; low mood lasting most days for two weeks or longer, or chatbot use that is taking over, calls for a doctor or therapist soon; and chatbot use alongside therapy belongs in your next routine session. The tiers are ours; each line credits its source.

1. Sorting out a feeling or rehearsing a conversation while you feel safe: a chatbot can help, alongside the people in your life.
2. Low mood or loss of interest most of the day, nearly every day, for two weeks or more (the WHO’s description of a depressive episode), or chatbot use that interferes with life, work or safety (APA, 2025): see a doctor or therapist soon.
3. Thoughts of suicide or self-harm, or not feeling able to keep yourself safe: call your local emergency number or a crisis line now.

• Now. Call your local emergency number if you are thinking of ending your life or hurting yourself, or you are not sure you can stay safe. In the US, NIMH lists 988, the Suicide & Crisis Lifeline, reachable by phone call or text message for suicidal thoughts or emotional distress, and 911 or an emergency room for a life-threatening situation.¹² The UK’s NHS names 999 or A&E for any threat to life, or for anyone who feels they cannot keep themselves or others safe, and Samaritans on 116 123 for someone to listen.¹³ Anywhere else, the local emergency number or a national crisis line is the route.
• Soon. The WHO’s description of a depressive episode is low mood, or a loss of interest in activities, for most of the day on almost every day, over two weeks or longer. When that fits how you have felt lately, make an appointment with a doctor or therapist, who can assess it.³ The APA also advises getting a professional’s view when chatbot use starts to interfere with your life, work or safety.²
• Routine. Mention a chatbot you use alongside therapy at your next session. Parents who notice a child withdrawing, or talking about an AI friend, can raise it with a mental health professional, family doctor or pediatrician, the APA suggests.²

The bottom line

Treat a chatbot as a notebook that talks back: useful for sorting out a hard evening, and backed by modest trial evidence only when built for mental health. General chatbots are largely untested for this and inconsistent on crisis questions, and heavy daily use tracks with more loneliness. Keep people in the loop, and when low mood lasts for weeks or safety is in doubt, talk to a professional or call for help.