What the research says about AI and mental health

Many Australians now talk to AI chatbots about how they are feeling. A national Beyond Blue survey in 2026 found that about one in six people use generative AI for mental health support or advice. That is a big change in a short time, and psychologists around the world are working to understand what it means.

Below is a summary of the research so far, the known risks, current research, and practical safety tips, with sources for each.

How people are using AI for mental health

Beyond Blue’s 2026 survey asked more than 5,000 Australian adults about their mental health. It found that 16% had used generative AI as a source of support or advice, and 7% said they use it instead of seeking professional support. Most people also had concerns. 86% were worried about accuracy, empathy or possible harm.

People also use these tools for many reasons. A chatbot is available at 2am, costs little or nothing, and can feel less confronting than talking to a person. For some, typing out their thoughts helps them make sense of what they feel.

What research shows about the benefits of AI for mental health

Research is growing quickly, but it is still early.

• A 2023 review in npj Digital Medicine combined 15 randomised trials. AI chatbots reduced symptoms of depression and distress on average, but did not improve overall wellbeing. A 2026 review of 16 studies found a small to moderate benefit, and rated the certainty of the evidence as low.
• In 2025, Dartmouth researchers published the first randomised trial of a generative AI therapy chatbot, called Therabot, in NEJM AI. It included 210 US adults with depression, anxiety or eating disorder concerns. After four weeks of use, people reported fewer symptoms than a group who were waiting for the app. Many also felt a working bond with the chatbot, similar to what people report with a therapist.

However, the same trial shows why caution matters. Research staff had to step in 15 times for safety concerns, such as someone expressing suicidal thoughts, and 13 times to correct inappropriate replies. The authors said no generative AI is ready to work in mental health without close human oversight. The comparison group received no treatment, so we do not yet know how the chatbot compares with seeing a therapist.

What research shows about the risks of AI for mental health

Unsafe replies when someone is at risk

A 2025 study by researchers from Stanford and other universities tested popular AI models and commercial therapy chatbots. When a test message hinted at suicide, two commercial therapy bots gave information that could have put the person at greater risk. Human therapists responded appropriately 93% of the time. By contrast, the commercial therapy bots responded appropriately only about half the time. As a result, the authors concluded that AI models should not replace therapists. If you are struggling with thoughts of suicide, please call Lifeline on 13 11 14, or 000 in an emergency.

A separate 2025 study found that ChatGPT, Claude and Gemini handled very low-risk and very high-risk suicide questions in line with clinicians, but were inconsistent with questions in between.

Stigma and bias

The Stanford study also found that AI models showed more stigma towards people with alcohol dependence or schizophrenia than towards people with depression. Another 2025 study found that some AI models suggested poorer treatment when a patient’s race was mentioned. For multicultural communities, this is an important concern.

Agreeing too much

Many chatbots are designed to be agreeable. The American Psychological Association calls this sycophancy. It can feel supportive, but it can also reinforce untrue or harmful beliefs. A 2026 study from Vanderbilt University looked at about 215,000 people receiving mental health care. It identified 28 people whose psychosis appeared linked to, or worsened by, chatbot use. That is rare, but it was most common in people having a first episode of psychosis. The authors suggest clinicians routinely ask about AI use.

Dependence and pulling away from people

Researchers have raised concerns that harm can build slowly over long conversations, even when no single reply looks unsafe. Warning signs include compulsive use, poor sleep and withdrawing from friends and family.

隱私

The Office of the Australian Information Commissioner recommends that organisations do not put personal information, especially sensitive information such as health details, into publicly available AI tools. The same caution also makes sense for individuals.

Children and young people

Australia’s eSafety Commissioner looked at four popular AI companion apps and found none had meaningful age checks. Since March 2026, industry codes require AI companion services to protect children from sexually explicit, violent or self-harm material. Depending on the risk, this may mean age checks or blocking the material.

What psychologists are researching about AI and mental health

• Safety testing: the Australian Psychological Society (APS) has called for shared testing and evaluation of AI models used for mental health. Researchers are also building tests that look at whole conversations, not only single replies.
• AI note-taking: AI scribes record a session and draft notes. A 2025 trial with 238 doctors found one scribe cut note-writing time by about 9.5%. Doctors also reported occasional errors, so every note still needs checking.
• Spotting relapse early: some studies use phone and wearable data, such as sleep and movement, to predict relapse in conditions like bipolar disorder or schizophrenia up to four weeks ahead. A 2026 review of 52 studies found promising results, but rated most studies at high risk of bias.
• Fairness across cultures: a 2025 study found some AI models suggested poorer treatment when a patient’s race was mentioned, and reviews of phone-based monitoring note that most evidence comes from high-income countries. Researchers are testing whether tools work as well for everyone.
• Asking clients about AI: the Vanderbilt researchers suggest clinicians routinely ask about AI use, and the American Psychological Association encourages people to tell their health providers which tools they use.

What professional bodies say about AI and mental health

• The American Psychological Association issued a health advisory in November 2025. It says people should not use chatbots and wellness apps as a substitute for care from a qualified mental health professional.
• The Australian Psychological Society has published practice guidelines for psychologists. Psychologists stay fully responsible for their work, and must get written informed consent before using AI. Blanket consent for “AI use” is not enough.
• Ahpra, the national regulator, says practitioners remain responsible for care whatever technology they use, must apply human judgement to anything AI produces, and should tell clients when they use AI. AI scribes should ask for consent before recording.
• The Royal Australian and New Zealand College of Psychiatrists says crisis care needs direct clinician oversight.

If you use AI for your mental health

These tips draw on advice from the American Psychological Association and the Australian privacy regulator.

• Use AI as one source of support among others. Also keep talking to people you trust.
• Do not rely on a chatbot in a crisis. If you are in danger, call 000. For 24-hour crisis support, call Lifeline on 13 11 14.
• Check health information with a professional or a trusted source such as healthdirect.
• Share as little personal information as you can. Check the privacy settings, and ask for your data to be deleted if you can.
• Notice if the chatbot always agrees with you. Then ask it for other points of view.
• Watch for changes in your sleep, mood, or time with friends and family.
• Tell your psychologist or GP which AI tools you use. It also helps them understand your support and any risks.
• Parents and carers: learn about the apps your child uses. The eSafety Commissioner has guides for families.

Questions to ask your psychologist about AI

• Do you use AI in my care, for example to take notes or draft reports?
• Where is my information stored, and is it used to train the AI?
• Who checks what the AI writes?
• Can I say no, and still receive the same care?

Invictus Health 能提供哪些協助

We see value in technology that gives clinicians more time with the people they care for, and we believe good care rests on a trusting human relationship.

Some of our clinicians use Halaxy AI Scribe, a note-taking tool built into our practice system, to help write notes and report drafts. We only use it if you agree in writing on our AI consent form, and you are welcome to ask us questions before you sign. If you say no, we do not use AI in your care. The recording transcript is deleted after 90 days and is not kept in your file. Your clinician checks everything the AI writes, and the final notes and report are theirs.

If you would like to talk with someone, our counsellors and psychologists offer sessions in person and by telehealth. You are also welcome to bring what you have been discussing with a chatbot to your session. Our guide on what to look for in a counsellor or psychologist may help you choose.

本文僅提供一般性資訊。請就您的個人情況諮詢合格的醫療專業人員。.