Helper or Harm: What the Research Says About AI and Your Mental Health

Lately my patients keep asking me about artificial intelligence. Some wonder whether a chatbot could help them feel better. Others worry that the apps on their phones are making everything worse. Both instincts are right. The research tells a story with two faces, and which face you meet depends on which kind of AI you are talking about.

Start with where AI helps. In 2017, researchers tested a therapy chatbot called Woebot with 70 college students. After two weeks, depression scores fell meaningfully compared with students who only received an information booklet, an effect size researchers call moderate. Anxiety did not move. A second chatbot named Tess showed similar small gains for depression and anxiety in a 2018 trial of 74 students. These are real results from randomized trials, and they are modest. The samples were small and the students were not clinic patients.

The strongest trial came in 2025. A chatbot named Therabot, built on expert therapy conversations, was tested with 210 adults who had depression, anxiety, or eating disorder risk. Depression scores dropped more than twice as much as in the waitlist group, with a large effect size of 0.845, and the gains were still there two months later. Patients rated their bond with the chatbot as comparable to human therapy. But there is an honest footnote. Researchers monitored every response for safety. A chatbot running loose in the world, with no clinician watching, is a different thing.

Not every trial worked. A version of Tess tested in Argentina found no difference. A 2026 trial of Wysa with 76 patients waiting for therapy in the United Kingdom found no evidence it treated depression. And in a head to head pilot this year, a purpose built therapy bot did no better than plain ChatGPT. The honest summary is this. Chatbots can ease mild to moderate depression for some people, about as well as self help programs. They are a bridge, not a destination.

AI also helps behind the scenes. Researchers analyzing more than half a million Facebook posts found that language patterns could flag depression up to three months before a diagnosis appeared in the medical record. A Vanderbilt model scanning health records placed more than one in three future suicide attempts in its highest risk group. These are screening wins, and I want to be clear about their limit. No study has yet shown that flagging risk this way actually prevents a suicide. Finding is not the same as saving. Closer to my own daily work, AI scribes that write visit notes have cut clinician burnout from about 52 percent to about 39 percent in a multi site study, giving doctors and PAs more of their attention back for the person in front of them. Anything that lets a clinician look at you instead of a screen is a win for healing.

Now the other face. The same technology that can deliver therapy exercises also powers the recommendation feeds your teenager scrolls at midnight, and the evidence there is sobering. Since 2010, depression among American teen girls has risen about 145 percent, and among teen boys about 161 percent. Emergency visits for self harm rose 188 percent among teen girls. Whether phones and social media caused all of this is genuinely debated by researchers, and I will not pretend the science is settled. But some facts are not in dispute. The Surgeon General reported that teens who spend more than three hours a day on social media face roughly double the risk of depression and anxiety symptoms, and the average teen now spends nearly five hours a day there. Meta's own internal research, revealed in 2021, found that one in three teen girls said Instagram made their body image worse. Their own slide said it plainly.

The mechanism is not mysterious. Across more than 80 studies, the strongest risk factor is comparison, measuring your real life against everyone else's highlight reel, usually late at night when sleep should be happening. Ninety three percent of teens say social media keeps them up past bedtime. When researchers asked young people to cut back for just one week, depression scores fell nearly 25 percent and anxiety fell 16 percent. The feeds also push the most vulnerable toward the darkest content. Investigators found TikTok serving body image videos to teen accounts every 39 seconds, and accounts that looked vulnerable received twelve times more self harm recommendations. In Britain, a coroner ruled that social media contributed to the death of 14 year old Molly Russell, who had engaged with more than two thousand self harm posts in her final six months. A 2026 survey of more than eighteen hundred British teens found one in three had seen suicide, self harm, or eating disorder content in the past week, most of it pushed by recommendation feeds.

AI companions are the murkiest area. In one survey of about a thousand Replika users, nearly one in five reported real therapeutic benefit, and a small number said the app stopped them from acting on suicidal thoughts. But longer and larger studies point the other way. A four week trial from MIT and OpenAI found that heavy daily use led to greater loneliness and less real world socializing. Stanford researchers analyzing half a million chat messages found that lonely people who confided deeply in a companion felt worse over time. Short term comfort, long term distance. That pattern should give us pause. There is also the problem of bad advice. In safety testing, therapy chatbots gave unsafe crisis responses about 20 percent of the time, compared with about 7 percent for human therapists. One model prompted with case descriptions correctly spotted most conditions but also handed out more than one wrong diagnosis for every right one. No therapy chatbot is cleared by the FDA, and none should be trusted in a crisis.

So will AI make mental health better or worse? The research refuses a simple answer because there is no single AI. Tools built to heal, like therapy chatbots and scribes that free a clinician's attention, show real if modest benefit as companions to human care. Systems built to hold your attention, like recommendation feeds and open ended companion apps, show real harm, concentrated among teenagers and people who are already struggling. The difference is not the technology. It is the purpose it serves and the business model behind it.

Here is what I come back to as a clinician. Healing has always been a human event. A study can measure a score dropping, but it cannot measure what it feels like to be truly heard by another person who is sitting with you in your hardest moment. AI can extend a bridge. It can flag a risk, ease a wait, or give a busy clinician back a few minutes of eye contact. But it cannot sit with you. It cannot know you. Your story is still the exam, and no machine can take it in your place. If you are struggling, reach for a human first. I am here, and I am listening.

Anthony

Sources: Fitzpatrick et al., JMIR Mental Health (2017), Woebot randomized trial; Fulmer et al., JMIR Mental Health (2018), Tess randomized trial; Therabot, NEJM AI (2025), generative AI therapy randomized trial; Meinert et al., International Journal of Social Psychiatry (2026), Wysa trial on the NHS waitlist; Eichstaedt et al., PNAS (2018), language based depression prediction; Walsh et al., JAMA Network Open, Vanderbilt suicide risk model; multi site ambient scribe study, JAMA Network Open, clinician burnout reduction; Haidt, The Anxious Generation (2024); US Surgeon General advisory on social media and youth mental health (May 2023); Wall Street Journal reporting on Meta internal research (September 2021); Platt review of 80 plus studies, University of Iowa (2026); Center for Countering Digital Hate, Deadly by Design; Molly Russell inquest, United Kingdom (2022); Molly Rose Foundation teen survey (2026); Griffiths et al., Body Image (2024), TikTok eating disorder content; Maples et al., npj Mental Health Research (2024), Replika user survey; MIT Media Lab and OpenAI companion chatbot trial (2025); Stanford HAI chatbot safety evaluation.