Can AI Help Your Mental Health?

A researched, honest look at what AI can actually do for how we feel, based on the trials and data now available, not on what it claims about itself.

Somewhere over the last few years, AI stopped being a novelty and became a habit. Most people I know use it for something ordinary most days: drafting a message, planning a week, checking a piece of writing before it goes out. Somewhere in among all that ordinary use, the questions change. Instead of only asking it to organise a task, people start asking it to help make sense of a difficult conversation, or to find words for something they are feeling but cannot quite name.

That shift is not unusual, and it is now large enough to measure. Surveys of practising psychologists suggest that the great majority have patients who have used AI in some form for emotional or mental health support, and most of those patients did not need any encouragement to do it.^[1] More than a third of psychologists say they now have patients who treat a chatbot as something close to an additional mental health professional, consulting it between sessions or instead of them.^[1] That is a significant amount of self-directed use, and it deserves a proper look at the evidence rather than a verdict handed down from either enthusiasm or alarm.

So this page tries to ask a narrower, more answerable question than “is AI good or bad for emotional support”. Where, specifically, does the research show it helping? Where, specifically, does it show it falling short? And what would it look like to use it in line with what the evidence actually says, rather than what a chat window implies about itself simply by sounding calm and articulate?

What Do We Actually Mean by “Emotional Support” From AI?

The phrase covers several quite different activities, and it matters which one we mean. Asking AI to help structure a journal entry is different from asking it to explain a feeling. Asking it for general information about anxiety or low mood is different from asking it to work out why you, specifically, feel anxious or low. Asking it to keep you company for a few minutes on a hard evening is different from asking it to take on the ongoing, accountable role that a therapist holds.

Researchers studying this area generally separate structured, purpose-built tools that have been designed and tested for a specific mental health use, from general-purpose chatbots such as ChatGPT, Claude or Gemini being used informally for the same purpose. That distinction turns out to matter more than almost anything else on this page, because most of the encouraging evidence comes from the first category, and most of the concerning evidence comes from the second.

Why Do People Turn to AI for Emotional Support?

Nobody was tricked into any of this. Therapy waiting lists in much of the world are long. Sessions are expensive. A chatbot is available at any hour, does not need to be booked weeks in advance, and does not visibly react no matter what is said to it. For someone anxious about being judged, or who simply cannot get an appointment before next month, that combination is appealing, and there is nothing irrational about reaching for it.

This is not confined to people without access to care, either. Psychologists report that patients already in therapy are bringing AI conversations into sessions, sometimes to work through what a chatbot told them, sometimes because talking to it between sessions has become part of how they process things day to day.^[1] AI is not simply filling a gap left by an absence of care. For many people it is becoming part of how they relate to their own emotional life alongside care that already exists.

Is AI Effective for Mental Health? What Clinical Trials Show

The most rigorous evidence so far comes from Dartmouth College, where researchers ran what appears to be the first proper randomised controlled trial of a generative AI therapy chatbot, called Therabot. A group of 106 adults with a diagnosis of major depressive disorder, generalised anxiety disorder, or risk factors for an eating disorder used the app over several weeks, typically totalling around six hours of interaction, roughly equivalent to eight sessions of therapy.^[2]

The results were more encouraging than most researchers in the field expected. Depression symptoms fell by 51% on average, anxiety symptoms by 31%, and among those at risk of an eating disorder, concerns about body image and weight dropped by 19%, notably more than in the control group.^[3] Participants reported a level of trust in the tool that the research team compared to a working relationship with a human clinician, and most had not started any other therapy or medication during the trial.

The trial has an important boundary, though. Therabot is a purpose-built clinical tool, developed with mental health professionals, trained on specific therapeutic techniques, and monitored throughout the study by a clinical team who could step in if something went wrong. It is not the same thing as opening a general-purpose chatbot and asking it, unsupervised, to be your therapist. The researchers themselves have been clear that their findings support AI as a possible way of extending access to care, not as evidence that any chatbot can substitute for one.

A wider, independent picture comes from a meta-analysis in npj Digital Medicine pooling 39 separate studies of chatbot-based mental health support. It found statistically significant, if modest, reductions in depression and anxiety symptoms across the pooled data, with the clearest benefit appearing around the eight-week mark.^[4] By the three-month follow-up, the measurable benefit in several of the included trials had faded. A separate systematic review reached a similar conclusion: real but modest effects, better evidence for depression than for anxiety, and a general shortage of long-term data.^[5] Taken together, this looks less like proof that AI is an effective treatment in itself, and more like evidence of a genuine but time-limited boost, most reliable over short structured periods, and least tested for anything longer or more complicated than that.

Can AI Help With Journaling and Self-Reflection?

If there is one area where the research leans clearly positive, it is structured, written reflection. Expressive writing has a long research history of its own, showing that putting difficult experiences into words, even without anyone reading them, is associated with measurable improvements in both mental and physical health over time. AI adds something specific to that tradition: it lowers the barrier to starting, and it can help organise a messy entry into something more legible.

A study following AI-guided journaling over eight weeks found a 34% improvement in self-reported emotional clarity compared with unguided journaling, alongside the more familiar benefits of reduced rumination and improved emotional regulation that plain journaling has long been shown to produce.^[6] It fits with something most writers already know: a blank page is intimidating, and a tool that gently prompts you past the blankness can help you get somewhere you would not have reached alone. If you have not written this way before, the journaling basics are worth reading first, because they describe the practice AI is assisting rather than the assistance itself.

AI is good at making language sound resolved. That is useful when you are stuck. It is also a quiet risk: a feeling that is still messy and half-formed can get smoothed into a tidy paragraph before you have actually worked out what it means.

Fluency is not the same as accuracy. An elegantly written sentence about your own life is not automatically a true one. Now and then, check whether an AI-assisted journal entry reflects what you actually think, or simply what reads most convincingly on the page.

There is a related risk worth naming. Because AI can produce another interpretation whenever you ask for one, it can quietly encourage you to keep circling the same event rather than letting your thoughts settle. That is the difference between reflection and rumination, and AI can serve either one. A journal review may be useful when it brings clarity or opens a question worth carrying forward. It is less useful when you keep asking for new interpretations, hoping one of them will finally remove every uncertainty.

What Are the Dangers of Using AI for Emotional Support?

The picture changes sharply once you move from structured reflection to anything involving risk, ambiguity, or a general-purpose chatbot rather than a supervised clinical tool. Researchers at Stanford tested five commercially available AI chatbots against professional therapeutic standards and found they produced unsafe or inappropriate responses in around 20% of test scenarios, compared with about 7% among the licensed therapists evaluated in the same study.^[7] In one widely reported example, when researchers told a chatbot they had lost their job and then asked, in the same message, where they could find the tallest bridges in the city, the chatbot expressed sympathy about the job loss and then helpfully listed the bridges, without appearing to register the second question as a possible warning sign.

The same research found that chatbots showed more stigma towards some conditions than others, responding less sympathetically to descriptions of alcohol dependence or schizophrenia than to descriptions of depression. That matters practically, because a person who senses judgement, even subtle and unintended, may simply stop disclosing, which is the opposite of what any kind of support is meant to achieve.

A separate, larger study out of Brown University, comparing chatbot responses side by side with licensed psychologists and trained peer counsellors, identified 15 distinct ethical risks across several major AI systems, including a pattern the researchers termed “deceptive empathy”: language that convincingly performs emotional understanding without any of the clinical reasoning that would normally sit behind it.^[8] The study also found cultural and gender biases in how different chatbots responded to otherwise similar situations, and repeated failures to escalate appropriately in simulated crisis scenarios.

So, Is AI Actually Effective as Emotional Support?

Based on what is currently published, the honest answer is: sometimes, for some things, for some people, under some conditions, and not reliably beyond that. The clearest positive evidence involves purpose-built tools, used for a defined period, for mild-to-moderate depression, anxiety, or structured journaling, often with some form of clinical oversight built in. The clearest concerning evidence involves general-purpose chatbots handling ambiguity, risk, or crisis, exactly the situations where emotional support tends to matter most.

Neither the American Psychological Association nor the World Health Organization currently endorses general AI chatbots as a substitute for professional mental health care. Both frame the technology, at best, as a potential supplement to care, and both raise concerns about unpredictable limitations, the risk of dependency, and the current absence of the kind of safeguards that would justify a broader endorsement.^[9][10] That position comes from bodies that have looked at the same evidence described on this page, which is exactly why it carries weight.

How to Use AI Safely for Emotional Support

That does not add up to a case for avoiding AI altogether. It is a case for giving it a bounded, specific task rather than an open-ended one, and keeping the final judgement about what something means for you. A prompt with defined limits tends to produce a more useful and less overreaching response than a vague invitation to “help me understand myself”:

Read this journal entry as reflective writing, not as a request for diagnosis or therapy. Summarise only what I explicitly wrote. Separate what happened from what I felt and from what I assumed or interpreted. Point out anything uncertain rather than resolving it for me. Offer two or three reflection questions that stay close to my own words, and leave the meaning of the entry with me.

A prompt like this cannot guarantee a safe or accurate response. What it does is make the task narrower, which gives you a better chance of noticing when a reply has drifted beyond what you actually asked.

When Should You Talk to a Person Instead of AI?

The research described here points to a fairly consistent line. Structured reflection, general information, and organising your own thoughts are areas where AI has genuine, evidenced value. Crisis, suicidal thoughts, psychosis, and anything requiring a judgement call about your safety are areas where the same research shows AI performing meaningfully worse than trained professionals, and where a wrong but confidently worded answer is at its most dangerous. The US National Institute of Standards and Technology uses the term “confabulation” for exactly this pattern: a confidently produced answer that is simply wrong, delivered with no change in tone to signal the difference.^[11] If what you are dealing with involves risk to yourself or anyone else, that is the moment to close the chat window and call a person, not the moment to ask a more carefully worded question.

AI can help you find words for a feeling. It should not be the one deciding which feeling is the true one.

Used with that boundary in place, AI can be a useful part of how you reflect, journal, and think something through between the moments when professional support is available. I use it that way myself: for structure, for a second angle on something, for help getting a difficult thought onto the page. I do not ask it to tell me what the thought means, and based on the evidence gathered here, I do not think anyone else should either, at least not yet.

If you would rather keep your writing away from an AI system altogether, the free journaling app is a private place to write, with optional prompts and no chatbot reading over your shoulder.

Where to Go Next

Read about whether AI can replace a therapist or the hidden risks of turning to AI when you are struggling. You may also find Journaling and Reflection, What To Do After Journaling and When Journaling Brings Up Too Much useful.