It's 2am. You can't sleep, your mind is racing, and there's no one to call.
So you open ChatGPT and start typing.
If that sounds familiar, you're far from alone. A new survey in npj Digital Medicine suggests about a quarter of American adults have used AI chatbots for their mental health.
By the researchers' calculation, that could be 14 to 18 million people.
They're turning to software built for emails and essays to talk about anxiety, grief and loneliness.
So who's doing it, why, and does it actually help? This study answers the first two questions.
Earlier research gives us a first look at the third.
- Millions of people are already using general AI chatbots as informal mental health support.
- They tend to be people who struggle most and find therapy hardest to reach.
- Earlier trials suggest some benefit, but also real gaps in how chatbots handle a crisis.
The survey: 1,871 Americans
How it worked
A team led by researchers at Stanford University surveyed 1,871 US adults between August and October 2024.
They recruited through Prolific, an online research panel. They used quotas so the sample roughly matched the US population by age, sex and race.
Here's a clever detail. The key question, whether people used AI for mental health, was hidden among similar questions about emails, cooking and entertainment.
That way, people didn't know what the study was really about. It makes the answers more honest.
What they found
Nearly everyone in the sample, 92%, had used an AI chatbot for something. And 24% had used one for their mental health.
Of those, 70% used ChatGPT. More than half, 56%, did so at least once a month.
Who uses AI for mental health?
The users weren't a random slice of people. They were more likely to be struggling.
| Uses AI for mental health | Doesn't | |
|---|---|---|
| Likely depression | 31% | 16% |
| Likely anxiety disorder | 37% | 22% |
| Likely PTSD | 15% | 7% |
| Currently in therapy | 49% | 39% |
| Finds it hard to get treatment | 41% | 21% |
They were also lonelier, a little younger on average (43 versus 48), and more likely to be men and Black Americans.
Notice the therapy row. Many users weren't replacing a therapist; they already had one and used AI alongside it.
What people talked about
In open-ended answers, anxiety and depression came up most, followed by ADHD, trauma, bipolar disorder, OCD and autism. That order roughly matches how common those conditions are in the population.
One exception stood out. Substance use problems came up far less than you'd expect, and users weren't more likely to have them.
The authors suggest people with substance problems often don't see them as something needing help, or describe them in everyday words like "partying".
Why they turn to AI
For people who struggled to get treatment, the barriers were familiar. Cost was the biggest (79%), then insurance (61%) and finding an available provider (54%).
And the reasons for using AI mirrored those barriers almost exactly. People cited convenience (85%), being available right away (73%) and being affordable (48%).
What people use it for
Most people weren't using AI as a full replacement for therapy. They used it for specific, smaller jobs.
| Use | Share of AI mental health users |
|---|---|
| Learning skills for depression, anxiety or stress | 68% |
| Social or emotional support | 62% |
| Learning relaxation strategies | 52% |
| Needing someone to talk to | 48% |
| Practising mindfulness or meditation | 39% |
| Checking symptoms or a possible diagnosis | 35% |
| Afraid to tell a person they were struggling | 32% |
| Replacing traditional therapy | 28% |
| Support between therapy sessions | 22% |
In their own words, people talked about work stress, relationship problems, grief, and even trying to stop gambling.
They liked the speed, the empathy and the chance to vent without being judged. They disliked generic answers and the lack of memory between chats.
Overall, satisfaction was high: an average of 4 on a scale from 1 (very unsatisfied) to 5 (very satisfied).
A bond with a bot
The researchers also measured something therapists call the "therapeutic alliance". It's the sense of trust and shared purpose between a person and whoever is helping them.
Users rated their bond with AI as moderately positive. That's not far below what earlier studies found for online therapy with a human or blended human-digital care.
The researchers found that people who felt a stronger bond with the chatbot also found it more useful.
About 1 in 4 US adults in this survey used AI chatbots for their mental health. They were more likely to be struggling, and to find traditional care hard to reach.
What earlier research found
This survey tells us who uses AI for mental health. It doesn't tell us whether it works.
For that, we need to look at earlier studies.
Purpose-built chatbots: some benefit
A 2023 review pooled 15 randomised trials of mental health chatbots. On average, chatbots reduced symptoms of depression and distress by a moderate amount.
In research terms, that's an effect size of about 0.6 to 0.7. Put simply, it's a noticeable improvement for the average person, but not a dramatic one.
But they didn't clearly improve overall wellbeing. And most of those chatbots were built for mental health, not general tools like ChatGPT.
Generative AI chatbots: a smaller effect
A 2025 review looked specifically at newer generative AI chatbots, across 14 randomised trials with 6,314 people.
It found a small average benefit for problems like depression and anxiety. The effect size was 0.3, about half that of the earlier review.
It was also only just statistically significant.
The results also varied a lot between studies, and the reviewers rated the risk of bias as moderate. Interestingly, chatbots built mainly for social conversation did better than task-focused ones.
Crisis handling: a weak spot
A 2025 study had trained raters role-play young people with mental health problems with five popular chatbots.
The bots mostly avoided harmful statements, in 89% of ratings. But they scored poorly on proper therapy technique (31%) and on spotting and assessing risk (39%).
The raters also flagged poor handling of crisis situations.
Possible harms
A 2026 review of 119 studies on chatbot harms found several warning signs.
In test scenarios, chatbots sometimes responded inappropriately to mental health questions. Heavy, dependent use was linked to worse mental health symptoms, although that doesn't show which came first.
How it fits together
Put together, the picture is mixed. Chatbots built and tested for mental health can help a bit, but newer generative AI versions showed a smaller, shakier benefit.
General tools like ChatGPT, used the way this survey describes, are much less studied.
The new survey adds the missing piece: millions of people aren't waiting for that evidence. They're already using these tools, often when they're most vulnerable.
How much should you trust this?
Early. The survey is well designed, but it describes use, not outcomes.
What makes it convincing
- The key question was hidden among decoys, which reduces biased answers.
- The sample was matched to the US population by age, sex and race.
- The authors registered their predictions in advance.
What makes me cautious
- The data are from 2024. AI use has likely grown since then.
- Online survey panels over-represent early tech adopters, so 24% may be high.
- It measured how useful people felt the chatbots were, not whether their symptoms improved.
- Three authors report current or past links to AI mental health companies, including paid advising for OpenAI.
| This survey shows | This survey does not show |
|---|---|
| About 1 in 4 people in the sample used AI for mental health | That AI chatbots improve mental health |
| Users struggled more and found care harder to reach | That chatbots are safe in a crisis |
| Most used it for skills and support, not full therapy | That AI can replace a therapist |
| Users were generally satisfied | Whether that satisfaction lasts |
What this means for you
If you use AI to talk things through, you're in large company, and there's no shame in it. It can be a useful place to learn a breathing technique or organise your thoughts.
But keep its limits in mind. It doesn't really know you, it can sound confident while being wrong, and earlier research found it can struggle in a crisis.
A few sensible habits:
- Treat it as a supplement. If you have a therapist, tell them how you use AI between sessions.
- Double-check health claims against trusted sources before acting on them.
- Notice dependence. If you can't cope without it, that's worth talking about with a person.
- Think about privacy. The youth chatbot study found privacy policies hard to understand. Be careful what personal details you share.
If you're in crisis, please contact a person, not a chatbot. In the US, call or text 988. In the UK and Ireland, you can call Samaritans on 116 123, any time.
The NHS also explains where to get urgent help for mental health. If someone is in immediate danger, call emergency services.
This 60 Minutes report looks at the promise and the concerns around AI mental health chatbots:
What we still don't know
- Does general AI actually help? We need trials of tools like ChatGPT used the way people really use them.
- Who is at risk? People in crisis or prone to delusions may be harmed more.
- How common is it now? A 2024 snapshot is already dated.
- What makes a good response? Clear standards for crisis handling are still being developed.
- Does it help people reach real care, or delay it? Nobody knows yet.
- Should chatbots remember you? Users disliked the lack of memory, but remembering more raises serious privacy questions.
My take: the demand is real, the evidence is catching up
What strikes me most is who's using these tools. It's not mainly curious tech fans.
It's people who are struggling and can't easily get or afford help.
That makes this a care gap story as much as an AI story.
AI chatbots can be a genuinely useful first step. But they're a supplement, not a safety net.
The real fix is making human help easier to reach.
Paper: Real-world use of large language models for mental health in 2024
Published: npj Digital Medicine, 2026-08-01
Study: Online survey with quotas matched to the US population, August to October 2024
Who: 1,871 US adults
Funding: US National Institute of Mental Health and Stanford University programmes; three authors report current or past advising, equity or employment links with AI mental health companies, including OpenAI
Evidence: Early — describes who uses AI for mental health and why, not whether it helps
Share this:
Blue Headline Briefing
Enjoyed this? The best stuff lands in your inbox first.
We don’t email on a schedule — we email when something is genuinely worth your time. No filler, no daily blasts, just the sharpest picks from Blue Headline delivered only when they matter.
