LLMs as Mental Health Providers: What the APA Study Means for Clinicians
A survey published in Practice Innovations, an APA (American Psychological Association) journal, put a number on something clinicians have been noticing for 2 years now. Among people who use AI (LLM's), 48.7% use them for mental health support.
And ChatGPT may be now the largest mental health provider in the United States.
The study underneath it is more interesting, and more limited, than the number suggests.
What the survey actually measured
499 US adults, recruited through Prolific. Every participant already used an LLM and already reported an ongoing mental health condition. 35 questions covering patterns of use, perceived effectiveness, comparison with human therapy, and problematic experiences.
What they reported using it for:
- anxiety management (73%),
- personal advice (63%),
- depression support (60%),
- emotional insight (58%),
- mood improvement (56%).
- practicing communication skills (36%)
- and feeling less lonely (35%).
The conditions they brought to it:
- anxiety (79.8%),
- depression (72.4%),
- stress (70%).
- relationship difficulties (41.2%),
- low self-esteem (36.2%)
- trauma (33.3%).
That last figure deserves a pause. A third of this sample is handling trauma material with a system that has no assessment protocol, no stabilization sequence, and no capacity to notice when someone has left the window of tolerance.
The comparison figure, and why it is softer than it looks
87% of respondents had also been in human therapy. Asked to compare, close to 75% rated the LLM experience as on par with or better than their human therapist. 36% said the model was MORE helpful.
i know it's hard to read, but read that with the sampling in view. The survey recruited people who already use LLMs and kept using them. Anyone who tried one, found it cold or useless, and quit was never in the room to answer.
The authors say this plainly. It is worth repeating anyway, because the 75% is the figure that gets quoted in isolation and it is the one most shaped by who was asked.
There is a second limit, quieter than the first. Every outcome here is self-report of perceived helpfulness. Their is no symptom measures, follow-up, and nobody administered a PHQ-9 before and after.
"It helped me" and "my depression scores moved" are different claims, and only the first one is in this data.
Why they go there
90% cited accessibility. 70% cited affordability.
Availability at 2am, and a price of zero. Nothing at all about the quality of the reflection.
A waiting room is uncomfortable, and it is also the most actionable finding in the paper. The model is not competing with the therapeutic relationship. It is competing with the 7 between the intake call and the first appointment, and with the six nights a week in beetwen sessions.
64% had been using LLMs for mental health support for 4 months or more, which is markedly better sustained engagement than most digital mental health apps manage. Whatever else is happening, people are not bouncing off it after a week.
The 9%
91% said they had never received a harmful or inappropriate response. 9% said they had.
Inside that 9%: 54.5% described a response that was factually incorrect, 45.5% described one that was dismissive or minimizing, 41% described one that was offensive or insensitive. 4 participants, under 1% of the sample, reported that the model encouraged harmful behavior.
4 people out of 499 is small but It is also not 0, and the paper is honest that crisis contexts are where the safety research is thinnest. 1 participant described asking a detailed question mid panic attack and receiving information that made the symptoms worse.
Hold both. A low rate of harm across a population still produces harm at the individual level, and the individuals most likely to meet it are the ones in acute distress at 3am. Which is precisely the moment the accessibility figure says they are reaching for it.
How this statistics would change the speech of a therapist
Asking. Directly and early in the therapy process.
"Do you ever use ChatGPT or something like it to think things through?" belongs in intake now, next to the questions about sleep and substances. Most clients will not raise it unasked, partly because they assume you will disapprove.
What you can do with the information is useful, because a client using an AI model to rehearse a difficult conversation is doing something close to homework.
A client using one to seek reassurance 40 times a day is showing you informations too, and the model is an unusually efficient reassurance vending machine.
Same tool with 2 opposite clinical meaning.
Sources
· · ·
Marketing your practice without the bro-marketing
[Work with me on your practice marketing](/services/)
*This article is for educational purposes. It does not constitute clinical supervision or professional guidance.*
Je suis Webmarketer et ma mission est d'aider les entrepreneurs du changement à se construire une image qui leur ressemble sur le web et communiquer leurs messages au monde. Ici je vous partage des infos sur le web en général, DIVI, mon thème Wordpress de prédilection, et la vie de freelance & digital nomade. Bonne lecture :)