Large
Language Models (LLMs; unless otherwise
noted, quotes below taken from Rousmaniere, T.,
Zhang, Y., Li, X., & Shah, S. (2025).
LLMs are AI
systems designed to generate text by identifying and
replicating patterns learned from enormous amounts
of data—typically billions of words from books,
articles, websites, and other textual sources
(Alammar & Grootendorst, 2024 ). Unlike
traditional computer programs that operate based on
explicitly programed rules, LLMs use the statistical
relationships between words and concepts to predict
what comes next in a conversation or text. Training
an LLM involves feeding it enormous data sets and
fine-tuning its performance through human feedback,
a process known as reinforcement learning with human
feedback. During this process, humans rank or rate
responses, guiding the model toward more helpful and
contextually appropriate outputs.
Additionally,
LLMs can be fine-tuned for specialized domains to
further enhance their effectiveness in various
contexts and applications. While LLMs do not possess
genuine understanding or consciousness, they can
simulate human-like dialog convincingly by
estimating the most likely response based on
previous examples encountered during training and
what they have statistically learned from vast
linguistic data on the internet. This allows LLMs to
imitate empathy, self-awareness, and personal
experience (Alammar & Grootendorst, 2024 ; for
video tutorial on how LLMs are trained (3h30min)
Karpathy, 2025 [https://www.youtube.com/watch?v=7xTGNNLPyMI]).
How
Widespread is the Use of LLMs in Mental Health?
·
In a national study of US adults,
Scholich et al. (2025) report “A striking 24% of
surveyed participants use LLMs for mental health. These
users are more likely to be young, male, and Black
and have poorer mental health and quality of life.
They report difficulty accessing traditional
mental health treatment – particularly due to cost
and insurance coverage – and use LLMs because
they are free, convenient, and available when
needed. Users rely on LLMs for social and emotional
support, to learn therapy skills and tools, and to
supplement existing therapy.”
·
Rousmaniere et al. (2025) report
that 49% of LLM users who self-report an
ongoing mental health condition use LLMs for
mental health support.
o
73% use LLMs for anxiety management,
63% for personal advice, 60% for depression support,
58% for emotional insight, 56% for mood improvement,
36% to practice communication skills and 35% to feel
less lonely.
o
63% of users report that LLMs
improved their mental health, with 87% rating
practical advice as helpful or very helpful.
o
90% cite accessibility and 70% cite
affordability as primary motivations for using LLMs
for mental health support.
o
39% rate LLMs as equally helpful to
human therapy, while 36% find LLMs more helpful than
human therapists.
o
34% of participants indicated
ambivalence about the helpfulness of the LLMs, and
9% report encountering harmful or inappropriate
responses, highlighting the need for careful safety
research, especially in crisis usage contexts.
o
64% have used LLMs for mental health
support for 4+ months, showing stronger sustained
engagement than typical digital mental health
applications.
Problems
with Mental Health LLMs themselves
(emphases below, added; unless otherwise noted,
quotes below taken from Rousmaniere, T., Zhang, Y.,
Li, X., & Shah, S. (2025). The various citations
within the quotes below can be found in the
Reference list of Rousmanier et al., 2025)
