Large Language Models in Mental Health Support¶
J. Rantanen — M.Sc. thesis, Aalto University, August 2024
Part of Large language models and generative AI.
A note on this page. Most theses in this directory are described through the three components of a machine learning problem: the data, the model and the loss function. This one is a literature review: it trains nothing and fits nothing, so that structure would be an empty form. It is described here by its question, its sources and its findings instead.
Demand for mental health care exceeds the supply of professionals almost everywhere, and conversational artificial intelligence (AI) is being proposed to close the gap. The proposals have arrived faster than the evidence. This thesis surveys what has actually been demonstrated about large language models in mental health support, and what remains assertion.
The question¶
How can large language models (LLMs) be utilized in mental health support? The question is examined across efficacy, real user experience, and the ethical, privacy and legal constraints on deployment.
What it surveys¶
The review builds on the transformer architecture and its attention mechanism, then works through thirteen case studies. Four of them carry most of the argument:
- Psy-LLM, combining a pre-trained LLM with professional Q&A and curated psychological articles, intended as a front-end tool for healthcare professionals and a screening aid for urgent cases. Evaluated by perplexity (a language-model fit score) and human assessment; judged coherent and relevant, but short of human-level performance.
- ChatCounselor (Liu et al., 2023), fine-tuned on real client–psychologist dialogue: 260 hour-long interviews distilled into the Psych8k query–answer set and used for instruction tuning of a Vicuna base model. It surpassed open-source baselines and approached commercial-model performance, which the review reads as evidence that domain-specific fine-tuning of a modest base model beats a larger general one.
- Diagnosis of Thought prompting, which decomposes clinical reasoning into subjective assessment, contrastive reasoning to surface cognitive distortions, and schema analysis. It improved cognitive-distortion detection and, notably, produced rationales human evaluators found interpretable.
- MindfulDiary, a journalling application with an AI-driven interface and a clinician dashboard, tested in a four-week field study with 28 patients with major depressive disorder and five psychiatrists — one of the few clinical evaluations in the set.
The review also draws on first-hand accounts posted publicly by users, which is unusual for a thesis of this kind and gives it access to failure modes that published evaluations tend not to report.
What it found¶
LLMs improve accessibility and immediacy: they are available at 3 a.m. and have no waiting list. There is some evidence of reduced anxiety and depression symptoms. But effectiveness varies sharply with design and implementation, which is another way of saying the category "LLM for mental health" does not predict much on its own.
The constraints are where the thesis is firmest. Data security and informed consent are unresolved in a setting where the disclosed material is about as sensitive as it gets, and the regulatory position is unclear. The review calls for long-term studies and explicit governance frameworks rather than further capability demonstrations, a conclusion that has aged well.
Watch the defence¶
Full text¶
The thesis record is at Aaltodoc. The full text is not open access.