Generative AI at AI-M: From Research to Clinical Translation
At AI-M, we use generative artificial intelligence as a translational tool across psychiatry, education, and clinical care. Our work spans the full lifecycle – from foundational research and evaluation to responsibly governed integration into real-world settings – ensuring that generative AI supports mental health research, training, and care delivery in clinically meaningful ways.
Across applications, we focus on using generative AI to augment clinical and scientific workflows while maintaining human oversight. Tools are evaluated for safety, interpretability, equity, and clinical fit, with clear limits on autonomy to ensure alignment with professional judgment and patient care.
We advance generative AI across psychiatry by translating research into responsibly governed tools that augment clinical care, education, and discovery.
In clinical care, generative AI can support mental health and psychiatry through documentation assistance (e.g., drafting progress notes), symptom and risk screening prompts, patient-facing psychoeducation, and decision support that helps clinicians navigate guidelines and differential diagnoses. AI chatbots can extend access between visits by delivering structured check-ins, skills practice (CBT/DBT-informed exercises), and triage guidance, provided they are clearly positioned as adjuncts, not substitutes for professional care. Emerging “agentic AI” systems go further by coordinating multi-step care workflows (e.g., monitoring patient-reported outcomes, flagging deterioration patterns, proposing outreach, and preparing clinician summaries), which raises critical requirements for oversight, escalation pathways, audit trails, and rigorous safety evaluation.
Can LLMs generate PHQ-8 directly from interview transcripts?
What we did:
Used the DAIC-WOZ dataset (clinical interviews + PHQ-8 scores)
Compared GPT-4o, Llama 3, Cohere, Gemini
Zero-shot learning, no fine-tuning
What we found:
GPT-4o performed best (Accuracy ~76%, F1 ~0.74)
Llama 3 excelled at detecting anhedonia
Cohere was best at psychomotor changes
Gemini helped with appetite/failure items
Model ensembles may offer the most robust results
Why it matters:
Could complement clinician assessments
Potential for scalable, automated MBC
Generative AI & Education
In education, generative AI can personalize learning for trainees and clinicians by generating case vignettes, simulating patient interviews, offering feedback on clinical reasoning, and creating targeted refreshers on psychopharmacology, psychotherapy principles, and legal/ethical scenarios. AI chatbots are particularly useful as low-stakes “practice partners” for motivational interviewing, risk assessment phrasing, and culturally responsive communication, when designed with clear guardrails and faculty-reviewed content. Agentic AI can support structured competency development by tracking progress across modules, scheduling spaced-repetition review, and tailoring scenarios to observed gaps, while maintaining transparency about sources and ensuring that learners do not over-trust fluent but potentially incorrect outputs.
Generative AI & Research
In psychiatry research, generative AI can accelerate hypothesis generation, protocol drafting, literature synthesis, and analytic workflows, especially for unstructured data such as clinical notes, transcripts, wearable streams, and ecological momentary assessment responses. AI chatbots can be deployed as research instruments to deliver surveys, collect longitudinal qualitative data, and test intervention engagement strategies, while carefully managing consent, privacy, and the risk of inadvertently providing clinical advice. Agentic AI is especially relevant to “closed-loop” digital mental health research: an agent can plan and execute sequences like identifying eligible participants, monitoring adherence and outcomes, adapting prompts/interventions, and alerting study teams, necessitating strong governance around bias, confounding, reproducibility, model drift, and human-in-the-loop adjudication for safety-sensitive decisions.
Generative AI & Quality Improvement
For quality improvement, generative AI can help teams identify process bottlenecks, summarize incident reports, detect variation in care pathways, and produce actionable dashboards and narrative summaries for stakeholders. In mental health settings, chatbots can support routine measurement-based care by collecting symptom scales, side-effect checklists, and experience-of-care feedback, then summarizing trends for clinicians and services. Agentic AI can automate QI cycles by continuously monitoring key metrics (e.g., follow-up timeliness, no-show risk, medication monitoring compliance), proposing targeted interventions, and generating Plan-Do-Study-Act documentation, while requiring careful calibration to avoid reinforcing inequities and to ensure that recommendations are traceable, reviewable, and aligned with clinical realities.
Generative AI & Information/Misinformation
Generative AI can improve mental health information access by translating complex concepts into plain language, tailoring content to reading level and culture, and providing structured guidance on when to seek urgent help. At the same time, it can amplify misinformation, hallucinated claims, overconfident medical advice, stigmatizing narratives, or persuasive but incorrect explanations of symptoms and treatments, especially in chatbot formats that users may treat as authoritative. Agentic AI increases both promise and risk: systems that proactively push content or guide decisions across platforms must be constrained by verified sources, clear uncertainty communication, robust safety policies (including crisis response), and monitoring for harmful trajectories such as reinforcement of delusions, disordered eating, or self-harm ideation.
