Reimagining Mental Healthcare in the Age of AI

Session Focus

Fava opened with the workforce arithmetic. The United States is short more than 50,000 psychologists and over 14,000 psychiatrists, and those gaps cannot be trained away in the years available. That is the backdrop against which the panel examined what AI in mental health actually means today, where it helps first, and what safety requires.

General Purpose Versus Purpose-Built

Sabine Wilhelm PhD drew the distinction that organized much of the session. Most people with mental illness currently receive no care, and those who do often do not receive good care, which is why they look for help wherever they can find it, including general purpose models.

Her analogy: a general purpose model is a Swiss Army knife that can do all kinds of things but does not necessarily help with mental health concerns well. Purpose-built tools are a different category because they go through safety testing and carry the efficacy and privacy requirements healthcare expects.

Luis Voloch added structure beneath that. Among tools actually built for mental health, some are made primarily for clinicians, assisting with training, scribing and notes, which most mental health clinicians have now used. Others are patient-facing, some FDA regulated and some marketed as wellness products.

Within the patient-facing category he described a spectrum. At the conservative end, tools focus on the bookends around a care encounter, preparing for an appointment and following up afterward. At the other end, they attempt to deliver direct behavioral interventions and protocols.

What Consumers Are Already Doing

Paul English urged the panel not to be dismissive of the horizontal platforms, while agreeing purpose-built software does the job better.

His figures: more than 50 percent of American adults have used ChatGPT as a therapist, and two-thirds of American teenagers have AI companions.

He was specific about why, despite hallucination and the risk of acting clinically without qualification. It is available at 10 p.m. on a Tuesday or at 2 a.m. during a panic attack. It is free or very inexpensive. And some patients report feeling safer describing something they are vulnerable about to a machine than to a human, because they are embarrassed by a behavior or a pattern of thought.

His proposed path forward had two parts: embrace these tools and bring expertise to improving them so they do less harm and more good, while developing purpose-built applications grounded in the evidence-based research institutions like Mass General are known for.

Wilhelm reinforced the point with her own data. According to Harvard Business Review, the most common use of ChatGPT in recent months was emotional support. She attributed that partly to the provider shortage and partly to something else: not everyone wants what the system currently offers. Seeing a clinician face to face every Tuesday at 3 p.m. does not fit many lives, given jobs and childcare, and for some the stigma of attending a named institution is itself a barrier.

Her conclusion was about flexibility rather than replacement. Some patients will always want to drive to her office and sit with her, which is fine. Others would much rather have a digital tool. Some want both. The goal is a system considerably more flexible than the current one.

Where AI Helps Today

Wilhelm noted that clinicians, herself included, already use AI as an augmentation tool. It handles administrative tasks that drive burnout, particularly note-taking. It supports clinician training and decision-making. And it surfaces patterns a clinician might miss, such as flagging that a patient’s depression scores have worsened over several weeks.

On patient-facing treatments, she was measured. Some standalone options show early promise, but the clinical trials are not yet at the standard typical of clinical care, lacking long follow-up and adequate comparison groups.

Mill Brown MD described deployment at scale, across all 50 states and roughly 200 countries within one ecosystem.

One tool collects information through a task-focused chat interface in the days between a patient booking a first appointment and attending it. That information is presented to the provider in digestible form within the EHR, so the first appointment is not spent collecting routine history and the clinician arrives knowing where to focus.

His adoption data for AI scribing is the most instructive detail in the session. On initial deployment, close to 70 percent of patients consented to real-time transcription of their sessions. Providers opted in at only 20 to 25 percent.

Closing that gap took a long time and required working through what the value proposition was and addressing clinician safety concerns. Providers were protecting patients, while the patients were already opting in. Over roughly a year, provider adoption rose to 65 to 70 percent.

What moved them was not faster notes. The tool helps write better treatment plans and captures what the clinician asked the patient to do between sessions, which other tools can then support.

Voloch named the least glamorous and most immediately useful application: basic care continuity and reminders for patients, which emerged from extensive user interviews with patients and clinicians when his company was starting.

Safety Layers

Brown described building safety into both directions of the interaction, which he identified as what distinguishes purpose-built tools from general models.

On the input side, when a person speaks to an AI tool, safety components detect whether someone may be suicidal, may be a victim of interpersonal violence, or presents other safety concerns.

On the output side, controls check that the tool responds appropriately, is not hallucinating, is not going off task and is not offering clinical advice it should not give.

Asking Patients About Their AI Use

Wilhelm made the clinical case for treating AI use as something to ask about directly, the way clinicians routinely ask about sleep, coffee and alcohol.

Her reasoning is that use varies enormously. Some patients use these tools well. In her OCD program, others spend hours seeking the same reassurance repeatedly, which is the compulsion itself finding a new outlet rather than any benefit.

Payment and Regulation

Voloch argued the binding constraint now is payment rather than capability, and named a recent CMS step as a beginning.

His concern is regulatory fragmentation. A patchwork of state-level rules creates uncertainty where clear federal regulation would let developers build with confidence.

Key Takeaways

1. The workforce gap cannot be trained away. More than 50,000 psychologists and 14,000 psychiatrists short.

2. Consumers are already there. Over half of American adults have used a general model as a therapist.

3. Availability and anonymity are real advantages, not just convenience. 2 a.m. access and reduced embarrassment matter clinically.

4. Clinicians adopted scribing more slowly than patients. Provider opt-in rose from 20 to 25 percent up to 65 to 70 percent over roughly a year.

5. Safety layers run in both directions. Risk detection on input, and controls against hallucination and unqualified advice on output.

6. AI use belongs in the clinical interview. The same tool can help one patient and feed another’s compulsion.

7. Payment, not capability, is the current constraint, with state-level regulatory fragmentation adding uncertainty.