Every story in the brief carries a small colored label. Those labels are a claim about
how this field works: forces shape what gets built, what gets built shows up in a handful of
distinct ways, and those deployments land on people. Here is the whole map.
Ten categories, three stages. Tap any one to see what it covers.
InputsApplicationsEffects
·
1
Inputs
Three forces decide what gets built and what is allowed,
General-purpose AI models and the companies building them. Decisions made here for
reasons that have nothing to do with mental health show up later in a crisis conversation.
Looks like
A new model release or a change to safety guardrails
Age verification or teen-account rules
How a model is trained to respond to distress
Policy & courtsWhat is legally allowed, and who is liable▶
Laws, regulation, and litigation shaping what can be built and deployed. Right now this
is where most of the field's boundaries are actually being set.
Looks like
A state law restricting AI therapy
An FDA decision or a consumer-protection inquiry
A wrongful-death suit against a chatbot maker
Industry & moneyWho is funding what, and for whom▶
Funding, partnerships, and business moves steering what gets built and who it reaches.
Capital decides which ideas get a second year.
Looks like
A funding round or an acquisition
An employer or insurer rolling a tool out to members
A company shutting a product down
shape
2
Applications
which reaches people as four kinds of tool, each with a different role,
Consumer chatbots & companionsNo clinician anywhere in the loop▶
AI people use directly for health advice, emotional support, or companionship. The
largest category by usage and the least supervised.
Looks like
A general assistant used as a de facto therapist
A companion or romantic AI app
Grief, journaling, and reflection bots
Clinically supervised therapy toolsAI therapy with oversight or evidence behind it▶
AI therapy deployed with clinical oversight, a real safety architecture, or trial
evidence, as opposed to a general chatbot people happen to confide in.
Looks like
A purpose-built therapy bot in a clinical trial
A prescription digital therapeutic
A health-system pilot with clinician escalation
Clinician support toolsAssists a therapist, does not replace one▶
AI that helps a clinician do the job, with a human in the loop. Includes the training
and administrative layers wrapped around care.
Looks like
A documentation scribe or note generator
Triage routing and treatment planning support
Simulated patients for trainee practice
Screening & risk detectionWatches and predicts rather than treats▶
AI that screens, monitors, or forecasts risk. It often acts on people who never asked
for it, which is what makes it its own category rather than a kind of treatment.
Looks like
Voice or text markers used to flag depression
Suicide-risk detection on social platforms
Relapse prediction from medical records
land on
3
Effects
and lands on three groups, who absorb the consequences.
Effects on individualsBenefits, harms, and attachments▶
What AI use is doing to the people using it. The fastest-moving level, and the one that
produces the headlines.
Looks like
More people seeking help, and loneliness relief
Delusion reinforcement or a missed crisis
Parasocial attachment and displaced relationships
Effects on the workforceWhat it does to clinicians' jobs▶
How AI is changing therapists' work, their skills, their liability, and the labor
fights it is starting.
Looks like
Deskilling and over-reliance on AI suggestions
Professional bodies pushing back
New hybrid job titles appearing
Effects on the health systemAccess, cost, trust, and who gets in▶
Shifts in access, trust, cost, and who actually receives care. The slowest level, the
stickiest, and the least watched.
Looks like
The total cost of care moving up or down
Public trust in the profession shifting
Triage rebuilt around a new front door
▶And then it loops back: effects reshape the inputs
A harm surfaces→lawsuits and regulators move→platforms add guardrails→what can be built narrows
The map is a circle, not a line. Today the loop mostly runs after the fact and with a lag:
the field tends to be reshaped by its failures rather than ahead of them. Commercial success
loops too, pulling money toward whatever worked last. The system-level loops are the slowest
and the hardest to reverse once set.
Asked of every story
The ten categories say where a story sits. These four questions decide whether it
is worth your time. They apply to every box on the map, which is why they are not categories of
their own.
ValidityIs the claim backed, or does it outrun the proof?▶
A randomized trial, a preprint, a company white paper, and a press release are
not the same thing. Every item in the brief carries a reliability rating for this reason.
Ethics & safetyIs this safe, consented, and trustworthy?▶
Data protection and secondary use, minors, bias across groups, how crises get
escalated, and how easily the guardrails come off.
Access & equityWho actually benefits?▶
Whether a tool closes the treatment gap, or delivers its best version to people
who already had options and a thinner version to everyone else.
What counts as careIs a bond with a machine real care?▶
If simulated empathy still helps, does the hollowness matter, and who gets to
decide what good care is: the client, the clinician, the system, or the model.