How does AI therapy work?
AI therapy works on three layers. A language model that talks. A memory system that holds your story. A safety classifier that escalates when something serious surfaces. None of these on their own is enough — together they make a tool that meets you where you are.
By The Calm AI Therapy editorial team. Not reviewed by a licensed clinician, and not a substitute for one. How we research and write.
The model: trained for stance, not just chat
The underlying model in Calm AI Therapy is prompted into a specific therapeutic stance — validate-and-change, not yes-person. Anti-sycophancy rules. Anti-reassurance in anxiety contexts. Cultural humility. The model is the tool; the prompt is the training. Both matter.
The memory: structured, longitudinal, yours
When you share a name, a relationship, a recurring feeling, that fact is stored against your account as a memory node. On every new session, relevant nodes are surfaced into the conversation. Not by being read aloud — by being available for the agent to reference if the moment calls for it.
The safety system: rule-based first, AI-augmented second
Every message you send is screened by a classifier that tags it with a tier from 0 (normal) to 3 (immediate danger). Tier 1+ activates a different prompt — one written by clinicians — that surfaces region-specific resources and follows a script the model cannot freelance through.
Common questions
Does AI therapy use ChatGPT?
Calm AI Therapy runs on a model from OpenAI — and the difference is the prompt, memory, and safety architecture wrapped around it. The model is one input; the system is the product.
Is the memory accurate?
It's only as accurate as what you've shared. If you contradict something, the system updates — you're always the source of truth.
Does AI therapy follow CBT or DBT?
Calm AI Therapy draws from CBT, DBT, ACT and IFS, and adapts to the user — it doesn't impose one framework.
Sources
- The Supportive Accountability Model in Digital Health: a scoping review — Journal of Medical Internet Research, 2025
- Baumel et al., Objective User Engagement With Mental Health Apps — Journal of Medical Internet Research, 2019
See it in one message on the landing page.
Try one sentenceRelated