How our AI works

EMDR Assist does not train, fine-tune, or build its own AI model. We use general-purpose foundation models and constrain them tightly: protocol-specific instructions, structured output formats, and the client record you have already entered. The clinical intelligence comes from the scaffolding and from you — not from a model that learned on therapy data.

Everything the AI produces is labelled as a draft. Nothing is written into a client record, a treatment plan, or a note without a clinician accepting it.

This is stated in Section 6 of our Terms and repeated in our Privacy Policy: we do not use therapist or client content to train AI models, and we do not permit third-party model providers to train on your content.

To be precise about what kind of commitment this is: it is a contractual and policy commitment, backed by the model providers' API terms, not a cryptographic guarantee. We think you should evaluate it on those terms rather than on marketing language.

Each AI tool is bound to the structure of the clinical task it supports. A Calm/Safe Place script generator is instructed to follow the Shapiro-tradition Phase 2 installation sequence. A target-sequence tool follows three-pronged past/present/future logic. A session summary follows a fixed set of clinical headings.

In practice this means the model is filling in a clinically-shaped form rather than inventing a format. Language is instructed to stay consistent with Shapiro (2018) and standard AIP framing, and output is validated against a schema before it reaches you.

Drafts are generated from the material already in the client record and any notes you pass in for that specific request — history, targets, negative and positive cognitions, SUD and VoC values, session events. That record is the source of truth.

When the record is thin, the tools are instructed to produce less, not to fill the gap: fewer items, "not yet documented" placeholders, or a flexible template with explicit [INSERT] markers.

Every clinical prompt carries explicit prohibitions. Among them:

• Never fabricate diagnoses, history, or symptoms not in the notes.
• Never infer or invent measurement values — in particular, never claim a SUD reached 0 when the field is empty.
• Never fabricate references, DOIs, page numbers, or attribute a claim to an author who did not make it; prefer canonical, checkable works.
• Ground each suggested target or part in something the therapist actually said.
• Leave fields blank rather than guessing.

These constraints reduce fabrication substantially. They do not eliminate it. Citations, clinical claims, and any numbers in a draft should be verified before they enter a record.

AI surfaces in EMDR Assist are review surfaces. Suggested targets appear as suggestions until you press the button that adds them. Scripts open as editable drafts. Summaries are attached to a session for your review and edit before export. Specialized protocols carry training-required notices.

The clinical judgment, the diagnosis, the treatment decision, and the final content of the record remain yours. The AI drafts; you practice.

• We do not diagnose, and tools are instructed not to assign diagnoses.
• We do not present AI output as a clinical decision or as reviewed advice.
• We do not require identifying client information; the product is designed to be used with pseudonymous client labels.
• We do not sell your data, and we do not use it to train models.