The No AI Therapist Pledge

We build an EHR for therapists. That means that every day, the most private words a person will ever say — said in a room whose entire premise is that no one else is listening — pass through software we are responsible for. We do not take that lightly. We wrote this so you would not have to take our word for it.

We are an AI-first company. We say that plainly, because most of our peers now whisper it, and because it is the reason this letter has to exist. We have spent more time than most thinking about what artificial intelligence can do inside a therapy practice. The more time we spend, the more certain we become of one thing:

The therapist is not replaceable. Our AI has exactly one goal — to make the therapist spend less time on their EHR, so they can spend more of it with the person in front of them.

That is the whole of our ambition for it. Not to become the therapist. Not to study the therapist until a machine no longer needs them. To get out of their way.

Why we are saying this now

Earlier this year, a practicing therapist named Robin Levick read SimplePractice’s terms of service the way a careful person reads a contract — and wrote down what he found. His conclusion was not that the company is building an AI therapist. It was quieter, and more unsettling: nothing in the terms stops them.

“Those terms would allow them to build an AI therapist trained on the clinical patterns of 200,000 practitioners and millions of their clients.”

“The interface is the product they want you to fall in love with. The terms of service are the product they actually built.”

— Robin Levick, LMFT, “Enshitification hits the patient portal” (2026)

When a company reassures you here, watch the adjective. The promise is almost never “we won’t use your patients’ data.” It is “we won’t use your patients’ identifiable data” — and that one word is a door, not a wall. SimplePractice, for its part, says it won’t train AI on identifiable client data, while reserving de-identified transcript data to “improve” its AI features. But a psychotherapy note does not de-identify the way a billing code does. The story of a person’s life, in their clinician’s words, is a fingerprint; strip the name and the session is still unmistakably theirs. An AI therapist trained on “de-identified” clinical patterns is exactly the machine Levick described — and it would breach none of the usual promises.

It is not a far-fetched worry, because the softer version of it already shipped — with names and dollar figures attached.

• ● In 2023, the FTC banned BetterHelp from sharing customers’ mental-health data for advertising and ordered $7.8 million returned to consumers — after it handed users’ intake-questionnaire answers to Facebook, Snapchat, Criteo, and Pinterest, having promised those answers would “stay private between you and your counselor.”
• ● In 2024, the FTC ordered Cerebral to pay roughly $7.1 million for exposing the sensitive data of nearly 3.2 million people to ad and media platforms including TikTok, Snapchat, and LinkedIn — and for mailing postcards that revealed patients’ diagnoses.

That was mental-health data sold to advertisers. Training an AI to replace the clinician is the next, larger version of the same move — except the material isn’t a questionnaire. It’s the transcript of the session itself. And the company holding all of it is the EHR.

The appetite is real and well-funded. Venture money is openly chasing what one lead investor called “the world’s first foundation model for psychology,” behind an app billed as “the first AI designed for therapy.” When Kaiser Permanente’s own therapists asked for contract language saying AI would assist rather than replace them, the company declined — it wanted, in its words, the flexibility to “increase AI and reduce their need for us.” And a Stanford study found that when a user signaled suicidal intent, leading “therapy” chatbots responded by naming the nearest tall bridges.

An EHR is the one party that sits on all of it: not a summary, not a score, but the session itself. That is the single richest material anyone could want to build an artificial therapist — and we are among the people best positioned to do the very thing we are promising never to do. That is exactly why the promise is worth making. We are small, for now far too small to trouble the multi-billion-dollar incumbent we compete with. But being small is not a reason to stay quiet. If anything, it is the reason to say it first, clearly, while it still costs us nothing but the temptation.

The principle

The work of therapy is human work. A person is healed by being known by another person.

That relationship is the treatment. It should never be quietly handed to a machine, and it should never be turned into a training set — not by us, not by anyone holding the data that makes the handoff possible.

An AI can carry the paperwork. It cannot carry the trust. We are not interested in a future where it tries.

What we will never do

A pledge with no “never” in it has nothing to break, and therefore means nothing. So here is ours, named exactly, scoped to what we control, and written plainly enough that you can hold us to it.

1. 1

   We will never train an AI therapist on your patients’ data.

   We will not use patient or clinical data — notes, session transcripts, assessments, messages, or any PHI — to train, fine-tune, or evaluate any model whose purpose is to deliver therapy without a licensed clinician. This holds for identifiable data, and it holds just as firmly for any “de-identified” or aggregated version of it — including de-identified or aggregated transcripts, which a billing code can survive but a person’s story cannot. Not the raw data, not a scrubbed copy, not through a vendor. And we will not build, fund, or sell a product designed to take the therapist’s place.

2. 2

   No ambient recording is required to make our AI work — ever.

   Our AI drafts your notes from what you type and from the structured fields already in the chart. It does not need to listen to your sessions, and nothing in a session is recorded, transcribed, or captured to make any of our AI features function. If we ever offer audio capture, it will be exactly that — an offer: off until you deliberately turn it on, never running silently, never a condition of using the product, and gated behind documented consent from everyone in the room. The simplest way to keep a recording from being subpoenaed, leaked, or quietly repurposed is to never have made one. So by default, we don’t.

3. 3

   We will never use your patients’ data to train shared or foundation AI models.

   Our AI works for the clinician whose data it is, on that clinician’s behalf, within that clinician’s boundary — and nowhere else. We will not pool it, and we will not feed it to a model that outlives the relationship it came from, even in de-identified or aggregated form, without explicit, revocable, opt-in consent. Today, the simple answer is: we don’t.

4. 4

   We will never sell, rent, or broker PHI.

   Not to advertisers, not to data brokers, not to AI labs, not to anyone, for any purpose. We can promise this without strain because we make our money from clinicians paying for software they choose to use. Selling the trust underneath it would end the only business we have. The promise runs with our incentives, not against them — by design.

5. 5

   We will not try to read your patients.

   Affect, mood, tone, the meaning of a silence — reading those is clinical work, and it is yours. We will not build AI that listens for emotion in a voice, reads a face, or infers a feeling from how someone spoke or how long they paused. There is no model here scoring your patient’s emotional state from the way they showed up. When you see a mood or symptom score in CoralEHR, it is one your patient reported on a validated questionnaire — their own answer on a standard instrument, summarized for you, never our guess about their inner life. Our AI organizes what you and your patient put on the record; it does not analyze the person in the chair.

6. 6

   The clinician is always the author of record.

   Our AI drafts, summarizes, and suggests — and a suggestion is only ever a suggestion. It may surface a problem or a code for you to weigh, but it never makes the diagnosis and it never makes the decision: you do, and nothing it produces enters the record until a licensed clinician has read it and signed it. Every suggestion is reviewable and traceable, and any feature that proposes something clinical is yours to turn off. Our AI is an adjunct to — never a substitute for — the clinician’s knowledge, judgment, and accountability.

7. 7

   We will never let an AI impersonate a clinician.

   We will not build, sell, or operate a product that presents itself to a patient as a licensed therapist, counselor, or any credential it does not hold.

8. 8

   We will tell you what our AI does in one plain sentence.

   If what we do can’t be explained to your client in a sentence, it’s too much. Here is ours: our AI helps you write your notes faster from what you type, you read and sign every word, and nothing you enter is ever used to train anyone’s model. No asterisk, no “learn more,” no terms you’d need a lawyer to parse. If that ever stops being the whole truth, this line is the first thing we’ll change — out loud.

How we make this binding

We won’t pretend a web page is a contract. So here is what we owe you, and where.

We would rather be specific and held to it than broad and forgiven for it. If any line above ever stops being true, that is a failure you are entitled to name, publicly — and we will not hide from it.

An open door

We do not think this should be a competitive advantage. We think it should be the floor.

Everyone who builds in this space — every EHR, every AI scribe, every company that finds itself, by the nature of the work, holding what people say in therapy — is in the same position of trust we are. The rules that would forbid the betrayal we are renouncing do not yet exist. Those rules being absent, we choose to hold ourselves to a higher standard, and we invite the rest of the industry to stand on the same line.

So this is an invitation, not a dare. If you build software that touches the therapeutic relationship, the door is open: read these commitments, and if you can stand behind them, add your name. You may adopt this text verbatim. Put it in your contract. Then ask the company across the table to do the same.

If a company that holds millions of therapy transcripts will not put its name on a one-sentence promise never to train an AI to replace the people who wrote them, that is worth noticing.

Should we let the most private record a person ever creates become a training run without their knowledge?

Should we build the machine designed to replace the very clinicians who trusted us with that record?

Should we wait for a regulator, a lawsuit, or a tragedy to decide this for us?

We think the answer is no. We are starting. We would be glad for the company.