Picture the session. A client finally says the thing they’ve never said out loud — the abuse, the affair, the relapse, the thought they’ve been too afraid to name. It’s the breakthrough you’ve spent months building toward. On some platforms, it’s also a data point in a dataset that’s being used to make AI sound more like you.

This isn’t a hypothetical. AI companion and “AI therapy” apps already exist, already talk directly to clients, and already improve by learning from real therapeutic language — the pacing of a reflective statement, the phrasing of a validating response, when to ask an open question versus sit in silence. Every one of those apps needed training data to get good at sounding therapeutic. Some of it came from clinical literature. Increasingly, some of it comes from real session recordings, captured by the same “helpful” AI notetaker you might be using to save yourself twenty minutes of charting.

We read the fine print. It’s worse than you think

We’re not going to name names. But we went through the terms of service and privacy policies of a number of AI notetaker platforms currently marketed to therapists, and here’s what stood out: several are written vaguely enough — around “aggregate data,” “de-identified data,” or data used “to improve our services” — that they leave the door open for your session content to be used to train models, without ever using the word “training.”

And even the platforms with reassuring language today can rewrite the rules tomorrow. Almost every one of these contracts includes a clause letting the company update its terms at any time, often with nothing more than a notice buried in an email you’ll never open. The privacy policy that made you feel comfortable when you signed up isn’t a fixed promise — it’s a snapshot. What you agreed to in January isn’t necessarily what governs your clients’ data in December, or what trains whatever model that data feeds next year.

The part that should actually keep you up at night

Here’s what almost nobody in this field has fully sat with: the more session data these companies collect, the less they eventually need you. Not because you’re replaceable — the therapeutic relationship, attunement, years of clinical judgment aren’t things a transcript captures — but because a model trained on enough sessions doesn’t need to be as good as you to be good enough to sell. It needs to be good enough to compete on price, availability, and convenience, to a client who can’t get an appointment for three weeks and doesn’t know the difference yet.

Every recording you hand over — for the sake of saving yourself a note — may be one more brick in the wall of a product built to be a cheaper, always-available substitute for what you do. Your clients trusted you with their worst moments so you could help them. They didn’t sign up to help train the thing that competes with the profession.

What we’d want, if it were our own sessions

We built NoteNest so this dilemma never has to exist in our own practice. No microphone. No audio file sitting on a vendor’s server, wondering what it might become useful for. No session content that could ever end up as a training example for a tool built to replace the work you spent years training to do. No Ai! Just clinical documentation you write yourself, generated fast, with nothing captured that didn’t come from your own judgment.

See how NoteNest keeps every session exactly where it belongs — between you and your client, and nowhere else.