Using AI for clinical documentation responsibly: a checklist for therapists

The checklist

• Consent: the client knows AI is used, what it does, where their information goes, and that they can say no without any change to their care.
• Where the data goes: you know whether audio and transcripts stay on your computer or go to a company, which company, for how long, and whether they are used to train a model.
• Every line reviewed: nothing is signed until you have read it all against your memory of the session.
• Nothing the session did not support: no diagnosis, severity or risk wording that was not said or assessed; no invented quotes, dates or history.
• Nothing important missing: risk asked about and the answer, consent, safety plans and anything unusual are in the note.
• Bias checked: language about culture, gender, disability, class or religion is accurate, respectful and the client’s own where possible.
• Your voice and judgement: the assessment and plan are yours, not a summary the tool chose.
• You sign: the note is signed by you, and corrections after signing keep the earlier version.
• Your rules: you have checked what your regulator or professional body says about AI and recording where you practise.

Reviewing a draft well

Language models write fluently, and fluent text is easy to trust. The World Health Organization’s guidance on large multi-modal models names automation bias, the habit of accepting what a system produces without enough scrutiny, as a risk for health professionals. A few habits keep the review honest:

• Review soon after the session, while you still remember it.
• Check every quote, number, name and date against the transcript or your memory.
• Read the assessment and plan slowly: they are where an unsupported conclusion does the most harm.
• Delete what you would not have written yourself.

Bias

AI systems learn from data that carries the patterns of the world that made it. In a widely cited study, Obermeyer and colleagues found that an algorithm used to manage the care of millions of patients underestimated the needs of one group, because it predicted health costs rather than illness. A note-writing model can do something similar on a small scale: describing the same behaviour differently depending on who the client is. Read for it, and correct it.

How Olive Grove helps

In Olive Grove, speech recognition and the AI that drafts notes run on your computer, and there is no online AI provider. Every draft is marked as an AI draft. The review bar counts Risk, Conflicts, Unsupported and Missing items; Show sources puts beside any sentence the transcript lines it most likely came from; and a second pass flags risk the note missed and statements the transcript does not support, without changing the note. Sign note asks you to confirm you have read it and it is accurate, and a signed note is locked: Unlock to amend keeps the signed version and records why. Check my notes looks across your notes for diagnosis, severity or risk wording the session never used.

See Review, correct and sign a draft note, Check and fix your notes in bulk, Notes, Where the data goes and Private AI for therapy notes.