AI scribe consent in South Africa: what HPCSA Booklet 20 requires

Most guidance on AI scribe consent is written for a US audience. It talks about HIPAA, state wiretapping statutes, and Business Associate Agreements. None of that maps directly onto a South African practice.

HPCSA's updated Booklet 20, covering the ethical use of artificial intelligence in healthcare, is the guidance that actually applies here. It sets out clear obligations around disclosure, consent, and patient autonomy when AI tools are used in care. For any practice using an AI scribe, or considering one, this is the standard to work from.

What HPCSA Booklet 20 requires

HPCSA's Booklet 20: Ethical Guidelines on the Use of Artificial Intelligence (September 2025) connects AI use to HPCSA's existing ethical framework rather than treating it as a separate category. Three requirements matter most for AI scribing specifically.

Patients must be informed when an AI tool is used in their care. This isn't satisfied by a line in a general terms-of-service document a patient signed on their first visit. It means the patient understands, at the point an AI scribe is being used, what that involves.

Patient autonomy, confidentiality, and informed consent must be preserved. Consent has to be real, not assumed. A patient hearing "I'll just record this so I don't have to type while we talk" has not been asked. They've been told.

A patient who declines AI use must not be disadvantaged or refused care as a result. This is the requirement most practices haven't fully worked through, because it isn't just about the consent conversation. It's about whether the alternative to AI is actually as good.

Why "no disadvantage" is the hard part

Getting consent right is a training and process problem. Most practices can solve it with a script and a consent form.

Avoiding disadvantage is a systems problem, and it's easy to get wrong without realising it.

If a practice's documentation workflow is built around AI as the primary path, with manual note-taking as a slower fallback, then a patient who declines has, in practice, made the clinician's job harder and the record slower to produce. The clinician may not resent the patient for it. But the incentive is there, and over enough consultations, it shapes behaviour. A rushed manual note is not equivalent care, even if nobody intended it that way.

The same logic applies to the clinician's own preference. A system where declining AI means losing access to structured templates, auto-populated patient details, or consistent formatting isn't offering a real choice. It's offering AI, with a worse-supported alternative attached.

For a documentation system to meet Booklet 20's standard, both paths, AI-assisted and manual, need to produce a record of comparable quality, speed, and completeness. That's a product design question as much as a policy one.

What this looks like by specialty

The practical shape of consent and choice changes depending on the kind of consultation.

In psychology and counselling, session content is often sensitive by design, and a patient's willingness to speak openly can depend on their certainty about what happens to their words. Some patients who are otherwise comfortable with technology will still decline a recorded session, and that decision deserves the same weight as any other clinical boundary a patient sets.

In general practice, consultations are short, and the consent conversation has to happen quickly without derailing the visit. A workflow that makes AI opt-in, rather than something the patient has to actively refuse, tends to produce more genuine consent than one built around an assumed yes.

In allied health disciplines involving physical assessment, such as physiotherapy or chiropractic care, the clinician is already dividing attention between the patient and the documentation task. An AI scribe can reduce that split. But it also means the audio capture is running during a physical examination, which is a different privacy consideration than a purely conversational consultation, and worth naming explicitly in the consent conversation rather than assuming it's covered by a general policy.

None of this makes AI scribing unsuitable for any of these settings. It means the setting changes what a genuine consent conversation needs to cover, and a single generic disclosure line won't do that work.

Clinician choice sits alongside patient choice

Booklet 20's language focuses on the patient. But a clinician's decision not to use an AI scribe deserves the same structural protection, for reasons that have nothing to do with the patient in the room.

A clinician might be unsure about reviewing AI-generated content thoroughly enough under time pressure. They might simply document better by writing in their own words as they go, rather than editing someone else's draft afterward. Professional judgement about how to document care sits with the person doing the documenting, and a system that quietly nudges every clinician toward AI, through faster defaults or friction on the manual route, is overriding that judgement rather than supporting it.

Where Bookem supports AI scribe consent by design

Bookem's AI scribe sits inside the practice management system, not in front of it. A clinician can record, dictate, or upload a session and let the AI scribe draft a SOAP note, referral letter, or medical aid motivation from it. Or they can write directly into the same document templates, with patient details, clinical alerts, and prior notes already populated, with no recording involved at any point.

Both paths use the same structure, the same patient record, and the same templates. Neither is the stripped-down version of the other. That matters in the consent conversation itself: a clinician can tell a patient, honestly, that declining the AI scribe changes nothing about the completeness or quality of their record. Bookem is also built on POPIA-compliant data handling, which sits alongside Booklet 20's requirements rather than replacing them - consent and data protection are related questions, but they're not the same one.

Ready to see how Bookem supports genuine AI scribe consent, for your patients and your clinicians? Book a demo with Bookem

Frequently asked questions

Do patients have the right to refuse an AI scribe in South Africa?

Yes. HPCSA's updated Booklet 20 on the ethical use of AI in healthcare requires that patients be informed when AI tools are used in their care and that their consent be genuine and voluntary. A patient who declines cannot be disadvantaged or refused care, so the practice must have a manual documentation workflow that works just as well.

What must a practice disclose before using an AI scribe under HPCSA Booklet 20?

A practice must tell the patient that an AI tool will be used in their care, in plain terms, before it's used, not buried in a general consent form signed at intake. The disclosure needs to give the patient a genuine opportunity to decline, and that decision needs to be documented and respected without affecting the care they receive.

Does declining an AI scribe mean a patient gets a lower-quality clinical record?

It shouldn't, and if it does, the practice's system is the problem rather than the patient's choice. In Bookem, document templates auto-populate patient details, clinical alerts, and history whether the content comes from a recording or is written manually. The structure and completeness of the record don't depend on whether AI was involved.

Can a clinician decline to use an AI scribe even if the practice has adopted one generally?

Yes. HPCSA's guidance centres patient consent, but the same principle of genuine choice applies to the clinician deciding how to document care. A system that makes manual documentation slower or less supported than AI-assisted documentation is effectively removing that choice, regardless of official policy.