You Cannot Sue An Algorithm, But You Can Sue The Doctor

When diagnostic AI is wrong, South African law puts the harm on the human who relied on it, not the machine that misled them.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/you-cannot-sue-an-algorithm-but-you?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

A diagnostic algorithm can be brilliant and unaccountable at the same time. It can read a scan more accurately than a tired radiologist at three in the morning, and when it is wrong, it cannot be sued, struck from a register, or called before a disciplinary committee. The human who relied on it can be all three.

CONTEXT AND BACKGROUND

AI diagnostic tools are moving into South African radiology, pathology and clinical decision support, and many of the accuracy claims are real. The governance framework has only just begun to respond. In November 2025 the Health Professions Council of South Africa published Booklet 20, its first ethical guidance on the use of AI, and the South African Health Products Regulatory Authority set out requirements for AI and machine-learning medical devices, including software that detects tumours in radiological images. The tools are arriving in wards faster than most hospital boards have worked out what they have taken on.

INSIGHT AND ANALYSIS

Booklet 20 is explicit on the point that matters most. The practitioner must always make the final decision on patient care, with accountability and oversight remaining human rather than delegated to the tool. The legal position set out by Werksmans reinforces this from the other direction, because South African law does not recognise an AI system as a legal actor, which means the software cannot hold a duty of care, be named as a defendant, or answer for its own mistakes the way a person can. The accountability has nowhere to settle except on the clinician, and through the clinician, the hospital.

This matters more because the tool’s errors are not evenly spread. Imaging models that perform strongly in their country of origin lose accuracy on external populations, with specificity falling by as much as 24 percentage points, and chest X-ray models have been shown to underdiagnose Black patients and even to infer a patient’s race from an image. Little of the diagnostic AI in local wards has been validated on a representative South African population, which is exactly the kind of thing a board should test rather than assume. The clinician who accepts the recommendation inherits an error manufactured upstream, in training data they never saw, then signs their name to it.

IMPLICATIONS

The exposure runs in an uncomfortable direction. Medical negligence still turns on fault measured against the reasonable practitioner, and the hospital is drawn in through vicarious liability for its clinicians. Yet where harm flows from an opaque model and no clear human error can be traced, that framework may struggle to place blame anywhere. Where it could not reasonably have discovered the defect, a hospital sued for AI harm may even be able to raise a complete defence under the Consumer Protection Act. This is not comforting.

An unusable algorithm owns no assets and compensates no one, so a harmed patient may be left with no recourse, and the search for a defendant returns to the person who clicked accept. A board that has never asked who answers when the tool is wrong has not removed that risk. It has only left it unowned.

CLOSING TAKEAWAY

The answer is not to refuse a technology that can genuinely save lives. It is to govern it as a clinical act rather than an IT purchase. Before deployment, a board should be able to say whether the tool was validated on a population resembling its patients, where patient data is processed, who carries the liability when it errs, and whether a clinician can meaningfully override it rather than defer under time pressure. The algorithm will never stand before the disciplinary committee. The doctor will. Hospitals should govern with that asymmetry in full view.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net