AI uses that are banned outright

Why this list matters more than the rest of the Act

Most of the EU AI Act is about doing things properly — documenting, testing, disclosing, overseeing. This part is different. It is a list of things you may not do at all, however carefully you do them.

Three features make it the first thing a South African business should look at. It has been in force since February 2025, so there is no runway left. It carries the highest penalties in the Act. And it is the part most likely to be breached by accident, because the exposure usually comes from a feature someone switched on in a platform you bought — not from a decision anyone consciously made.

Several items on the statutory list read as exotic — real-time biometric surveillance in public spaces, predictive policing based on personality. Those are aimed at state actors. Five are not.

You may not use AI to infer how someone feels in the workplace or in education, where it does so from their biometric signals — their face, their voice, their physical reactions. There is a narrow exception for medical or safety reasons, such as detecting a driver falling asleep.

This is the prohibition most likely to reach an ordinary South African business, because the capability ships as a standard feature rather than an exotic one. Contact-centre platforms offer agent emotion analytics; video-interview tools offer emotion scoring on candidates. Where those work off voice or facial signals in a workplace or education setting, they are in prohibited territory. The boundary matters, though — see the note below on what is probably not caught.

You may not evaluate or classify people over time based on their social behaviour or their personal characteristics where the resulting score leads to them being treated badly — either in a context unrelated to where the data came from, or out of all proportion to what they actually did.

People assume this is aimed at governments. It is not limited to them, and private scoring schemes can be caught. But note that scoring on its own is not the offence: the prohibition requires the score to produce detrimental treatment of one of those two kinds. The realistic risk sits in loyalty, risk-rating and customer-tiering programmes that quietly borrow signals from an unrelated part of the business and then penalise people on the strength of them.

You may not build or expand facial-recognition databases by untargeted scraping of facial images from the internet or from CCTV footage.

For most businesses this is a supplier question rather than something they did themselves. If you buy facial recognition, ask where the training images came from and get the answer in writing. "We do not know" is not an acceptable answer when the penalty tier is this high.

You may not use AI techniques that are deliberately manipulative or deceptive, or that exploit someone’s vulnerability — their age, disability, or economic or social situation — in a way that materially distorts their behaviour and causes, or is likely to cause, significant harm.

One correction worth making to how this is usually described: the prohibition is not limited to subliminal techniques operating below conscious awareness. Purposefully manipulative or deceptive techniques are covered in their own right, whether or not the person notices them. What keeps ordinary marketing outside the ban is the rest of the test — the behaviour must be materially distorted, and significant harm must be caused or likely. Nothing here stops you recommending products well. For most businesses the practical value is knowing the outer limit exists on how aggressive an optimisation loop may become, and it is a good question to put to any vendor selling “behavioural” anything.

You may not use biometric categorisation systems that classify people individually on the basis of their biometric data in order to deduce or infer their race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation.

This one is easy to trip over without intending anything of the kind, which is why it belongs on a business list rather than a civil-liberties one. Any product that categorises people from facial images or other biometric signals should be checked for whether the categories it produces — or the ones a customer could derive from them — fall into this list.

New from December 2026

The amending regulation that came into force on 27 July 2026 added one further prohibition, applying from 2 December 2026: AI systems that generate non-consensual sexually explicit or intimate content, or child sexual abuse material. It carries the same top penalty tier. For most businesses this raises no operational question — but it is a useful reminder that the banned list is capable of growing, and that a compliance position taken once is not a compliance position forever.

These follow your output into Europe

A South African business might reasonably ask why a European prohibition is its problem. The answer is that the bans apply to providers and deployers in third countries whose systems’ output is used in Europe.

A South African supplier of interview-analytics software used by European employers cannot shelter behind its location. Nor can a South African contact centre whose agent-sentiment scores are delivered to a European client. The prohibitions travel with the output.

And if nothing you produce reaches Europe? The Act’s bans do not apply to you — but check your vendor terms before relaxing. The major AI and cloud providers have written the same restrictions into their global acceptable-use policies, so you may be bound to them by contract regardless of what the law requires. More on the vendor-terms channel.

A six-point screen for your own stack

This is the version to hand your operations team. Go tool by tool through your AI inventory and answer six questions. Anything that returns a yes and touches Europe should stop until it has been properly assessed.

Frequently asked