People already type legal questions into general-purpose assistants. The answers are fluent. That is the risk, not the feature.
A model trained on the open web can name a real Act, then attach a section that was never enacted. It can attach a real section to the wrong problem. It can mix the Basic Conditions of Employment Act with the Labour Relations Act, or with another country’s hours rules, and still sound certain. The paragraph looks like research. It is a completed sentence.
Lawyers have been sanctioned for filing invented citations. A generated citation is not a source you can open. The point does not depend on naming any one product.
Why does this happen? Those tools are built to continue a sentence from patterns in training data. They are not built to retrieve a licensed copy of the statute, show the portion they found, and stop if they found nothing. When the training data is thin, as it often is for South African legislation compared with United States or English material, the model still answers. It fills the gap with something that looks like a section number.
Any tool that answers from memory of the internet, not from a retrieved statute, can do the same. The risk is higher in a smaller jurisdiction. South African legislation is less represented on the open web. A model that has seen many foreign code sections will still try to sound local. It will guess.
NOMOS Legal AI is a South African product, built in Midrand and served at nomoslegalai.co.za. When you ask a question, NOMOS retrieves licensed legislation first. Answers are built from those excerpts. If retrieval cannot find an on-point source, NOMOS says so. It will not invent a section to finish the paragraph.
That is still not legal advice. Research assistance means you get a trail: Act, portion, excerpt. You open the Act. You check that the portion is the one that governs the facts in front of you. You do not paste a generated paragraph into a brief because it arrived with a citation mark.
Retrieval can miss the right section. A heading can look relevant and not be. Always open the source. The limits of coverage and the refusal behaviour are on Trust. If you want to ask a South African question against licensed collections, open the workspace.