Intellectual property is where AI collides with the law most sharply — and where South Africa is genuinely, and confusingly, out in front. SA is the only country to have granted a patent naming an AI as the inventor (DABUS, 2021). But read the reason carefully: it happened because SA is a non-examining patent office, not because anyone ruled that a machine can invent. The same application was refused in the UK, the US, and at the EPO. That gap — between a headline and what it actually means — is the whole leaf: who owns AI output, whether you can train on protected works, and why a global tool will apply the wrong law to all of it.
South African IP rests on four statutes: the Copyright Act 98 of 1978, the Patents Act 57 of 1978, the Trade Marks Act 194 of 1993, and the Designs Act 195 of 1993. Copyright is automatic on creation; patents, trade marks, and designs are registered through the Companies and Intellectual Property Commission (CIPC). For most of their history these Acts asked human questions about human creations.
AI forces two new ones, and SA answers both differently from the jurisdictions a global tool learned from. Who owns what an AI produces? — a copyright-authorship and patent-inventorship question. And can you lawfully train a model on copyrighted works? — a text-and-data-mining question the current Act never contemplated. The SA answers are distinctive, unsettled, and exactly where a confident foreign model is most likely to mislead.
On 28 July 2021 the CIPC published the acceptance of South African patent 2021/03242, listing DABUS (Device for the Autonomous Bootstrapping of Unified Sentience) as the inventor, with its developer Dr Stephen Thaler as the patentee. It was the first time any patent office in the world had granted a patent naming an AI as inventor. Headlines followed; some read it as South Africa deciding that machines can invent.
That reading is wrong, and the honest version matters. South Africa runs a non-examining (formalities-only) patent system: the CIPC checks that the application is formally in order — the right forms, fees, and disclosures — but does not substantively examine whether the invention is novel, inventive, or whether an AI can lawfully be an inventor. The same DABUS application was refused in the UK, the US, and at the EPO, where examiners do decide those questions. The SA grant is a feature of the filing system, not a considered ruling of law.
"SA recognises AI inventors" is the confident, shareable, wrong summary — and it's exactly what a generative tool will produce. The accurate statement is narrower: SA granted one AI-named patent because it doesn't examine for these things, the question remains legally untested in an SA court, and building a strategy on "AI can be an inventor here" is building on a headline. This is the leaf's warning in miniature: the flashiest AI-and-IP fact about SA is the one most often stated backwards.
The live global fight — is training a model on copyrighted works infringement? — has an uncomfortable SA answer today: probably not clearly permitted. The current Copyright Act's exceptions are narrow "fair dealing" categories, and using copyrighted content to train a model may fall outside them. There is no explicit text-and-data-mining exception. So an SA developer training on scraped or even lawfully-acquired protected works is on uncertain ground under the law as it stands.
That may change. The Copyright Amendment Bill (B13F-2017) proposes a US-style fair use clause (a four-factor test: purpose, nature, amount, market effect) that could open room for transformative training. The Bill has been passed by Parliament and is before the Constitutional Court on its constitutionality — so the rule that governs whether SA can build models on local data is, right now, undecided and in the highest court. That is a genuinely consequential open question a foreign tool won't even flag.
The Copyright Amendment Bill has been in motion for years and its fate is in the Constitutional Court. Treat any statement about SA fair use — including a confident one from an AI — as a claim to check against the current status, not settled law. What's certain is the shape of the question; the answer is still being written.
Set aside the AI-and-ownership questions and there's ordinary IP work AI genuinely speeds up: prior-art and patent search across large corpora, trademark screening for similar marks, drafting first passes of patent specifications and IP clauses, portfolio management (renewals, classifications, deadlines), and summarising long technical disclosures. As everywhere in the tree, the tool reads and drafts; the registrability call, the novelty judgement, and the infringement opinion stay with a person — and the CIPC filing still has to be right.
The single most likely wrong answer: that SA law recognises AI inventors. It granted one AI-named patent because it doesn't examine. A strategy built on the headline is built on sand — the question is untested in an SA court.
SA's computer-generated-works rule helps, but ownership still turns on a real human "arranger" and originality. For AI-assisted deliverables, who owns what must be set in the contract — not assumed because you ran the tool.
There's no TDM exception in the current Act, and fair dealing is narrow. Training a model on copyrighted works in SA is legally uncertain now — the fair-use Bill that might change it is still in the Constitutional Court.
An ungrounded model answers SA IP questions with US or EU rules. On authorship, inventorship, and fair use those are precisely the areas where SA differs. Confident, fluent, and jurisdictionally wrong.
IP work involves unfiled inventions and confidential know-how. Putting them into a model whose terms allow training on inputs can destroy novelty or leak the secret. The tooling choice is an IP-protection decision. See the Data privacy & POPIA leaf on residency.
Search and drafting: prior-art sweeps, trademark screening, specification first drafts. The tool compresses the research; the registrability opinion, the claim strategy, and the CIPC filing stay with the attorney who signs them.
Where the ownership question bites: who owns the AI-assisted code, design, or content, and can it be protected. SA's computer-generated-works rule gives more to work with than most jurisdictions — but only if the arrangements and the contract are set up to claim it.
Where the training question decides the model: what data can lawfully be used, under an Act with no TDM exception and a fair-use Bill still in the Constitutional Court. The safe posture is lawfully-licensed data and a watching brief on the Bill — not a US fair-use assumption imported wholesale.
SA's IP-and-AI story is unusually rich — a world-first AI-inventor patent, a statutory author for computer-generated works, a fair-use Bill in the highest court. But every one of those facts is narrower or more uncertain than the headline. That's the SA context in a sentence: the country is genuinely ahead on the questions, and the confident answers a global tool gives are the ones most likely to be wrong.
Unlike settled areas, AI-and-IP in SA is live: the DABUS question is untested in court, the authorship rule is under strain from autonomous output, and fair use is sub judice. Advice has a shelf life. The discipline is to state the current position, flag what's moving, and verify before relying — the opposite of a model's confident, timeless tone.
Novelty can be destroyed by disclosure, and trade secrets by a leak. Feeding unfiled inventions or know-how into a model that trains on inputs is an IP risk before it's a privacy one — and the residency answer the tree gives applies: keep it in country, on a tool that doesn't train on your data. See Data privacy & POPIA.
IP is the practice where the SA position is most distinctive — and where the confident foreign answer is most likely to be exactly wrong.
The Acts and the registry — and note the Copyright Amendment Bill is before the Constitutional Court, so verify its status before relying on the fair-use position.