Your AI Vendor Says You Own the Output. South African Law May Disagree.
A contract cannot override the Copyright Act, and in-house counsel should draft for the gap between the two.
Sign up for my Substack daily AI newsletter here.
See my AI Training course portfolio for corporate Business Leaders here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Most AI platforms tell business users that they own what the system produces. Marketing copy, code, reports and designs are assigned to the customer in the terms of service, and most businesses stop reading there. Under South African law, that promise may be worth less than it appears. A contract can transfer rights that exist. It cannot create rights the law does not recognise, and it cannot override a statute that may place ownership elsewhere. For in-house counsel, the gap between the vendor’s promise and the Copyright Act is where the drafting work now lies.
CONTEXT AND BACKGROUND
South Africa’s Copyright Act already addresses computer-generated works. It treats as author the person who undertakes the arrangements necessary for the work’s creation. Webber Wentzel notes that this may point to the developer of the AI system rather than the person who typed the prompt, and that platform terms stating the user owns the output cannot override the statute. A second difficulty runs deeper. South African courts require a work to reflect sufficient skill, labour and judgement before copyright arises. If a prompt alone does not meet that test, there may be no copyright in the output at all, and as PPM Attorneys puts it, businesses would then have nothing to own, license or enforce.
Vendor contracts rarely fill the gap. An analysis of AI agreements using TermScout’s contract data found that 92% claim data rights beyond what the service requires, while only 33% indemnify customers against third-party intellectual property claims.
INSIGHT AND ANALYSIS
The consequence is not that businesses cannot use AI output. They plainly can. It is that output they assume is a company asset may be unprotectable, owned by someone else, or exposed to infringement claims. Those problems tend to surface at the worst moments: when a competitor copies a campaign, when an investor’s due diligence asks what intellectual property the business holds, or when a rights holder claims that an output reproduces its work.
The answer lies partly in the contract and partly in how the work is done. Better agreements grant the customer a perpetual, royalty-free licence in generated output while the provider keeps its rights in the underlying model. A licence of that kind, combined with an undertaking that the vendor will not assert rights against the customer, protects the business whether or not copyright exists. The rest depends on people. The more genuine skill and judgement staff contribute, through detailed briefs, careful selection and substantial editing, the stronger the case that a human author exists at all.
I have previously written about this. In an article on training data, I argued that businesses should require AI vendors to document what their models were trained on and the legal basis for acquiring it, commitments missing from most current agreements. Output ownership is the other end of the same supply chain.
IMPLICATIONS
In-house counsel should insist on four things. The first is a perpetual, irrevocable, royalty-free licence to use, modify and commercialise outputs, drafted to operate whether or not copyright subsists, with an undertaking that the vendor will claim no rights in them. Where the vendor offers an assignment instead, the clause should provide that it takes effect as such a licence if no assignable copyright exists. The second is an indemnity against third-party claims that outputs infringe existing works, since only a third of AI contracts currently offer one. The third is a commitment that the vendor will not reuse the customer’s prompts and outputs to train its models. The fourth is an internal practice that records human contribution to important work, because no contract can manufacture authorship. Counsel need not start from scratch. The European Union’s model clauses for procuring AI offer a template, and Cliffe Dekker Hofmeyr notes they can be adapted to South African law, including POPIA. Until the Copyright Act is updated for generative AI, ownership has to be managed through drafting.
CLOSING TAKEAWAY
A vendor’s promise that you own the output is sincere, but it is only as strong as the law behind it, and South African law has not settled the question. Businesses that treat AI output as secure intellectual property may discover otherwise in a dispute or a deal. The prudent course is to secure the right to use what AI produces, whoever the law ultimately says owns it, and to keep enough human judgement in the work that the question has a good answer.
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
Comments