At first glance, the Constitutional Court’s latest copyright judgment has little to do with artificial intelligence.

The Court was not asked to rule on AI, nor did it decide the legality of AI training. Yet the judgment’s implications for AI are difficult to ignore. By confirming that copyright is property protected by the Constitution and that fair use is constitutionally valid, the Court has established principles that are likely to shape future disputes involving AI and other emerging technologies.

Most modern technologies, from legal research platforms to image generators, are built on copyrighted works such as books, articles, images, lectures, and software code. This article explores what the Court’s reasoning could mean for those who develop, train, and deploy these technologies.

On 26 June 2026, the Constitutional Court decided the constitutionality of the Copyright Amendment Bill and the Performers’ Protection Amendment Bill. Three findings sit at its centre, and together they reshape the ground on which anyone who builds technology from existing works now stands.

What the Court decided

Copyright is constitutional property. For the first time, the Court held that copyright is property protected against arbitrary deprivation under section 25(1). The works that feed technology development, including textbooks, novels, journal articles, photographs, illustrations, recorded lectures, music, and source code, are almost all copyright works, and their protection now carries express constitutional weight.

Fair use is valid and built to flex. Section 12A replaces the old closed “fair dealing” list with an open, United States-style “fair use” test, and the Court upheld it. Fair use limits an owner’s exclusivity and is therefore a deprivation of property, but not an arbitrary one, because a four-factor fairness test structures it: the nature of the work, the amount used, the purpose and character of the use, and the substitution effect of the use on the potential market for the work. The Court described this open approach as better able to evolve alongside creative and cultural developments than the closed list it replaces.

Market substitution is the line, and “useful” is not enough. The Court struck down section 12D(1)-(5), which would have allowed broad reproduction of works for education where a licence was unavailable or unreasonable, because it undermined the market for those works and shifted the burden onto authors and publishers. A worthwhile public purpose does not save a limitation that is disproportionate in its effect on the owner’s market. The factor the Court kept returning to is the fourth one: whether a use “eats into” the commercial market for the original work.

What it means for building AI and other technology

This is not an AI judgment. The case concerned issues such as education, disability access, quotation, and parody, and the Court did not consider artificial intelligence, training data, or model outputs. Nevertheless, the principles it established are highly relevant to technology development. While the Court did not rule on the legality of AI training or deployment, it confirmed the constitutional framework within which those questions are likely to be assessed in future. As a result, the judgment provides important guidance for developers, rights holders, and businesses operating at the intersection of technology and copyright. It allows a number of broader conclusions to be drawn with some confidence.

Your inputs are protected property. Building a product from scratch still means building it on something. A model learns from a corpus, and that corpus is usually made of copyright works: books, articles, images, lectures, or code. After this judgment, those inputs are not just protected by statute; they are constitutionally protected property. Using them without permission is a deprivation of the owner’s rights. The question is no longer whether a deprivation has occurred, but whether it is fair.

Fair use is the gateway and it can reach new technology. This is the helpful part for developers. Under the old closed list, a use had to fit a fixed category or it infringed.

Fair use asks a different question: is the use fair, judged on the four factors? Because the list of purposes is open (the section says “such as”), and the Court read that openness as designed to cover uses not yet contemplated when the law was written, there is room for novel, technological uses to qualify.

For a South African developer, that is a materially better position than a rigid regime that never imagined machine learning. But the openness has a limit the Court was careful to draw: an unlisted purpose must share the character of the listed ones, that is, public benefit or private and personal use. A model trained for research or scholarship maps onto that more easily than a commercial product deployed to the market. How you characterise the purpose of a use matters.

Run each use through the four factors, the way a court will.

• Nature of the work. Highly creative works (novels, photographs, artwork, music) attract stronger protection than factual or functional material. A tool built on primary legal materials or technical data sits differently from one built on copyrighted textbooks, commentary or stock images.
• Amount used. Using a limited part weighs towards fairness; reproducing whole works weighs against it. This is the awkward factor for AI, because training typically ingests entire works, not extracts.
• Purpose and character. The section asks whether the use serves a different purpose from the original and whether it is commercial. The Court’s picture of fair, transformative use is specific: using a limited part of an original work to create a new, original work. That is narrower than the broad “transformative purpose” idea used in some other jurisdictions, and whole-corpus training does not fit it neatly. Commercial character does not defeat fairness on its own, but it weighs in the balance.
• Substitution effect on the market. The decisive factor. If your product competes with, or displaces demand for, the works it was built on (a research tool that substitutes for the publisher’s own database, an image generator that replaces licensing the source images) the use is at the hardest end of the test. If it produces something those works do not sell, its position is stronger.

“Innovation” and “access” are not, on their own, a defence. The clearest signal from the section 12D strike-down is that the Court will not let a socially valuable goal justify wholesale copying that guts a market. Arguments that a product democratises access, advances a profession, or drives innovation carry weight in the balance, but they do not override a disproportionate effect on the owner’s market. A developer who plans to lean on public-benefit framing should assume it will not, by itself, carry a use that substitutes for the source works.

What the judgment does not resolve. Several development questions remain open. The judgment does not decide whether ingesting whole works to train a model is fair; the amount and substitution factors pull in different directions, and no South African court has yet applied them to training. It does not address who owns the output of an AI system, which turns on separate authorship rules. And it does not speak to outputs that reproduce protected expression from the training data. These will be litigated inside the framework the Court has now set, but they are not answered by it.

Where the Bill stands

The Bill is not yet law. Parliament passed it, the President referred it in October 2024, and because section 12D(1)-(5) fell, it cannot be signed as it stands and must return to Parliament for amendment. The rest of the framework, including fair use and the market-substitution test, has been endorsed, and the Court held that its constitutional findings bind the courts now, before enactment. Treat the fair-use architecture and the market-harm test as settled, and section 12D(1)-(5) as the part that will change.

What to do now

• Build a copyright position into product design, not after launch. Identify what your model learns from, and whether those inputs are copyright works, before you train.
• Document the purpose and the market effect of each use. Those are the two factors that decide close cases, and a contemporaneous record is worth far more than a reconstruction later.
• Prefer inputs you can license, own, or that carry weaker protection. Properly licensed corpora, primary legal materials and factual data are safer ground than scraped creative works.
• Do not rely on “transformative” as a slogan. The Court’s version of transformative use is narrow and assumes limited copying, which sits awkwardly with whole-work training.
• Do not assume public benefit rescues a product that competes with its sources. Where a use would displace the market for the works behind it, licensing may be the only lawful route, not merely the safer one.
• If you hold rights in works being used to train others’ systems, the property characterisation and the market-substitution factor are now your strongest arguments.

This article is general information and reflects the law at the time of writing. It is not legal advice. For advice on a specific matter, please speak to us.