AI and Employment: Legal Questions on Training Your Replacement

Information security and privacy consultant Anthony Olivier. Consider Nomsa. Over 20 years and three employers, she has become an exceptional compliance specialist. She knows which regulator to telephone, which ambiguity to leave alone, and which minor exception sinks an audit. Her employer asks her to spend 18 months teaching an internal AI agent how she works. She corrects it, explains her reasoning, feeds it the exceptions. The agent eventually handles most of her file. The company retrenches her, and keeps the agent. I am not asking who owns the AI. The employer owns the servers, the licences and the model. The harder question is this: has an employer acquired a permanent, reusable copy of an employee's personal professional capability merely because that employee exercised it at work? South African law gives Nomsa a stronger position than most executives expect. In the case Automotive Tooling Systems v Wilkens, the Supreme Court of Appeal held that a person's skills and abilities are a part of himself, and that the know-how the employer sought to protect was nothing other than skills, which do not constitute a proprietary interest vesting in the employer. Money spent on training does not buy the competence. This is not a museum piece. The Labour Court applied the principle in Nuvest Chemicals v Meyer in January 2026, holding that training employees affords an employer no proprietary interest in the worker, his know-how or his skills. Copyright is the wrong instrument here, and any board relying on it is relying on the wrong thing. The High Court reached comparable conclusions in Tapati Investments v Malatji in 2025 and Equity Medical Technologies v De Villiers in 2026. The doctrine was built, however, for a world in which knowledge left the building inside the employee's head. Distillation, meaning the deliberate extraction of a person's working method into a model, breaks that assumption. The capability stays. The person leaves. Not yet in South Africa, as far as I can establish. Elsewhere the outlines are appearing. In 2026, the Hangzhou Intermediate People's Court found the dismissal of a quality assurance supervisor unlawful after his employer cited AI disruption and offered him a lesser post at a 40% pay cut. A Beijing arbitration panel had already found in 2025 that an employer replacing a worker with AI had shifted the cost of its own technological transformation onto that worker. In Lehrman v Lovo, decided in New York in July 2025, two voice actors sued after recordings supplied for academic research were turned into commercial synthetic voices. The instructive part is what failed. The copyright claims fell away, because copyright protects the fixed recording rather than the abstract qualities of a voice. What survived were the contract claim and the personality rights claim. Copyright is the wrong instrument here, and any board relying on it is relying on the wrong thing. Apple's trade secret action against OpenAI, filed in July 2026, raises the adjacent problem: once information is absorbed into a model, an order to hand back the documents no longer restores the position. In July 2026, Yang Chen published the first systematic legal account of what he calls employee distillation and cloning: feeding a worker's mail, chats, documents, meeting transcripts and code reviews into systems that reproduce not merely outputs, but working style, tacit judgement and professional persona. His conclusion is uncomfortable. Privacy law, copyright, trade secret doctrine, personality rights and contract each appear to offer protection, and each dissolves on inspection, producing what he calls a rights mirage. Chen argues not for prohibition, but for notice, genuine consent or bargaining, limits on use, compensation and collective governance. Valerio De Stefano of Osgoode Hall calls the uncompensated capture of workers' digital activity a quiet transfer of value. In a survey of data workers by the Alphabet Workers Union with the C