Under South African law, legal professional privilege is unlikely to protect confidential client information that a lawyer inputs into a publicly available generative AI tool. The foundation of privilege is confidentiality, which is destroyed by disclosing information to a third party. Inputting data into a public AI system constitutes a disclosure to the AI provider, whose terms of service often permit them to use the data for model training, thereby destroying any reasonable expectation of confidentiality.

The output generated by an AI tool is also not automatically privileged. It cannot be protected by legal advice privilege because an AI system is not a qualified legal adviser. The argument for protection under litigation privilege (as work product) is weak, as the legal status of AI-generated works is uncertain, particularly since South African copyright law requires a human author. Privilege would likely only attach once a lawyer has applied their professional skill and judgment to verify and adopt the text, transforming it into their own work product.

The analysis is different for secure, private enterprise-grade AI tools that contractually guarantee data confidentiality. In this scenario, one could argue that privilege is not waived, but this position remains untested in South African courts. Practitioners are, in all circumstances, bound by their duties of confidentiality under the Legal Practice Act and data protection obligations under the Protection of Personal Information Act. Furthermore, South African courts have shown zero tolerance for AI-generated fictitious citations, establishing that practitioners are fully responsible for the accuracy of their work product and cannot blame technology for errors (Mavundla v. MEC: Department of Co-Operative Government and Traditional Affairs; Northbound Processing (Pty) Ltd v. South African Diamond and Precious Metals Regulator).