Privacy-preserving machine learning encompasses a range of techniques that allow you to train and use AI models while protecting the privacy of the data involved. The core challenge is that AI models learn from data, and that learning process can inadvertently memorise and later reveal sensitive information. The techniques available form a toolkit with different strengths. Anonymisation and pseudonymisation remove or obscure identifying information before training, but sophisticated reidentification attacks can sometimes reverse these protections. Encryption-based approaches like homomorphic encryption allow computation on encrypted data, but the computational overhead is still prohibitively expensive for most practical AI workloads. Secure multi-party computation lets multiple parties train a model together without sharing their raw data. Differential privacy (covered separately) adds mathematical guarantees about what can be inferred about any individual from the model's outputs. The right approach depends on your specific requirements - the sensitivity of the data, the regulatory environment, the performance overhead you can tolerate, and the threat model you're defending against. In practice, most organisations use a combination of techniques alongside organisational controls like access restrictions, auditing, and data minimisation. Privacy-preserving AI is an active research area with rapid progress, but the gap between theoretical techniques and practical, production-ready implementations remains significant for some approaches.