Anonymity in the Age of AI
Parisasadat Shojaei, Nabi Zameni, Rezza Moieni · Open Journal of Social Sciences · 2025
Artificial intelligence (AI) is eroding traditional de-identification practices by enabling accurate re-identification of images, text and behavioural traces. A systematic review of 64 peer-reviewed studies published between 2013 and 2025—47 on technical privacy-enhancing technologies (PETs) and 17 on the EU General Data Protection Regulation (GDPR)—shows that no single safeguard withstands modern adversaries. The most resilient configurations layer differential privacy, federated learning and partial homomorphic encryption, maintaining < 2% accuracy loss on medical benchmarks while blocking current model-inversion attacks, though at notable computational cost. The legal literature reveals a coverage gap: GDPR protections are strong during data collection and preprocessing but weaken during training, inference and post-deployment reuse, when AI-specific risks peak. Article 22 offers only partial defence against model-inversion and prompt-leakage and learned embeddings or synthetic corpora often fall outside the regulation’s definition of personal data. Effective anonymity in the AI era, therefore, requires end-to-end PET adoption and regulatory updates that specifically address behavioural telemetry, embeddings and synthetic datasets.