GENERATIVE AI
“ The generative model never, ever sees real data,” he explained.“ However, the discriminator has seen some real data in the past... Eventually, over enough iterations, the generator, without seeing any real data, will actually start to create stuff that looks real.”
Privacy by design Synthetic data’ s privacy-preserving capabilities rest on mathematical foundations. Differential privacy adds controlled‘ noise’ to datasets, theoretically preventing synthetic data from being reverse-engineered to expose original information.
“ The benefit of synthetic data in that guise means that you can start to create data that’ s privacy-preserving, it means it can be shared,” Iain noted.“ And more importantly, there’ s ways to then collaborate across not just internal systems but external.”
But recent research reveals troubling gaps between theory and practice. An audit of a financial synthetic dataset – despite implementing differential privacy with supposedly conservative settings – uncovered actual customer names and sensitive financial details embedded in the synthetic data. The audit was completed overnight by two graduate students using only local
100 December 2025