Synthetic data is increasingly being treated as a primary input for model development as access to high-quality, compliant real-world data tightens. The practical shift: teams can scale training sets faster while reducing exposure to sensitive records—if they can prove utility and control leakage.
Synthetic data adoption accelerates as real-world data gets harder to use
Industries are ramping up synthetic data generation for AI training, driven by a mix of data scarcity and regulatory pressure. The piece cites estimates that synthetic data could represent 80% of AI training data by 2028, up from about 5% five years ago—framing the change as a response to diminishing supplies of high-quality, ethically sourced, and compliant real data.
Examples highlighted include J.P. Morgan using synthetic datasets for fraud detection, Waymo using simulation to test autonomous-driving scenarios at massive scale, and healthcare organizations generating synthetic patient records to train diagnostic AI while staying aligned with HIPAA requirements. The article also points to synthetic data’s role in privacy and governance strategies as regulations evolve (including the EU AI Act), positioning synthetic data as a way to build and test models without directly exposing sensitive customer or patient information.
- Data leads: Synthetic data can reduce dependency on hard-to-access production data and shorten iteration cycles, but you’ll need clear acceptance criteria (task performance, coverage of edge cases) before it becomes “default” training input.
- ML engineers: Simulation and synthetic augmentation can improve robustness in rare-event regimes (fraud, safety incidents), yet model behavior can drift if the generator encodes the wrong priors—monitor for distribution mismatch.
- Privacy & compliance: Synthetic data can lower breach exposure and support privacy-by-design workflows, but it doesn’t eliminate risk; teams still need to validate leakage and re-identification resistance for the chosen method.
- Procurement & governance: Expect more “synthetic-first” vendor claims; require documentation on generation approach, evaluation methodology, and controls for memorization before approving use in regulated pipelines.
Estimate cited: synthetic data could make up 80% of AI training data by 2028 (up from ~5% five years ago).