After reviewing fifty-seven real-world use cases for Synthetic Data Systems, we developed a clear and practical way to describe how these systems work and how their quality can be evaluated. Too often, synthetic data is discussed in vague terms, relying on broad claims about realism or accuracy. Our taxonomy organizes Synthetic Data Systems into clear layers based on the types of outputs they produce, and the checks used to ensure those outputs are reliable. This structure helps providers, customers, and the wider industry assess synthetic data in consistent, comparable ways that reflect how it is actually used to support decisions.
The Modeled Answers and Insights layer delivers summarized conclusions directly to stakeholders, without requiring access to individual respondent data. The quality of these outputs depends on how well the underlying data is balanced and complete, and on whether the study design assumptions are realistic. When these controls are applied correctly, modeled insights remain credible and transparent, and clearly communicate uncertainty rather than masking it behind precise-looking numbers.
The Optimized Survey and Instrument layer focuses on improving how studies are designed before they go into the field. It produces guidance on feasibility, structure, and configuration to help ensure surveys can be executed as planned. Quality controls at this layer ensure that recommendations reflect real respondent behavior and operational limits, reducing the risk of designs that look good in theory but fail in practice.
The Decision Intelligence Layer brings together analysis-ready synthetic data with modeled trends and projections. This is the main layer used for analysis, reporting, and decision support. Quality controls here focus on making sure the data fairly represents the intended population and is not distorted by gaps or imbalances. By addressing these issues early, this layer helps ensure that downstream analyses and decisions are based on a solid and trustworthy foundation.
