Synthetic data is moving from “privacy workaround” to a governed data product. A new ERC-funded research project, a WEF policy brief, and fresh medical evidence on privacy-utility trade-offs all point to the same need: measurable risk controls, not assumptions.
New project to investigate societal consequences of using synthetic data to train algorithms
The University of York announced the launch of SYNDATA, a European Research Council-funded project led by Dr. Benjamin Jacobsen. The project will examine the practical, ethical, and political consequences of using synthetic data to train algorithms across sectors including healthcare and finance.
SYNDATA plans to use archival research, fieldwork, and case studies to understand how synthetic data influences society and power structures—particularly as generative AI blurs boundaries between “real” and “synthetic” data in deployed systems.
- Governance scope is widening: beyond privacy, synthetic data programs may be evaluated for downstream societal and political effects (e.g., who benefits, who is excluded).
- Procurement and assurance pressure: buyers in regulated sectors should expect more scrutiny of how synthetic training data is created, validated, and documented, not just whether it is “de-identified.”
- Regulatory signal: research like this can inform emerging standards for data ethics and algorithmic fairness, especially where synthetic data is used to fill gaps in real-world representation.
Synthetic Data: The New Data Frontier
The World Economic Forum published a briefing paper positioning synthetic data as a scalable response to data scarcity, privacy constraints, and AI training needs—particularly in sensitive domains such as healthcare and finance. The paper highlights use cases including testing, personalized AI, and red-teaming, while emphasizing that governance is required to manage accuracy, equity, and privacy.
The document frames synthetic data as an enabler when real data is limited or risky to share, but it also stresses multi-stakeholder collaboration to establish responsible practices and guardrails.
- Expect “synthetic-by-default” proposals: the WEF framing will reinforce vendor and internal platform roadmaps that treat synthetic data as core infrastructure for model development and evaluation.
- Governance becomes a product requirement: accuracy, equity, and privacy checks need to be operational (KPIs, test suites, sign-off), not just policy language.
- Red-teaming expands: teams can use synthetic data to probe model failure modes and security issues—if they can show the synthetic generator doesn’t introduce misleading artifacts.
Impact of synthetic data generation for high-dimensional cross-sectional medical data: privacy versus utility considerations
A JAMIA study evaluated three strategies for synthetic data generation (SDG) on high-dimensional, cross-sectional medical datasets, comparing privacy risk (membership disclosure) against data utility. The paper contrasts approaches that generate synthetic data from the full dataset versus subset-based strategies, and reports evidence-based recommendations for balancing privacy and utility in data-sharing platforms.
By quantifying trade-offs rather than assuming synthetic outputs are automatically “safe,” the study contributes practical guidance for privacy-preserving machine learning in healthcare contexts.
- “Synthetic” is not a privacy guarantee: membership disclosure risk needs explicit measurement, especially for high-dimensional medical data where re-identification concerns are acute.
- Platform design implications: full-dataset vs. subset SDG choices affect both utility and privacy; data-sharing services should treat SDG configuration as a controlled, auditable parameter.
- Compliance alignment: evidence on privacy-utility trade-offs supports more defensible decisions under regimes like GDPR and HIPAA—useful for DPIAs, risk registers, and data release approvals.