Research
Synthetic health data, taken seriously.
Research on privacy-preserving synthesis, differential privacy and health-data regulation, written for the teams who have to get it right. Including where synthetic data fails.
§ 01Publications
Published briefs, primers and working notes.
Policy & regulatory analysis
AI ActThe AI Act's synthetic-data clause has never been tested Article 4a lets providers use special-category data for bias detection only when synthetic or anonymised data cannot do the job. What a provider would have to show, and why the evidence on subgroups makes it hard.Oct 2026→ GDPRIs synthetic health data anonymous under GDPR? Not automatically. A reading of the EDPB’s July 2026 draft anonymisation guidelines and what they ask of synthetic data.Oct 2026→ EHDSWhat makes a synthetic release “adequate” under the EHDS? The EHDS lets an access body release non-personal outputs from a secure processing environment but never defines adequacy. Five checkable criteria an access body could apply.Sep 2026→ EvaluationFit for which purpose? How to judge whether synthetic health data is good enough Which utility measures match which use, why similarity-based privacy scores fail, and what a fitness-for-purpose label must carry.Sep 2026→
System & benchmark notes
TN-02Formal privacy costs the smallest groups most Measured on our own public release across 24 subgroups, with a random sample as control. The release reproduces the population mix inside every subgroup, and what a certificate must therefore say.Nov 2026→ Claims dataSynthetic Medicare claims data: from SynPUF to certified synthesis Why CMS built DE-SynPUF, where it falls short today, and what certified differentially-private synthesis adds.Sep 2026→
Foundational primers
PrimerDifferential privacy in healthcare, explained for non-cryptographers What ε and δ actually mean, why post-hoc noise does not count, and how to read a privacy certificate.Sep 2026→ PrimerSynthetic data vs. de-identification: why the difference is mathematical De-identification removes identifiers and hopes. Differentially-private synthesis generates new records under a provable bound.Sep 2026→ PrimerWhat is synthetic health data? Definition, how it is generated, when it can be trusted, and an honest table of what it is and is not suitable for.Aug 2026→
§ 02Open questions
What we are working to answer.
The questions that define the field. If you are working on one of them, we would like to compare notes.
Q1 / Evaluation
How should utility be measured for a stated purpose?
A feasibility study and a model benchmark need different fidelity. Which measures predict success for which use?
Q2 / Privacy
What does an attack-tested privacy claim look like?
Formal guarantees and empirical attacks answer different questions. A certificate should carry both.
Q3 / Regulation
When is a synthetic release adequate?
Access bodies under the EHDS will need criteria they can apply consistently. What should they check?
Q4 / Representation
How do we build cohorts where data is scarcest?
Populations under-represented in global datasets are often where privacy constraints bite hardest.