Paper taxonomizes synthetic data for privacy and competition law
Original title: Taxonomizing Synthetic Data for Law
The authors say no synthetic-data category stops group-inference privacy harms, and synthetic data rarely offsets incumbents' data advantages.
A cs.CR paper extends Michal Gal and Orla Lynskey's account of synthetic data's effects on privacy, competition, and data quality. It divides synthetic data into transformed data that preserves statistical properties of collected data, augmented data that uses assumptions to improve fidelity, and simulated data built mainly from background knowledge. The authors argue that ground-truth assumptions matter as much as the link to collected data. They conclude that none of the categories prevents privacy harms from group inferences, that augmented and simulated data can amplify bias, and that synthetic data rarely helps smaller competitors overcome incumbents.
Unconfirmed: The privacy, bias, and competition conclusions are the authors' arguments, not empirical measurements.
- Transformed data modifies collected data for a stated end use.
- Augmented data relies on assumptions about ground truth.
- Simulated data depends almost entirely on background knowledge.
- Authors urge policymakers to check those ground-truth assumptions.
From the original article
1 min read at arxiv.orgSynthetic data is increasingly important in data usage and AI design, creating novel legal and policy dilemmas. All too often, discussions of synthetic data treat it as entirely distinct from "real," collected data, overlooking the risks posed by different kinds and uses of synthetic data.
By Ignacio Cofone · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.