Limits of Synthetic-Data Attribution and Provenance
September 29, 2026
Research shows that while generator attribution is 98.7% accurate on original text, it drops to 53.1% after paraphrasing and 29.0% after style rewriting. The study demonstrates that identifying the source of synthetic data is not a reliable way to filter for quality or fitness in training sets.
HOW THIS AFFECTS YOU
●
researcherYou should be cautious about relying on provenance to manage the quality of synthetic training data.
●
policyYou can evaluate the difficulty of enforcing intellectual property rights when synthetic data is heavily modified.