arXiv:2610.27251 Maps AI Enclosure onto Medical Diagnostic Datasets

The 19 October 2026 arXiv submission applies historical enclosure analysis to information commons. Williams details how AI training pipelines extract value from open literature while flooding repositories with synthetic outputs that degrade search and citation integrity. In medical diagnostics this mechanism converts peer-reviewed imaging studies and outcome reports into closed model weights controlled by single firms.

English enclosure records from 1600-1800 demonstrate conversion of common land into rent-generating property without physical destruction. Equivalent patterns appear in diagnostic AI: models trained on public PubMed Central and radiology archives now gatekeep performance benchmarks. Bias audits published in Nature Medicine 2019 recorded 18-34 percent drops in accuracy for underrepresented populations when training corpora shifted from open to proprietary subsets.

Corporate control concentrates both data and evaluation. Diagnostic firms withhold training distributions while releasing only aggregate metrics, blocking independent verification of fairness claims. The paper identifies slop spam as an active erosion vector: generated review articles dilute signal in literature searches that clinicians and regulators use to validate new tools.

Operational consequence is accelerated divergence between public scientific record and private model capability. Regulators requiring post-market surveillance lack access to the original training distributions once enclosure completes.