Researchers' 2022 paper warned privacy issues like this were coming

thegautamkamath · x · 2026-09-09

Researcher Gautam Kamath says he has no information about the recent NS allegations, but notes privacy concerns like this tend to surface only after something bad may have happened—citing his 2022 arXiv position paper with Florian Tramèr and Nicholas Carlini.

The paper questions whether pretraining on large web-scraped datasets should count as privacy-preserving: publicizing such models as "private" could cause harm and erode trust in differential privacy; standard benchmarks poorly measure generalization to sensitive domains; and since pretraining mainly benefits models too large for users to run locally, deploying them may require outsourcing private data to third parties—a net loss for privacy.

Original post →

More from Safety

Safety channel →