Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Foundation Models And Ic topic
No spam. Unsubscribe anytime.
Panelists at PCLOB forum debate foundation models for intelligence use; agree on retraining and policy safeguards
Summary
Experts at the Privacy and Civil Liberties Oversight Board forum generally endorsed conditioned use of commercially trained foundation models by the intelligence community—citing cost constraints—but urged retraining, explicit policies on U.S.‑person data, and rigorous predeployment testing and human review.
A central debate at the board’s public forum focused on foundation (large) models and whether intelligence agencies should use commercially trained systems or develop their own. Board members and panelists agreed the question involves tradeoffs among capability, cost, privacy and security.
William Usher and other panelists argued that reproducing leading foundation models in‑house is likely to be prohibitively expensive and that the intelligence community will therefore rely on commercially trained models in many cases. That dependence, they said, requires strict standards for retraining, accuracy, and protections for U.S.‑person information.
Dean celis (former chief technology adviser at ODNI) framed the challenge bluntly: "these Foundation models are trained on huge volumes of data and in the intelligence community that means they're trained on data that involves us persons," and he added that agencies "should retrain them to our own standards." Panelists noted that retraining, auditing and clear policies about human review are ways to balance utility and civil‑liberties protections.
Panelists also highlighted distinct technical and operational risks: models can hallucinate, leak training‑data artifacts if exposed, and compound secrecy concerns when black‑box commercial software is combined with classified tradecraft. Speakers recommended context‑specific policies (for experimentation versus operational use), predeployment evaluations, and a requirement that high‑risk decisions be subject to human verification and civil‑liberties review.
No formal decisions were made at the forum. Panelists urged the board to press agencies to document assumptions, run measurable tests, and incorporate continuous privacy impact assessments as agencies implement the White House executive order and await the forthcoming national security memorandum on AI.
