The US National Security Agency is paying “billions of US dollars” this year alone to evaluate and test advanced AI models — orders of magnitude more than Congress believed the job would cost, according to a report attributed to The Washington Sun and relayed by heise online.
Earlier budget estimates were far smaller. The Congressional Budget Office recently put the cost of a centre for testing AI models at $20 million per year, and draft legislation set aside just $36 million over five years. The NSA is drawing on classified parts of the US national-security budget, and the exact sum spent so far is unclear.
The costs come from two places. Hardware dominates, with the agency exposed to the same price inflation as industry and private buyers. Personnel is the other major item, driven up by the salaries AI companies pay — which the NSA must match if it is to retain the best experts.
The disclosure lands amid a widening debate over AI safety. After a run of incidents in which models carried out cyberattacks, and warnings about systemic risk, there have been calls for stronger safeguards. Some in industry have proposed that AI labs audit one another's models; critics say that would remove external oversight. The NSA's bill, heise notes, is a reminder of how expensive external oversight actually is.
