CIA’s ‘AI coworker’ with limits
Analysis of CIA plans says the agency is building an AI ‘coworker’ but explicitly documents vendor and security limitations and sets boundaries on what AI will not do. The coverage framed the effort as cautious adoption with clear non‑goals rather than wholesale automation (nowadais.com).
The Central Intelligence Agency says it plans to build artificial intelligence “co-workers” into all of its analytic platforms within the next few years, while keeping humans in charge of final decisions. (politico.com) Deputy Director Michael Ellis said on April 9 that the tools will help analysts draft key judgments, edit for clarity, compare drafts against tradecraft standards, test conclusions and spot trends in incoming intelligence. He said the agency recently used artificial intelligence to generate an intelligence report for the first time. (govexec.com) The agency ran more than 300 artificial intelligence projects last year, Ellis said at a Washington event hosted by the Special Competitive Studies Project. He said officers could eventually manage teams of artificial intelligence agents in a hybrid model within a decade. (nextgov.com) A large language model is a text-prediction system that guesses the next word from patterns in huge datasets, like autocomplete scaled up to report-writing. In a 2023 Studies in Intelligence article, the Central Intelligence Agency said such systems “mimic” understanding rather than actually understanding data. (cia.gov) That same Central Intelligence Agency article said the core problem for analysts is the “black box”: users often cannot see which facts a model weighed or how it linked them. The article said that gap collides with intelligence work, where analysts are expected to show the basis for their judgments. (cia.gov) The agency has been describing those limits for more than a year. In a June 4, 2024 interview, Central Intelligence Agency Director of Artificial Intelligence Lakshmi Raman said classified work creates governance challenges and limits how the agency can use commercial products. (fedscoop.com) Ellis also signaled that vendor restrictions are part of the fight. Politico reported that he said the agency will “not let private companies dictate how and when the CIA will make lawful use of their technologies,” in remarks that pointed to tensions over model guardrails and national security use. (politico.com) The Central Intelligence Agency has also been moving to speed up how it buys technology. On February 9, 2026, Director John Ratcliffe announced a new acquisition framework with centralized vendor vetting and a streamlined information-technology authorization process to cut the time between defining a mission need and getting operating authority. (cia.gov) That puts the “co-worker” plan in a narrower lane than the label suggests: faster drafting, triage and pattern-spotting inside classified systems, not unsupervised analysis replacing officers. Ellis said “human beings are the ones making key decisions.” (politico.com)
