CIA: AI Produced First Fully Machine-Generated Report

Deputy Director Michael Ellis confirmed Thursday at a Special Competitive Studies Project event that an AI system produced the CIA’s first fully AI-generated intelligence report. He outlined a roadmap to embed AI “coworkers” in analyst workflows and predicted officers could oversee teams of AI agents within about ten years.

Ellis noted the agency ran more than 300 AI projects last year and that one produced an intelligence product created entirely by a machine. In the near term, the CIA plans to embed AI tools in its analytics platforms to help draft text, edit for clarity and check outputs against tradecraft standards, with human analysts retaining final sign-off.

The agency described the tools as “coworkers” intended to speed production and allow faster publication than a human-only process permits.

Looking ahead, Ellis used the term “autonomous mission partners” for teams of autonomous AI agents that officers would supervise. He estimated that model could be in place within roughly ten years and would expand collection and analysis capacity beyond what human teams can do alone.

Ellis addressed concerns about vendor dependence, noting the CIA “cannot allow the whims of a single company” to constrain its use of AI and that the agency is diversifying suppliers to remain operationally flexible. His remarks came after government actions that limited some commercial tools for sensitive applications.

The agency has increased its technology focus. Ellis said the CIA doubled its technology-focused foreign intelligence reporting to track how adversaries use AI across semiconductors, cloud computing and research and development. The Center for Cyber Intelligence was elevated to a full mission center and will be tied closely to AI development and defense.

Generative tools are already being used at the agency. An internal AI chatbot launched in 2023 to help staff parse surveillance data, and by 2024 senior intelligence leaders had publicly acknowledged using generative AI for content triage and analyst support. Ellis framed the approach as balancing automation with human judgment and added that humans will continue to sign off on results.

Agency officials declined to provide technical details about the system that produced the report, including which models or safeguards were used, citing operational security. They emphasized that tradecraft standards and human oversight will remain part of the process as automated tools are scaled.