The CIA is no longer treating artificial intelligence as a distant experiment. Under Lakshmi Raman, the agency's director of AI, the technology is being framed as a practical intelligence tool, with a strong emphasis on human judgment, legal compliance and responsible deployment.
That balance matters because the stakes are unusually high. AI can help analysts sort, search and interpret large volumes of information, but the same systems can also introduce bias, produce false claims or make sensitive uses of data harder to scrutinize.
From software development to CIA AI leadership
Raman joined the CIA in 2002 as a software developer after earning her bachelor's degree from the University of Illinois Urbana-Champaign and her master's degree in computer science from the University of Chicago. She later moved into management and eventually led the agency's overall enterprise data science efforts.
Her current role is to orchestrate, integrate and drive AI activities across the CIA. She describes the agency's approach as one that does not replace people with machines, but instead puts them together in the intelligence workflow.
"We think that AI is here to support our mission,"
Raman also said, "It's humans and machines together that are at the forefront of our use of AI." That line captures the agency's public argument for AI adoption: the technology is meant to help analysts work through information, not remove the need for human review.
Her career also sits inside a field that has historically been male-dominated. Raman said she benefited from women role models and predecessors at the CIA, and that she still has people she can ask for advice about leadership and career navigation.
What the CIA says AI can do for analysts
AI is not new inside the CIA. Raman said the agency has explored data science and AI since around 2000, especially in natural language processing, computer vision and video analytics. Those uses map closely to intelligence work: analyzing text, images and video at a scale that can be difficult for people to process manually.
The newer focus is generative AI. Raman said the CIA follows developments from industry and academia and uses that input to shape its roadmap. One area she highlighted is content triage, where generative AI can help users move through large amounts of material.
"When we think about the huge amounts of data that we have to consume within the agency, content triage is an area where generative AI can make a difference,"
She pointed to several specific uses: search and discovery aid, ideation aid, and generating counterarguments that may help counter analytic bias. In plain terms, the agency is looking at AI not only as a summarization engine, but also as a tool that can challenge assumptions and surface alternative lines of analysis.
The CIA is also using AI tools for translation and for alerting analysts during off hours to potentially important developments. Raman said the agency works with commercial services and wants to engage both well-known vendors and more specialized providers.
Osiris shows how intelligence AI is becoming operational
One of the clearest examples is Osiris, a generative AI-powered tool developed by the CIA. It is described as similar to OpenAI's ChatGPT, but customized for intelligence use cases.
Osiris summarizes data and lets analysts ask follow-up questions in plain English. For now, the source material is limited to unclassified and publicly or commercially available data.
The tool is already in broad use. Osiris is being used by thousands of analysts not only inside the CIA, but also across the 18 U.S. intelligence agencies. Raman did not say whether it was built in-house or with third-party technology, though she said the CIA has partnerships with name-brand vendors.
The pressure to move faster is also part of the picture. The source article points to geopolitical tensions, including threats of terror motivated by the war in Gaza and disinformation campaigns mounted by foreign actors such as China and Russia. It also notes that the Special Competitive Studies Project set a two-year timeline for domestic intelligence services to move past experimentation and limited pilot projects and adopt generative AI at scale.
The concerns are as important as the capability
The CIA's AI work is controversial because intelligence agencies already handle sensitive information. The source article notes that in February 2022, Senators Ron Wyden (D-OR) and Martin Heinrich (D-New Mexico) revealed in a public letter that the CIA has a secret, undisclosed data repository that includes information collected about U.S. citizens.
It also says that an Office of the Director of National Intelligence report showed that U.S. intelligence agencies, including the CIA, buy data on Americans from data brokers like LexisNexis and Sayari Analytics with little oversight.
If AI were applied to such data, the objections would be serious. The source article connects that concern to civil liberties and to the known limitations of AI systems.
Those limitations include bias and inaccurate outputs. The article cites studies showing that predictive crime algorithms from firms like Geolitica can be skewed by arrest rates and can disproportionately flag Black communities. It also says other studies suggest facial recognition produces a higher rate of misidentification for people of color than for white people.
Hallucination is another risk. Even strong AI systems can invent facts and figures. The source article gives the example of Microsoft's meeting summarization software occasionally attributing quotes to nonexistent people, then points out why that kind of failure is especially dangerous in intelligence work, where accuracy and verifiability matter.
Responsible AI depends on what users understand
Raman said the CIA complies with all U.S. law, "follows all ethical guidelines" and uses AI "in a way that mitigates bias." She described the agency's posture as careful rather than purely aggressive.
"I would call it a thoughtful approach [to AI],"
According to Raman, responsible AI at the CIA requires users to understand as much as possible about the systems they are using. She also said stakeholders need to be involved, including AI developers and the agency's privacy and civil liberties office.
That emphasis on user understanding is not abstract. The source article points to a North Carolina State University study finding that police were using AI tools, including facial recognition and gunshot detection algorithms, without being familiar with the technologies or their shortcomings. It also cites a reported NYPD example involving celebrity photos, distorted images and sketches used to generate facial recognition matches when surveillance stills produced no results.
Raman said AI-generated output should be clearly understood by users, including through labeling and explanations of how systems work. For the CIA, that means the future of AI is not just about which tools can be deployed. It is also about whether analysts, partners and oversight stakeholders know what those tools can and cannot be trusted to do.