A knowledge advantage can save lives, win wars and prevent disasters. At the Central Intelligence Agency, fundamental artificial intelligence – machine learning and algorithms – has long served this mission. Now generative AI is joining the efforts.
CIA Director William Burns says AI technology will augment humans, not replace them. The agency’s first Chief Technology Officer, Nand Mulchandani, is assembling the tools. There is great urgency: adversaries are already spreading AI-generated deepfakes aimed at undermining American interests.
Mulchandani, a former Silicon Valley CEO who led successful startups, was appointed in 2022 after a stint at the Pentagon’s Joint Artificial Intelligence Center.
Among the projects he oversees: a ChatGPT-like generative AI application that uses open-source data (i.e. unclassified, publicly or commercially available). Thousands of analysts from the 18 American intelligence services use it. Other CIA projects using large language models are, unsurprisingly, classified.
This Associated Press interview with Mulchandani has been edited for length and clarity.
Q: You recently said that generative AI should be treated like a “crazy, drunk friend.” Can you explain that?
A: When these generative AI systems “hallucinate,” they can sometimes act like your drunk friend at a bar who may say something that pushes you beyond your normal conceptual boundaries and sparks out-of-the-box thinking. Please note that these AI-based systems are probabilistic in nature and thus not accurate (they are prone to fabrication). So for creative tasks like art, poetry and painting, these systems are excellent. But I wouldn’t use these systems for precise math or designing an airplane or skyscraper just yet – “close enough” doesn’t work for those activities. They can also be biased and narrowly focused, which I call the “rabbit hole” problem.
Q: The only current use of an enterprise-level large language model that I know of at the CIA is the open-source AI called Osiris that it created for the entire intelligence community. Is that correct?
A: That’s the only one we’ve made public. It was an absolute home game for us. However, we need to broaden the discussion beyond just LLMs: for example, we process massive amounts of foreign language content across multiple media types, including video, and use other AI algorithms and tools to process that.
Q: The Special Competitive Studies Project, a powerful advisory group focused on AI in national security, is releasing a report saying that US intelligence agencies must quickly integrate generative AI – given its disruptive potential. It sets a two-year timeline to move beyond experiments and limited pilot projects and “deploy Gen AI tools at scale.” Do you agree?
A: The CIA is 100% committed to deploying and scaling these technologies. We take this as seriously as we probably do any technology issue. We think we’ve beaten the original timeline by a large margin because we’re already using Gen AI tools in production. The deeper answer is that we are just at the beginning of a host of additional changes, and much of the work is integrating the technology more broadly into our applications and systems. These are early days.
Q: Can you name your major model partners?
A: I’m not sure if naming the suppliers is interesting at this point. There is an explosion of LLMs on the market right now. As a smart customer, we do not tie our boat to a specific group of LLMs or a specific group of suppliers. We evaluate and use virtually all the high-quality LLMs out there, both commercial and open source. We do not view the LLM market as a unique market, where any single laboratory is better than the others. As you notice in the market, models are overlapping with each new release.
Q: What are the top use cases at CIA for major language models?
A: Primary is summary. It is impossible for an open source analyst at the CIA to handle the firehouse of media and other information we collect every day. So this has been a game-changer for insights into sentiment and global trends. Analysts then delve into the details. They must be able to annotate and explain – with complete certainty – the data they cite and the way they reach conclusions. Our profession has not changed. The additional pieces provide analysts with a much broader perspective – both the classified pieces and the open source pieces we collect.
Q: What are the biggest challenges in adapting generative AI at the agency?
A: There isn’t much cultural resistance internally. Our employees have to compete with AI on a daily basis. Clearly the whole world is on fire with these new technologies and the astonishing productivity gains. The trick is grappling with the limitations we have in terms of compartmentalizing information and the way systems are built. In many cases, data separation is not for security reasons but for legal reasons. How can we efficiently connect systems together to leverage the benefits of AI and keep it all intact? There are some really interesting technologies emerging to help us think about this – and combine data in ways that preserve encryption and privacy controls.
Q: Generative AI is currently about as advanced as an elementary school student. Intelligence work, on the other hand, is for adults. It’s about trying to break through an opponent’s deception. How does Gen AI fit into that work?
A: Let’s first emphasize that the human analyst takes precedence. We have the world’s leading experts in their fields. And in many cases of incoming information, a tremendous amount of human judgment is involved in weighing its importance and significance – including that of the individuals who may be providing it. We don’t have machines that mimic all that. And we’re not looking for computers to do the work of domain experts.
What we’re looking at is the co-pilot model. We think Gen AI can have a huge impact on brainstorming and coming up with new ideas. And in increasing productivity – and insight. We have to be very deterministic about how we do it because these algorithms, when applied properly, have a positive impact. But if you handle them incorrectly, they can really hurt you.