Former tech CEO who now manages AI for the CIA explains why it’s like your ‘crazy, drunk friend’
An advantage in knowledge can save lives, win wars, and prevent disasters. At the Central Intelligence Agency, basic artificial intelligence – machine learning and algorithms – has long served that mission. Now, creative AI is joining the effort.
CIA Director William Burns said AI technology will support humans, not replace them. The agency’s first chief technology officer, Nand Mulchandani, is organizing the tools. There is significant urgency: Adversaries have spread AI-generated deepfakes aimed at undermining US interests.
A former Silicon Valley CEO who led successful startups, Mulchandani was appointed to the position in 2022 after a stint at the Pentagon’s Joint Artificial Intelligence Center. Corner.
Among the projects he oversees: A ChatGPT-like synthetic AI application based on open source (i.e. unclassified, public, or commercially available) data. Thousands of analysts across America’s 18-agency intelligence community use it. Not surprisingly, other CIA projects using the broad language model were all secret.
The Associated Press interview with Mulchandani has been edited for length and clarity.
Q: You recently said that creative AI should be treated like a “drunk, crazy friend.” Can you build?
Answer: When AI systems create these “hallucinations,” they can sometimes behave like your drunk friend at the bar who might say something that pushes you beyond your conceptual boundaries your routine and inspire out-of-the-box thinking. Remember that these AI-based systems are probabilistic in nature and therefore imprecise (They tend to make things up). So, for creative tasks like art, poetry, and painting, these systems are great. But I wouldn’t use these systems for precise calculations or to design airplanes or skyscrapers – in those activities, “close enough” doesn’t work. They can also suffer from bias and narrow focus, which I call the “rabbit hole” problem.
Q: The only current application of large language modeling at an enterprise scale that I know of at the CIA is an open source AI, called Osiris, which is made available to the entire intelligence community. Right?
A: That’s the only thing we reveal publicly. That was really a home run for us. However, we should expand the discussion beyond the LLM—for example, we process large amounts of foreign language content across a variety of media, including video, and use algorithms and other AI tools to process that content.
Q: The Special Edge Research Project, a powerful think tank focused on AI in national security, has issued a report saying that US intelligence agencies must rapidly integrate computational AI. innovative—due to its disruptive potential. It sets a two-year timeline to move through trials and limited pilot projects and “deploy Gen AI tools at scale.” do you agree?
A: The CIA is 100% confident in using these technologies and scaling them up. We are taking this as seriously as we take any technology issue. We think we are well ahead of our original schedule because we are already using Gen AI tools in production. The deeper answer is that we are in the early stages of a large number of additional changes, and the bulk of the work is on integrating the technology more broadly into our applications and systems. These are the early days.
Q: Can you name your major language modeling partners?
A: I’m not sure if naming suppliers is interesting right now. There is currently a boom in the LLM market. As a smart customer, we don’t tie our boat to a particular LLM group or a particular set of providers. In fact, we are evaluating and using all the premium LLMs available, both commercial grade and open source. We do not view the LLM market as a single market in which a single lab is better than the others. As you have noted in the market, models are competing with each new release.
Q: What are the most important use cases at CIA for large language models?
A: That’s the summary. An open source analyst at the CIA cannot absorb the trove of media data and other information we collect every day. So this is a game changer for better understanding global sentiments and trends. Analysts then drill down into specifics. They must be able – with absolute certainty – to annotate and explain the data they cite and how they arrived at their conclusions. Our profession does not change. The supplements give analysts a much broader perspective – both the classified and open source pieces we collect.
Q: What are the biggest challenges in adapting general AI at work?
A: There is not much internal cultural resistance. Our employees deal with AI competitively every day. Clearly, the whole world is bursting with these new technologies and incredible productivity gains. The trick is to grapple with the limitations we face in terms of dividing information and how to build systems. In many cases, data separation is not for security reasons but for legal reasons. How do we effectively connect systems to take advantage of the benefits of AI while keeping all of that intact? Some really exciting technologies are emerging to help us think outside the box – and combine data in ways that maintain control over encryption and privacy.
Question: Generative AI is now as complex as an elementary school student. In contrast, intelligence work is for adults. It’s all about trying to penetrate your opponent’s deception. How does Gen AI fit into that job?
A: First, let’s emphasize that the human analyst has a superior role. We have the world’s leading experts in their fields. And in many cases where information arrives, a great deal of human judgment is needed to weigh its significance and significance – including the individuals who may provide it. We don’t have the machines to replicate any of that. And we’re not looking for computers to do the work of domain experts.
What we are looking at is the co-pilot model. We think Gen AI can have a huge impact in brainstorming and coming up with new ideas. And in increased productivity – and insight. We have to be very decisive about how we do that because, if used properly, these algorithms work well. But used incorrectly, they can really hurt you.
