William Marcellino is an expert at getting the most out of AI models. He oversees a growing portfolio of specially built tools that RAND is using to innovate, accelerate, and elevate its research.
He came to RAND as a linguist and was training computers to find patterns in words long before anyone had heard of ChatGPT or Claude. He's used them to identify Russian trolls operating in the shadows of social media, and to better understand what makes ISIS supporters tick. His latest research has analyzed how AI could develop in the coming years; what an artificial superintelligence would mean for global conflict and competition; and how China could use AI to flood social media with fake accounts.
“I love this stuff,” he said. “I use it, I'm fascinated by it, I do research on it. These AI models are very powerful, and they're very, very useful.”
Let's start with a prediction: 2026 is going to be the year when…?
…we have a massive cybersecurity incident.
Businesses are starting to really use AI agents, and a lot of those agents use something called MCP, or model context protocol. It's a way for these agents to talk to each other. They assume, “If you talk to me from a higher control level, I can trust you.” So if I can convince an MCP server to let me in one time, I can go all the way down through the system, and no one will stop me. Cyberattacks are nothing new, but now we've automated vulnerability.
But I also don't want to be too much of a pessimist. I think this year we're going to start to see really big productivity gains as well. People who incorporate AI into their workflows are going to have a major advantage.
What will those productivity gains mean for human employment?
That's the billion-dollar question. Right now, the same model that can win gold in a math competition will also say that you can exist in two places at once if you just join a video call. That ability to be both brilliant and stupid tells me we're going to need humans for a long time.
If you look back at the software revolution of the 1980s, knowledge workers became much more productive, and their median wages skyrocketed. I think we could see something similar here, just a massive increase in the value of information work. It's going to explode.
[AI's] ability to be both brilliant and stupid tells me we’re going to need humans for a long time.
You direct the development of AI tools at RAND. What are some really cool examples of how RAND is using AI?
We're working on tools that can help researchers design more effective wargames, or that can answer questions by searching through thousands of RAND research reports. We've developed a model that can collect and scan the existing research on a topic, so researchers know what's already out there. These models are allowing us to do much faster, higher-quality work.
One tool that's had an immediate impact analyzes interview transcripts. Let's say researchers go out and interview 100 or 200 people about a health intervention. The tool can go through all of those transcripts, pull apart the responses, and match them up. It'll come back and say, “Hey, I found some high-level patterns. Here's what really works well. Here are the problems. Here's why things work or don't work.” All of those people on the ground have little pieces of the puzzle. This allows us to put all those pieces together and get a really rich analysis of what's happening.
You've described current large language models as brittle. What does that mean—and what does it mean for the future of AI?
These models are built to do one thing, find patterns. They can identify a beak, a wing tip, some tail feathers, and know they've got a bird. But they have no memory. They can't handle symbolic thought. They can see millions of examples of math problems and learn the patterns, but they can't “do” math. They're powerful but brittle. They're always going to fail as complexity rises. They miss too many of the core things that we would need to trust them with to get into the really robust use cases envisioned for much more advanced AI. They know that “knife” and “sharp” and “cut” go together—but they have no model for why someone would freak out if a three-year-old has a knife.
What's something you've learned that could help people be more effective when they use LLMs?
LLMs tend to surface whatever was dominant in their training data. That usually means they're going to give you a standard answer, just whatever shows up the most in textbooks or online posts. It's not always wrong, but it's not complete either. If I ask a question about economics, I don't want to get fed the dominant position as the only answer.
So, I'll tell the LLM, look, I don't know anything about this topic. But I do know there are simple answers and there are complex, expert answers. So please outline beginning, intermediate, and advanced levels of understanding. Tell me what misconceptions beginners usually have. And then give me the specific terminology I need to search to understand the advanced level. Now I'm giving the model a chance to not just follow the dominant pathway but to retrieve what might be hidden or harder to find.
