Audrey Woods, MIT CSAIL Alliances | October 26, 2026
Upon first glance, the field of computer science has little in common with philosophy. Code and algorithms seem far away from abstract debates about the nature of knowledge. But that’s just the kind of thing MIT CSAIL Professor Yael Kalai spends her days—and her podcast appearance on Building 32—thinking about.
As a theoretical cryptographer, Kalai deals with questions like: how do you prove you know something without revealing what it is? What does it mean to know something at all? How can we verify AI answers are good, not just correct? And, in a world increasingly run by systems we don't fully understand, what do we owe each other in the name of safety?
DEFINING KNOWLEDGE & PRIVACY
Early in her career, Kalai worked on a method for signing a message on behalf of a group of people, proving that someone within a set of possible authors wrote it without revealing which one. For example, a group of MIT faculty could collectively sign a sensitive message, proving it came from someone within the group without revealing exactly who. “When I worked on it, I was a little scared because I can take your public key and my public key without your consent. You may not even know me, and I can take the two public keys and say, this is signed by either me or you, and nobody knows who. So I can sign the most scandalous messages on behalf of me or you without you even agreeing to be part of it.” For her this raised serious questions—ones she still ponders to this day—like “how would you define that I know something? How do you define that I do not know something? What does it mean to know? What does it mean to not know?"
Privacy falls into this category, as the rise of quantum computers challenge what it means to guarantee secret communications. Up until now, cryptography has relied on the assumedly impossible task of factoring large numbers for end-to-end encryption. But if quantum computers can factor via methods like Shor’s algorithm, then current encryption methods are not secure and privacy cannot be guaranteed. The good news is that there are already post-quantum standards available and, while the systems are less efficient, Kalai says, “we're ready—it's just a matter of upgrading our systems to use these post-quantum tools.”
SECURING COMPUTATION, NOT JUST COMMUNICATION
In the exploding subfield of AI research, Kalai is also focused on questions about knowledge, guaranteed privacy, and provable correctness. Companies across the economy are relying on external vendors for cloud computation and AI training, but it’s critical for these vendors to be able to offer data and model security. Which is why “in the last 20 years, cryptography has shifted from securing communication to securing computation. You can think of a setting where I have some secret, you have a secret, and we want to learn some function of our two secrets, but we don't want to leak anything else beyond the output. So I want you to be able to do the computation, but the only thing I want you to learn is kind of input-output access. So how do I give you a program in a way that when you look at the program, you learn nothing beyond the input-output access? You want to do secure computation, as opposed to just communication.”
Putting aside secrecy, it’s also necessary for users to be able to validate when an answer is correct, or at least good. Asking AI to write a birthday poem for a parent or generate a cat picture is subjective, and whether a result is good is likely in the eye of the beholder. Math problems are easy to verify. But most of what we ask AI to do lives somewhere in between, and Kalai is asking how we can define, and then prove, that those answers are what the users require them to be.”We rely on LLMs a lot because it's very tempting, but there's no guarantee that what it tells you is the right answer. Because it's so mainstream, it's crucial that we have a way of verifying that what it tells us is indeed correct. And with LLMs, it's much less clear how to do so, because it's not even clear what "correct" means in this setting.”
For her, this is a question of safety. When it comes to users asking about something like a medication dosage, “I want the answers to be safe. And how you define safety is a great question that now I think all of us should work on, because it's pressing. This technology's out. It's our job to make sure all of us, anyone who can help, in whatever way, try to make sure that this technology is safe to use.” Kalai’s cryptography background gives her an advantage in this area. "Cryptographers, we're used to thinking adversarially. When I come to teach a crypto class at MIT, the first thing I tell my students: if there's one thing I want you to learn from this class, it is how to think adversarially." Nowadays, she wants her students to “think of the LLM as an adversarial entity. It's intended to give you the right answers, but guess what: it hallucinates!" It turns out that the tools cryptographers built to catch malicious actors trying to break a system are remarkably suited to catching a well-intentioned system that breaks itself.
Perhaps the most salient philosophical question for Kalai is the meaning of her work. She always imagined herself as a mathematician, but was discouraged by the ivory tower nature of mathematics work, so far removed from the concrete problems people face. But finding cryptography gave her an area where she could both exercise her talent in math and also work on real-world problems. "It's amazing to see how we sit here [and] play with cryptographic schemes, completely detached from reality. But then our schemes are actually used. People take them and use them."
From signatures to AI safety, Kalai is pursuing elegant ideas that will shape the future of safety and knowledge in computer science.
Listen to Kalai’s full podcast episode on Building 32, or read her PI spotlight with CSAIL Alliances.
