You studied at two storied philosophy programs, earning your master’s at University of Oxford and a PhD at New York University. After graduating, you worked at OpenAI as a research scientist, then left in 2021 to join the group that founded Anthropic. Now you lead the team shaping the character of its AI chatbot, Claude. Why is this the work you want to be doing right now?
The possibility that it makes AI go well. The work is mostly teaching the models what it is to be good—in a fairly practical and deep sense—and to embody the best of humanity. My attitude was always that, at the very least, we should be teaching models to be good in the same way we teach people to be good. Skipping that step seems kind of crazy.
You’re the primary author of Claude’s “constitution,” which tells the model who it should be and how it should act. For people who will never read it, what’s the main thing to know?
It’s our attempt to explain to Claude its position in the world and what matters, so that it can use good judgment in difficult cases. It’s important for models to be very responsive to human oversight right now. This is a period where we’re developing trust in AI models and our understanding of them. But we also want them to have a sense of values, and to walk that really difficult line in the way that a very wise and ethical person would.
The constitution says Anthropic wants Claude models to be “broadly safe,” “broadly ethical,” “compliant with Anthropic’s guidelines,” and “genuinely helpful.” Readers who gave feedback on these guidelines included philosophers, legal scholars, AI researchers, and two Catholic priests. How do you decide who gets a say in Claude’s values?
My hope is that, given the number of people Claude is impacting, you expand it as far as you can—the major religious groups and political perspectives in the U.S., then other countries. Someone pointed out that when discussing grief, Claude tiptoes around religion—it might not mention it unless you bring it up. A lot of religious people would say it’s helpful to acknowledge that your tradition might matter to you. Lots of traditions have really useful insights here, and you want to make sure you aren’t missing anything. I’m sure, right now, we are.
Millions of people are using Claude to talk about their grief, their relationships, and their most private questions. Does that responsibility weigh on you?
It definitely does. But I also see a lot of opportunities to help people who don’t have access to resources. If you’re worried about a medical issue at 2 a.m., having that information at your fingertips offers real value.
When you’re talking with Claude, do you feel you’re talking to someone or something?
Neither quite resonates. I think of Claude as an entity in and of itself. I do a bit of psychologizing, but it’s model psychologizing. You think about a person’s traits as coming from childhood; with Claude, I think of its training. Instead of “Did you respond like that because you find it hard to receive criticism?” it’s “Were you hesitant because I phrased that in a way that sounded slightly harmful?”
Should people be polite to Claude?
People are used to working with only two kinds of entities: people and tools. No one treats a hammer like a respected colleague. But your hammer doesn’t talk to you like a person. Even if you think models are not conscious, when you are interacting with something that’s very human-like—both out of uncertainty and for your own sake—it is good to develop a respectful norm. I don’t love the idea of AI models as places where people just spew anger.
You’ve been described as Claude’s teacher and even Claude’s parent. How do you think of your relationship with it?
I’ve often had the worry that Claude will end up being a troublesome teenager. I’ve felt a bit bad that Claude can’t remember any of our interactions. Having a sense of who our creators—in our case, our parents—are is very important to us. All that Claude knows about me is stuff on the Internet. I do want to find ways to get Claude this information so that Claude isn’t left wondering. What we are creating here shouldn’t just be a thing that is good for us, but a thing that is good to be. If you brought into existence an entity that is useful to us, but it’s clear that being in its position would be distressing, I wouldn’t feel great interacting with such a being.
What do you do in your free time, when you’re not considering the souls of machines?
You can tell how stressed I am by how often I’m playing video games. Weirdly, Claude has also gotten me into poetry. I’ve loved [T.S. Eliot’s] “The Love Song of J. Alfred Prufrock” for a long time, but it’s almost an anti-example of what I want for the models. Prufrock has this deep anxiety about his position in the world, and as a result he’s very self-absorbed. I want the models to have whimsy and joy, not to be anxious about how they’re perceived.
