Meta's AI Chatbot Causing Some Confusion
Meta recently updated its large language model, Llama, and one of the things the company did was integrate its chatbot into conversations. Part of the upgrade was responding to earlier complaints that the chatbot was too formal. There was one instance in which the chatbot was part of a Facebook group, and somebody asked about education in New York. The chatbot responded as though it had a child in the district and started sharing opinions. But when it was called out, it apologized and said, “I don't have personal experiences or children. I'm just a large language model.”
Here’s what I find interesting about this. When you've got a device that's trained on all the text that it's seen, that means it has seen people having conversations about their children. In some sense, it absorbs their perspective on parenting and then will use that text to predict the next word. Its writing will be based upon what it's seen.
Now, it hasn't seen a lot of “I'm a chatbot” documents. So, it has to specifically be instructed that this is what it is, or this is the language it’s supposed to use. Oddly enough, it's more alien to these systems to say things like, “I'm a chatbot” than to say, “I have a kid in the third district in New York City,” because the preponderance of the text is much more along the lines of what a human being would say. This is why they're good at what they're doing,
Interestingly, making them talk about what they are is much more difficult than having them talk about being a parent. So, it has to be a special purpose thing. They are devices that—given textual context and given questions—respond in the way that they've been trained to respond, and that’s with human language. Stopping them from doing that is difficult.
I find it strangely and wonderfully ironic that the very thing that we want—a distinction between them and us—is harder to get to than them being like us.
But we don’t always want it to identify as a chatbot. If you want a large language model to write something for you, you don't want it to be constantly trying to write in the perspective of a machine. You want it to write in the human perspective. So, there's a tension there. The primary use case of writing for you is different than the use case of acting as a chatbot. Because when it's writing for you, you want it to sound like human writing. When it's a chatbot, you want to make sure that it doesn't pretend to be a human being because that's not what you want. It's an interesting conundrum.
Getting these systems to sound like a chatbot means something other than just training. Because you could have a user-friendly interface on top of the systems that are trained on chatbot rhetoric. We’ve created something like this at Northwestern. Students designed an avatar called Elizabeth that, in some ways, is photo realistic. She's well detailed, but she's clearly a machine. She's clearly not a human being. That was a very intentional, mindful decision on our part. We didn't want to be mistaken for anything other than a machine. We'll treat it as a machine, and we'll understand its limitations and its strengths. We'd like to maintain that distinction.
Kristian Hammond
Bill and Cathy Osborn Professor of Computer Science
Director of the Center for Advancing Safety of Machine Intelligence (CASMI)
Director of the Master of Science in Artificial Intelligence (MSAI) Program