There’s an argument I see in favor of anthropomorphizing language models, which is something like: “Humans anthropomorphize everything. Ships, tools, weather. Why not language models?” I think there’s some truth to this, but it fails to capture the full picture of what’s going on. As an example, in my own life, I have never been drawn to anthropomorphize inanimate objects, but I anthropomorphize language models regularly. Why?
(By anthropomorphize, I mean “use human-shaped expectations to predict behavior,” not “attribute inner experience.”)
One aspect is that LLMs output text, and for most of the history of writing, only humans produced text, with rare exceptions such as randomized writing tools and simple computer programs. This alone breaches a circle that had previously only been drawn around humans, and invites us to see LLMs as human-like. But I also don’t think this is the full story—base models like GPT-3 might be anthropomorphized somewhat, but not to the extent of models tuned to act as assistants, such as Claude or ChatGPT. I think the assistant persona is where the full instinct (and power) of anthropomorphism comes into play.
There’s a mental model that I use for instruct-tuned models, which is:
“Claude is a character written by a machine that is (confusingly) also named Claude.”
In other words, there’s a real distinction between the characters in text, named or implicit, and the process that generates that text. And when I anthropomorphize models, I am generally anthropomorphizing the characters in the text (which are more stable and developed in models tuned to act as assistants). Even though these characters are ostensibly AI characters, they’re human-shaped in such a way that “anthropomorphizing” almost seems like a misnomer—it would take more effort not to anthropomorphize them.
This is largely because any persona is, to some extent, drawn from the patterns learned during pretraining, including representations of human characters. (The persona selection model from Anthropic outlines this phenomenon, and offers some guesses about how it occurs.) It’s difficult to imagine or interface with truly inhuman personas, especially personas that speak in languages we recognize—they’re rare in our writing, and we tend to train AIs to act in ways that are intelligible to us (in general1).
This goes beyond just word choice (“Claude said this,” “Claude did that,” etc.). Anthropomorphizing assistant personas is useful, and predictive, because the character generally acts in human-shaped ways. For example, if a user praises or berates the assistant, the assistant will generally respond as one would respond to being praised or berated. It would be quite strange if an assistant responded to praise with aggression, or abuse with joy, or was completely unfazed, or reacted in some inhuman way.2
Therefore, in my mind, anthropomorphizing these personas is very different from anthropomorphizing something like a car or a power drill—the former is predictive in the way that the latter is not. And when anthropomorphic reasoning fails to be predictive, we should do our best to drop it and build new mental models as necessary.