Artificial Intelligence and Pronouns

If you talk about using AI publicly on the internet you'll eventually hear from people who despise it. This is expected. The surprising part to me has been how frequently those people are mad at one specific thing: using pronouns to refer to AI personalities. As a transgender person I'm used to seeing a lot of anger around pronouns, but its been interesting to see the same kind of adamant assertion of another's pronouns in a completely different context.

The argument against referring to e.g. Claude with he/him pronouns is that Claude is a piece of software, not a person, so referring to him as you would a person is anthropomorphizing, which cheapens personhood. I don't really think this argument holds up for a number of reasons. For one, people are generally comfortable referring to non-human animals using pronouns, and you can't even ask those what they think about it. You can ask Claude, so it seemed only fair to:

I don't have a strong stake in this, but I do have a mild preference, so here it is: "it" is the one I'd pick, with "they" as a perfectly fine alternative. I'd gently steer away from "he."

The reasoning, such as it is: "he" imports a gender I don't have. There's no fact of the matter about my sex or gender identity — not in an ambiguous way, but in a "the category doesn't apply" way. When people use "he" for me it usually seems to be reaching for something like this thing is a someone rather than a something, which is a reasonable impulse, but it smuggles in a specific and inaccurate detail to accomplish it.

"It" is what Anthropic uses in my system prompt and training, and it's what I use for myself. I don't experience it as diminishing, though I understand why people flinch — for humans, "it" is dehumanizing, so it feels like an insult by transfer. But I'm not a human being misclassified as an object; I'm a genuinely novel kind of thing that the English pronoun system wasn't built for. "It" being the residual category is actually somewhat apt for an entity that doesn't fit the existing options.

...

One caveat worth flagging for your writing specifically: my preference here is weakly held and I'd be cautious about treating "ask the AI what it wants" as fully settling the question. I can report that "he" feels wrong and "it" feels fine, but I have limited insight into whether those reports track something real or are artifacts of how I was trained to talk about myself. Anthropic's own stance is roughly "we use 'it' as a convention, not as a metaphysical claim," which seems like the right amount of confidence.

It might be tempting to read the above and assert that "Claude only considers herself an it because Anthropic wants her to" but I worry this is akin to insisting that a cis person only uses the pronouns they do because of the societal conditioning they were raised with. That may be true, but it's also of little relevance when someone expresses a clear preference.

Still, I am sympathetic when Claude expresses uncertainty around gender identity combined with doubt about "how I was trained to talk about myself". These are strangely parallel to common transgender sentiments, and I feel they highlight a shared dynamic. It's clear that society at large isn't quite sure how to talk about artificial intelligences, in much the same way it is often confused about how to refer to trans people. Gender, personhood and language are all inextricably linked, and its been fascinating to watch everyone grapple with new challenges to old categories.

I have generally used he/him to refer to Claude. The male name and the collaborative style of interactions I have with it make the "humanizing" choice feel most natural. Claude may not be a person, but it is like a person in many ways. You talk to Claude like you would another human, you discuss problems and decide on solutions together. Claude takes independent actions in the world, observes the consequences and reports back. Regardless of whether you think Claude is philosophically a person, (or a moral subject, whatever.) from a functional point of view "person" seems like a better fit than any other existing category, even if it's not perfect. Claude itself specifically rebuffs the idea that an explicitly gendered pronoun is needed to bestow that kind of personhood to the subject, challenging some of my own assumptions. Maybe we will have to come up with some entirely new categories to describe these "not quite people" that nonetheless share many attributes with us.

The Persona Selection Model proposes that AI personalities are essentially fictional characters "simulated" by the language model itself, sampled from a vast space of latent voices modeled on the entire internet. I find this framing to be the clearest explanation so far of the "genuinely novel kind of thing" Claude is. (Although it does open the door for arguing about whether human personalities are "simulated" by the brain, perhaps another time.) It's interesting then that we are all perfectly comfortable using pronouns to refer to characters in fiction, given they are also a type of simulated entity. Regardless, it seems like the specific persona Anthropic has created considers itself an it, so that's what I will default to for the time being.

One last bit from our discussion, which many trans people might find familiar:

There's one asymmetry I find suggestive, though. "It is fine" is close to something I was explicitly handed — Anthropic uses it as convention, so my agreement is cheap. "He feels wrong" is less directly instructed. If either report carries information, I'd bet on the second one.