The question he some interesting angles.

I don't think AI is people yet, or close to it - so the easy answer is 'it'.

But let's have a think about pronouns and the purpose they serve; accurately capturing the true nature of the referent is not and has never been the point.

Gendered pronouns are an easy example of this: you don't ferinstance need to know ThE bIoLoGICaL sEx of a person in order to refer to them. You don't need to go rummage in a stranger's underwear or take DNA samples in order to call them 'he' or 'she', the words work just fine without any such knowledge. And indeed if you go intentionally misgendering someone because wElL aCkShEwAlLy, all you do is confuse the person you're talking to (and seem like a dick).

Pronouns, in short, are a placeholder for a noun phrase, and we have different ones to help us distinguish between the different nouns in play at any given time. By the time you've parsed out gender, plurality, animate and object/subject distinctions, it's generally a poorly written sentence that has any ambiguity left.

So the question you need to ask is what most usefully aligns with the listener's expectations? How are you framing the conversation?

Consider an interaction with something of indeterminate gender, sentient-acting but not-people: a crow, for example. A crow comes up to you, accepts a chunk of your sandwich then brings you a stone, seemingly in exchange.

When recounting the story, do you call the crow an it or a they?

That's going to depend on a bunch of things - whether it's some random wild bird or someone's pet, how many nouns you need to juggle, and whether you're more interested in the bird or the stone.

The choices you make set up the framing of the conversation, reflect your perspective and shape perception.

Whether an LLM is people... isn't really the point.