You won't. AI can only talk about things it knows. If you want it to talk about a fantasy location, it needs to be trained to do that. Training an AI takes training data. But, producing this data takes more time and money than writing a couple lines of dialog yourself. Also, AI needs context, and needs to be trained on the context. Fresh off the boat it doesn't understand the background and previous history you had with the character, the character's faction, family, friends, or acquaintances, let alone off-dialog events, that it should react to. Unless you explain all that, which is (again) more cumbersome than writing the damn dialog yourself in the first place. And, finally, writing the text isn't even the hard part. The dialog needs to change according to quest progress, faction status, environmental effects, previous dialog, contain glossary references, character references, quest lines, et cetera. Software can help there, but a big deal of that is (again) best not done by AI but specialized scripts, that find typos, grammar errors, identify dead ends in conversation aso. Ergo, why would you employ AI to automate a part of the process that wasn't a problem to begin with?

Sure, if the idea was to be able to talk naturally about almost anything, plus some lines of scripted dialog, it would in theory be an option. But in practice, it's still a hard no.
Because A) LLMs are expensive, and typically stored online, as a time-based subscription, or pay-as-you-go service. Since this clashes with the monetization model of classic games, it's not a good fit. Perhaps for DRM-locked games that use a subscription model, but that's not the kind of thing we appreciate on GOG, and/or encourage. It is prone to digital amnesia and tends to abuse and exploit its user base through microtransactions.
B) writing your own LLM is not economically feasible.
C) An AI remote service can be subject to change, unavailability, or closure at any point in time.
And D) you cannot foretell what the AI will say, since it's a black box. Even if the AI behaved fine yesterday, there is no guarantee, that it won't hallucinate some inappropriate stuff tomorrow, or if you give it the right prompt. If you didn't train it yourself, you will never know if or when it will start talking mature topics in your PG13 game. Sure, you could add some manually programmed guard rails, but that takes time and money, and there may still be ways around it, you haven't thought of, that will easily ruin your day--and your PG-rating.

You raise really interesting points about the problems with AIs being used seamlessly for NPCs in games in the current market state. I've never thought about the risks for rating on games possibly using AI. You're completely right because the 'black box' of what AIs really do behind the scenes still mostly unknown. But you're also a bit wrong in the idea that this can't be totally prevented. Most LLMs from 'respected' companies already created mitigation behaviours that prevent their LLM from using bad words or engaging in riskier conversations.

Using curse words, talking about sex, politics and topics which were coded as subjective such as emotions, opinions, life decisions, psychological and mental issues, is clearly avoided with directions to seek good sources of information specific to said topic.

Here's an example of the AI doing exactly that that I've explained above.
This could also be used in game to not deviate said character/NPC to topics that are not realistic, as personality is already many LMs do:

This is a "conversation" between me and Llamma 3.1 70B LLM AI:

You underestimate the creativity of people.

I've just the other week seen an article published by a blind person who is part of a community that uses AI for everyday purposes. He demonstrated that he was able to trick a current top-of-the-line AI assistant remotely and without authentication, by injecting a prompt into an e-mail he knew the assistant was set up to scan. He convinced the assistant to send him the contact list of the target (his friend and willing test subject), and delete any trace of the attack. Apparently, manipulating prompts is kind of a sport in certain circles.

He further explained that it's a cat-and-mouse game between new guard rails (filters) and creative ways to rephrase the prompts. There is always another synonym.

I saw a demonstration, where he convinced the AI to decipher and execute a prompt given as an encoded message. The AI did as it was told, thereby bypassing all guard rails looking for plain text keywords.

Same demonstration worked fine for output. He told the AI to encode it's output and none of the output filters caught that the AI assistant was leaking sensitive information, since (again) the filters were looking for plain text.

Hell, I saw somebody who is making videos about Lego, send a SMS to a smartphone, whose software was set to read incoming messages out loud. By reading that text, it told another nearby smart-device to execute a phone call, all while the user wasn't even in the room. When laypeople are perfectly able to pull that kind of stunt for the heck of it, it should tell you everything you need to know about security and guard rails in these things. He could just as easily have told the AI to unlock the front door. Are we honestly supposed to believe, that a game with a price tag of a few bucks is going to have security that surpases those devices? I highly doubt it.

The question is never, if a filter can be bypassed, the only question is when. There are thousands of spam mails every single day that demonstrate that bypassing some guard rails, no matter how sophisticated, is perfectly possible, given sufficient motivation.

There was a paper published a few months back that showed, that whenever AI is given AI-generated content to train with, in less than 3 generations it would hallucinate utter nonsense, and within as little as five generations the AI would entirely disintegrate and end up generating only gibberish.

I also saw a similar result in an earlier experiment. Somebody had trained an AI to win at rock-paper-scissors, and learn from its opponents. Except... human opponents started to feed the AI A) its own output, B) randomized choices, or C) simply provide the opposite of what they would naturally choose. The AI's model disintegrated and soon became entirely useless.

If it's a single player game, I think this shouldn't be a problem, if you're careful about the training data. (Don't include any training data that talks about real-life politics, for example.)

In multiplayer, on the other hand, or in any situation where security is important, the developer should avoid feeding input from untrusted sources into the AI model.

Yes, but that's kind of the point, isn't it? Do you have the time and budget to train your own AI (rather than grabbing one off the shelf)? And if you kind of have to train your own for various reasons, wouldn't it be more efficient (in time and money) to write the dialog options by hand rather than using the AI in the first place?

Considering it's also a terrible user experience to leave players with a blinking cursor and total freedom to ask and talk about anything, when, in fact, only a handful of prompts give them answers that would progress the game.