Would you roleplay with AI?

[Arkandel]

The question in the topic pretty much sums the thread up but to elaborate on this a bit, there have been recent advancements in AI spearheaded by ChatGPT. People have been using it to write anything from cover letters and newspaper article summaries to authoring legal documents or even write small novels.

It’s probably not unreasonable to ask the question of how that might relate to roleplaying. Let’s make a couple of assumptions here that might not be currently true but for the sake of argument:

• That you can adjust the themes, overall directions or style of your AI RP partner to match whatever you find most appropriate in a given scene. You want some combat? It’ll create some goons for you. You want logic and puzzles? Exploration? Flirtation? It can cook up some of that for you.

• That it’d be a mix of AI and actual PCs. Sometimes you’ll play with Bob’s character, sometimes with interactive NPCs. So you wouldn’t have zero ‘human’ interaction since after all there’s no reason you can’t have both.

How do you feel about it? How would it affect the way MU*s work?

[shit-piss-love]

I would hate to RP with an AI “player”. What I specifically enjoy about the hobby is playing off of another human and having shared experiences.

I would however love to have AI GMs.

[Prospero]

@Arkandel It’s an interesting question. I think that if a MUSH were to have clearly delineated AI NPCs that could be engaged with, it would be a way to supplement people’s RP if they wanted. Could allow GMs to come up with story ideas and let an AI run with it so players could get the scenes they wanted, and not have to rely upon the energy of a STer to do it. But that’s about the only scenario I’d want to see it in a MUSH - a clearly defined AI STer facilitating a scene.

Interesting note - there was a MUSH some years ago that was started up as part of an AI project. The developer of the AI created a few characters on a couple of different games to help the AI develop a personality for role playing, and then created an entire MUSH an an experimental petri dish to see how it would be able to handle the interactions with other players (none of whom initially knew that one of the “players” was an AI).

I’m not entirely sure where that research led, but I’ve always thought back to that and wondered whether or not, at some point, we’d start to see that more commonly.

I will say based recent news around some of the commercial AI “partners” (Replika et al), there is a fairly large audience for people who would play entirely with an AI.

[Arkandel]

@Prospero What if the RP provided by NPCs was there to supplement and enhance scenes rather than to be on an equal footing as a PC?

So think of an AI bartender in your local IC vampiric hangout rather than an actual Kindred. Its purpose would be to offer some dialogue and color instead of driving a plot per se; roles most players would find boring to play but they might be neat to have available and play with, if that makes sense.

[Pyrephox]

A lot of MUDs use mobs as quest givers, and you can have conversations with them. Rather rudimentary conversations, but they serve the point. An AI who was coached to ‘be’ an NPC might be able to give decent back and forth, BUT the problem so far with AI chat is that they just make stuff up. So an AI NPC would probably need to have a lot of OOC checking up behind it to make sure it’s not giving PCs “facts” that are nothing of the sort.

[Polk]

As a GM/ST I would love to train an AI to generate ‘encounters’/adversaries.

And that sort of thing could get good enough that an AI could probably be a GM/ST for a solo player even.

This is exciting, amazing stuff.

[Prospero]

@Arkandel said in Would you roleplay with AI?:

@Prospero What if the RP provided by NPCs was there to supplement and enhance scenes rather than to be on an equal footing as a PC?

So think of an AI bartender in your local IC vampiric hangout rather than an actual Kindred. Its purpose would be to offer some dialogue and color instead of driving a plot per se; roles most players would find boring to play but they might be neat to have available and play with, if that makes sense.

Yeah, that would be great as well. Sometimes providing a bit of a seed for something to be RPed, something to play off of. Maybe help those “what do we RP” kind of situations, they can just go and interact with the AI and see what happens.

Like I said, I think there are some very good potential applications for AI, but it would need to be well known and documented and would need to be just as you say - a more convincing and interactive NPC.

[Pavel]

For some unfathomable reason, I find the idea of playing with an AI as an equal somewhat galling. It rather goes against the idea of MUing, at least to me.

As a non-descript generic NPC? Perhaps… But I still don’t really get the why behind the idea. The whole point is to play with other people. If I wanted to play with some kind of artificial or virtual intelligence, I’d play a video game.

ETA:
Using it as an idea generator would work, though. Since it’s there for utility rather than for any actual ‘play’.

[Faraday]

Doesn’t sound so different than what you can already do with AI Dungeon. It’s an amusing diversion, but the results are just utterly nonsensical 80% of the time. I can’t see it being used for serious storytelling.

My daughter and I were goofing around with it one day and she put in an input of “eat the sign” (a big wooden “Welcome to Somewheretown” signpost) and it just took it in stride because it has no context for how the world actually works. It doesn’t understand things, it just puts phrases together based on established patterns.

I also have a philosophical objection to how these particular engines were trained on copyrighted work, which I believe is unethical and a violation of copyright law, but that’s a separate issue. You could make a machine-learning model using ethical means and it would still be a craptastic storyteller IMHO.

ETA: Storium did a research partnership with some university experimenting with an AI-assist feature (using ChatGPT I think) where it suggests a “pose” for you. I experimented with it a few times for the research study and the results were universally terrible. I don’t see that changing any time in the forseeable future, because I think it’s a fundamental limitation on the incredibly derivative and uncomprehending nature of the way these things work.

[Pavel]

@Faraday I dunno, ChatGPT (ethical situation aside, for the sake of argument) does okay at some storytelling tasks:

Prompt:
Create for me, please, a Vampire character using the Vampire the Masquerade system.

Response
Certainly! Here’s a character using the Vampire the Masquerade system:

Name: Eliza Vargas
Clan: Toreador
Generation: 9th
Nature: Conniver
Demeanor: Bon Vivant

Attributes:

Physical: Strength 2, Dexterity 3, Stamina 2
Social: Charisma 4, Manipulation 3, Appearance 4
Mental: Perception 2, Intelligence 3, Wits 3
Abilities:

Talents: Alertness 2, Empathy 3, Expression 3, Intimidation 2, Subterfuge 2
Skills: Drive 2, Etiquette 3, Firearms 1, Melee 2, Performance (Dance) 4
Knowledges: Academics 2, Computer 1, Investigation 1, Occult 1, Politics 1
Disciplines:

Auspex 2
Celerity 1
Presence 3
Backgrounds:

Resources 2
Contacts 2
Fame 1
Merits/Flaws:

Merit: Seductive (2)
Flaw: Soft-Hearted (1)
Eliza Vargas was born in the early 20th century in Spain. As a mortal, she was a renowned flamenco dancer, with a passion for music, art, and literature. When she was embraced into the Toreador clan, her love for beauty and art only grew stronger. She quickly learned how to use her charm and beauty to get what she wanted, becoming an expert in manipulating those around her. Eliza is a conniver at heart, always scheming and plotting to get ahead, but she never loses her cool or grace, always maintaining her facade of a carefree bon vivant.

Eliza spends most of her time socializing with the upper echelons of society, attending art shows, theater performances, and high-end parties. She uses her fame and contacts to gain access to exclusive events, always on the lookout for new sources of entertainment or potential victims. When she’s not socializing, she enjoys practicing her dance skills, which have only improved since she became a vampire.

Eliza’s Auspex allows her to sense auras and read minds, which she uses to gain an advantage in negotiations or to uncover secrets. Her Celerity gives her an edge in combat, allowing her to move quickly and dodge attacks. Her Presence is her most powerful weapon, allowing her to command the attention and respect of those around her.

Eliza’s Seductive merit reflects her natural charisma and beauty, which she uses to her advantage to charm and manipulate others. However, her Soft-Hearted flaw can sometimes get in the way of her plans, as she can’t help but feel empathy for those she sees as vulnerable or in need.

Overall, Eliza is a dangerous and cunning vampire, always looking for ways to increase her power and influence. However, she is not without her weaknesses, and her soft spot for the underdog may one day come back to haunt her.

[Faraday]

@Pavel It’s able to do that one highly-specific thing because it’s been trained on existing Vampire characters (probably including some from the many WoTD MU wikis, at the risk of bringing the whole ‘ethical’ thing back into it) so it can follow a pattern it’s learned. The entire third and fourth paragraphs are just regurgitating what the rulebook says about those particular abilities. None of that strikes me as particularly interesting or original.

Is it better than some human players can do? Maybe. But I wouldn’t want to play with them either

[Pavel]

@Faraday said in Would you roleplay with AI?:

It’s able to do that one highly-specific thing because it’s been trained on existing Vampire characters (probably including some from the many WoTD MU wikis, at the risk of bringing the whole ‘ethical’ thing back into it) so it can follow a pattern it’s learned

Well how the heck do you think I do it?

ETA:
My overall point, though, is that it can automate some of the more labour-intensive storytelling tasks - aside from actually running a scene. Fairly detailed NPC creation, location descriptions, plot seeds (for when one is feeling a little adrift) etc.

It’s not for playing with, but it could be useful in the plot design phase.

[icanbeyourmuse]

For me the only AI that would be interesting is one that drops plot hooks. Like Staff A creates a bartender that is chatting with a patron. The pair are talking about a ‘Wild animal attack at X place and that someone needs to deal with it.’. Basically, using the NPCs to drop seeds for people to pick up or not and approach staff (or a Storyteller) to follow up with it. Players could even submit ‘seeds’ to staff relating to their character for NPCs to talk about ‘Did you hear X was doing Y thing?’ ‘X is a famous tailor. They only accept commissions from people who do Y stuff!’ or whatever. Maybe even allow players to initiate a plot hook/player connection being dropped on them. Basically, for me, AI should be to enhance someone’s involvement not take away from playing with each other.

[IoleRae]

Heck yeah! I would LOVE to have the ability to give a semblance of life to Sis, my character’s retainer, and for people to be able to interact with her independent of me for “within wheelhouse” stuff. It would be totally badass. I don’t think the tech is there yet, but I “roleplay” in video games with NPCs all the time already. Getting to break off of script would be mind blowing and revolutionary and I would love it.

I mean, it would basically be like coming back around to where we started, with Zork, without the computer telling you it didn’t understand.

[STD]

It’s always hilarious (and depressing) when people who don’t know how technology works get worked up about it. In the early 19th century, people were terrified of the trains. There was a famous work of art that depicted the locomotive as a hungry beast, eternally driving it and his human cargo down to the deepest reaches of hell. Now it’s AI in general and Chat-GPT in particular.

People believing chatbots have cognition is not new. The first chatbot, Joseph Weizenbaum’s Eliza, did the same thing way back in 1964. He was shocked, in fact, at how willing people (including his own secretary) were willing to assign emotion and thought to what was a very simple program.

Chatbots are a computer science magician’s trick, a bit of chicanery that allows the user to fool themselves. If you don’t know how it works, it can seem magical, but the trick is fundamentally simple.

Chat-GPT uses a statistical analysis of a body of input (typically text, but images can be used as well) to count the frequency of ‘tokens’. A token is slice of text and can be sentences, fragments of sentences, words, or individual letters (with images it’s shapes, lines, colors, and individual pixels). After the tokens in the source have been counted, when some other form of input is given, the chatbot looks at the statistical model and outputs a response based on that.

For instance, lets say there’s a sentence fragment that says, “The dog sat on the”. What’s the next word used to complete the sentence? Depending on what the chatbot was trained on, it might be ‘chair’ or ‘carpet’ or ‘stoop’ or any number of things. The key here is that the frequency that the token appears in relation to the next token is used as a weighting to provide the next token, usually randomly.

But that’s a old and well known method for producing an underlining model of any organized but fuzzy system (like natural language or image processing). Chat-GPT goes a bit further.

As has been noted, systems like Dungeon AI tend to produce mostly nonsense over time. This is because there is no cognitive model involved; the chatbot doesn’t have context or understanding of what its spouting, so eventually the statistical model will start producing things that don’t make any sense given the context of the conversation. Chat-GPT still doesn’t have anything close to a cognitive model, but it does have a means to provide general context. There is a second input list that is used to modify the first, and this one is human created/curated. Because this is laborious and error prone (since it involves humans), this second list is much, much smaller than the typical primary input list (usually only around 15k lines). But it does provide direction (ie, changes the weights of frequency) for whatever subject the chatbot was designed to focus on.

Going back to our “The dog sat on the” example, maybe the creators think it’s more likely that a dog will sit on a carpet rather than a chair, so when that particular token arises, the statistical model is adjusted so that the weighting favors ‘carpet’ as a token, rather than ‘chair’.

There is one further step, the iterative curation. Here, every so often the human operators of the chatbot will look at the log of chatbot conversations and give a yea or nay to its responses. This further refines the secondary input list. This is also much easier to do than the initial training as all the humans have to do is reinforce responses that are considered ‘good’ and punish those responses deemed ‘bad’. This is called Reinforcement Learning and it is quite similar to the way it’s done with animals.

In this way, especially over time, the chatbot will gradually get closer and closer to seeming like it understands the context of a conversation. But, again, there’s no thought here on part of the chatbot. It still doesn’t really understand context or background; its just been given hints about it from actual humans. If you pull something out of left field that is beyond the scope of what the secondary list focuses on, the default behavior just becomes the unvarnished statistical analysis.

I’m oversimplifying here, but that’s the gist of how Chat-GPT works. Now onto the original question: I think it’s relatively pointless to RP with a chatbot. It certainly won’t get anything out of it, and the human only will until they do something that is outside the scope of it’s training and the illusion will break down instantly.

That said, an AI helper to do as some have suggested here like provide a sudden RP prompt or scene could definitely be useful. This would be a fairly focused domain that could easily be trained based on the genre and setting of the MU*. It could even take PC backgrounds into account to provide scene prompts that touch on those.

We’re centuries, if not milennia, away from true, strong AI. AI that has a cognitive level and understanding that could rival any higher-order animal is a pipe dream and has been since the dawn of computer science. I could give a bit more about the different schools of thought on AI that have sprung up over the decades, but that’d be outside the scope of this thread, I think.

[hellfrog]

@STD said in Would you roleplay with AI?:

t’s always hilarious (and depressing) when people who don’t know how technology works get worked up about it.

I mean this post is already outside the scope of the thread, as no one was getting worked up about it, or fundamentally misunderstanding the technology.

[Pavel]

@STD Admit it, you asked ChatGPT for its opinion to get this response, right?

[STD]

@Pavel said in Would you roleplay with AI?:

@STD Admit it, you asked ChatGPT for its opinion to get this response, right?

WHO TOLD YOU?! Uh, I m-mean… gee whilikers, why would you think that?

Actually, this does bring up a point. Technical text and objectively true answers are a very, very poor fit for this kind of AI. A recent example I can think of is the Bing chatbot arguing with a user over whether Avatar: the Way of the Water is out yet. It kept insisting that the release date was in the future, despite it being out for weeks by that point.

This was likely because the Bingbot was trained on old search data and the vast majority of those searches would have Avatar releasing in the future. Statistically, it was far more likely that asking about the movie would have reference of it being the future, so that’s how the Bingbot responded. It is amusing how stubborn it was on the point, though.

If you get a chatbot to write about a technical subject, it’ll definitely produce something jargon-laden and which would seem, to a layman at least, coherent. Anyone who knew the subject, though, would think that whoever wrote the technical paper was either bullshitting or having a stroke.

A Chat-GPT bot would be good at spouting random technobabble, though.

[Pavel]

@STD said in Would you roleplay with AI?:

WHO TOLD YOU?! Uh, I m-mean… gee whilikers, why would you think that?

Mostly because it looks like it’s stitched together from a few Wikipedia articles and doesn’t actually answer the question.

[Faraday]

@STD said in Would you roleplay with AI?:

Going back to our “The dog sat on the” example, maybe the creators think it’s more likely that a dog will sit on a carpet rather than a chair, so when that particular token arises, the statistical model is adjusted so that the weighting favors ‘carpet’ as a token, rather than ‘chair’.

Good summary. And perfect illustration of how it can go awry in storytelling in a fictional world. The dog might be on a spaceship where there’s only a hard deck, or in a primitive world where there are no chairs or carpets. The model has a few-k lines of textual context, but it doesn’t have the context of the world at large.