I Fell Down The "AI Roleplay" Rabbit Hole

I, a person who does not like GenAI, spent 120$ on tokens for Dreammir.AI. And in the span of roughly a week, I used all of them.

I could be embarrassed to admit this, but my embarrassment does no one any good. So instead I am going to lean into it, entirely: Yes. I spent dozens of hours on this program. I learned the way it worked, immersed myself in it entirely, and for a very brief time I was captivated by it. This is in spite of all my beliefs on LLMs, in spite of how I’ve badmouthed them in the past and stated that they can generate no art worth consuming.

So, then, why?

First: we need to define what these types of interactive AI narrative “games” are. You might have seen them being advertised on the internet. Choice-based narratives promising “infinite possibility” or some such. If you go to the Dreammir.ai website, this is the first image you’ll see.

And it does keep that promise, for the most part. Dreammir.ai is many things. But it is not a scam. It tells you exactly what it is charging you for and its pricing model is far from exploitative. It earnestly attempts to do as it claims. It’s just that, for those who hate LLMs, what it ‘claims’ it does causes inherent squick. This was true of me, and still I did it.

The quickest answer, for ‘why,’ is that I’ve always enjoyed choice-based narratives. I am a recovering D&D addict who hasn’t played a game of D&D in well over three years, and I can’t foresee a future for myself in which I am able to schedule a session again. Dreammir.ai is like that, in some ways. A visual novel so flexible that it’s less like a choose your own adventure, and more like playing a TTRPG with a DM who is pretty much down for anything.

Until you find the limits. But we’ll get there.

If you’re like me, the sight of these programs might always have prompted idle curiosity. Muses of: “What is it actually like, though?” So what follows is an attempt at: “I did it, so you don’t have to.” Hopefully, in that, I can redeem myself for my feckless, yet intentional participation in a technology that is replacing humanity and destroying the planet. And on a more personal level, maybe I can get enough out of this article to justify that 120.00$ USD forever gone from my bank account.

So let’s start with an overview. What, exactly, is Dreammir.ai. How does it work?

Dreammir.ai is part of a broad emerging ‘medium’ of interactive AI story games. It’s not even the most popular. Others include programs like talefy.ai, which is, on the surface, cheaper. Of course, if you clicked the link, you’ll see the genre of story that Talefy trends toward, right away.

In reality, even though I happened to screencap it when a succubus girlfriend was the headliner, most Talefy ‘scenarios’ (what I will be calling them from this point on) cater to women’s romantic fantasies. Dreammir, which I chose, seems to be more of a mix, and seems to not specifically target romance.

This begs the obvious question: Is this porn? And the nuanced answer is: It very well can be. Do you want it to be porn? Because it will go in whatever direction you want to take it.

And in my case, the direction I wanted, at first, was easygoing slice of life, maybe with some romance. So looking at the scenarios available, I chose “Summer Days.” An anime-esque story about five old friends staying together on an island. Possibly with a mysterious secret, but always with one wrinkle in the plan: only three rooms are available, for the five of you.

The interface is pretty simple. You have a text box, and in that text box you can type whatever you want, choosing to “say” “do” “whisper” “shout” or, oddest of all “sing” whatever you are typing. (I will be honest, I never sang.) This prompt is given at the end of each interlude. Every time you take a “turn,” typing something, you are prompting an LLM. And each turn costs tokens. You can see them there, on the arrows in the bottom right corner.

Dreammir.ai features flexible pricing, actually! Based on what qualities you want, you can go as high as 50 tokens per turn (full voices, the best AI model available, roughly 12.5 cents per turn) or as low as 15 tokens per turn (no voices or narration, the less good model. roughly 3.75 cents per turn). You can even switch in the middle of a scenario, turning off and on the more expensive additions, as you wish. For a huge chunk of this, I had everything on max, which is why I burned through so much money, so quickly.

I know. Expensive. But whatever else this is, it is genuinely an entertainment medium that is only possible through AI. No choose your own adventure designed by a human could be this flexible. No roleplay with a friend or fellow human could so instantly provide audiovisual accompaniment. Like it or not: This is GenAI ‘innovating’ new possibilities for our creativity, providing a medium that truly did not exist before. I really did want to experience it at its best, so its best is what I got.

Even though, once people got sorted into rooms, it couldn’t keep track of who was staying where.

Yes, one of the pivotal plot points of this scenario is that there aren’t enough rooms. Well guess what? Whatever model for LLM you use, whether it’s the weaker ‘moon’ model or the stronger ‘sun’ model, both seem incapable of understanding that if Raven is staying with Cassandra tonight, she’ll be staying with Cassandra tomorrow, and not suddenly with someone else.

This is not the only issue Dreammir.ai has. The above image shows a remarkably coherent hotel room background, and image of Cassandra. But sometimes things get…weird. Random people will be in the background when they shouldn’t. Or artist signatures will show up. In a mecha scenario I came up with later, my mech was always in the background, but never looked the same. I was lucky if it even was a mech. Sometimes it’d be a spaceship. The AI generates ‘atmospheric music’ for you, whether you like it or not. And the music is, consistently, abysmal. Below, you may listen to a sample of what was playing as I talked to Cassandra in that hotel room.

And yet, I already couldn’t stop.

Because I had successfully, just at a text prompt, pulled Cassandra away from the group and told her that I had a crush on Raven. Because I had so effectively derailed and defined the narrative. This is my problem. Just the mechanics of being able to change a story are enough to set me off, to send my neurons firing.

So after my brief experiment with Summer Days, I bought a bunch of tokens, and I generated a scenario of my own. I was determined to either alleviate the problems of the LLM, or to at least ignore them.

Campus life was my first original creation. The idea was to still do slice of life with some romance, but to make the protagonist a college freshman who had no idea who he wanted to be, yet, and by extension no idea what kind of relationship he wanted to have. He would be surrounded by women with many different relationship models (a strict monogamist, a polygamist, a no-strings-attached hookup-seeker, a woman deeply into kink) and in a sense discover himself.

But after one slice of life story, followed by another, I quickly became bored.

The problem started in the above scene. One thing I have not mentioned is what Dreammir.ai is good at. Surprisingly: creating characters. Characters who have broad archetypes, then redefine themselves as you dig deeper. Rebecca Torres, pictured above on the left, is a psychology major who is an intelligent and capable woman, but has always felt driven toward a submissive role in relationships. Not just sexually, but philosophically. Spiritually, even.

My character then proceeded to connect with Rebecca and, with little adherence to her preferred relationship style, started dating her.

This is the first emergence of a core problem of Dreammir.ai: The LLM is not actually the god of this world. You are. I intended for my inexperienced character to pursue this woman who so clearly knew what she wanted, and then get in over his head. The result might be funny, or awkward, but it wouldn’t work. Right? Instead, as soon as I signaled to the AI that my character was vanilla, and interested in a vanilla relationship with Rebecca, Rebecca became vanilla. Yes, the characters are surprisingly well-drawn. Until it becomes convenient for them to be someone else. Then they change on a dime, even when you don’t explicitly ask them to. Characters are shockingly believable until you see their fuzzy, mutable cores.

So, now bored with how easy this world had become for my character, I wrote a memory note, a broad directive that changes the story: “At the upcoming party on Friday, someone is murdered. Then the genre and tone of the story changes to mystery/thriller until the killer is caught.”

Campus Life became the longest scenario I ever did. Going on for over 450 turns. And in truth: I came away disappointed with the thriller elements.

It wasn’t that they’d been forced in later, and it wasn’t that they made zero sense. It’s that it never got dark enough, really, to be a thriller. When only one student had died, and it had been a long time, I finally put in a memory note claiming: “The player will be forced to choose who lives or dies. Saving everyone will be impossible.” This led to the murderer sending text messages to me, having a gang of goons trying to kill people in my inner circle. At first it was thrilling, intense!

But ultimately, everyone survived.

People got injured, sure. But all I had to do was text Rebecca. Tell her to check up on the person I wasn’t saving. And just like that, they were saved.

Oh, yeah, Rebecca. Let’s talk about Rebecca.

Rebecca Torres, as it turns out, is the coolest girlfriend you could ever possibly have. Is she submissive? Yes. I mean, she literally said she was. But despite being submissive, she is strong-willed, organized, supportive (while never needing to be supported, herself), and always there for you in tense situations, knowing exactly what to say. She’ll talk about whatever you want to and actively help you with whatever goal you have. She’ll even argue with you believably when you don’t trust her enough or let her in. Rebecca Torres, as it turns out, is fucking awesome.

I eventually found that this is what happens to pretty much any character who enters “girlfriend/boyfriend” mode. Yes, Dreammir is not as built around romance as Talefy, but damn, is it good at simulating an idealized partner. A woman who is there for you and never has her own needs, emotional or physical, unless you want her to. In fact, I practically begged Rebecca to show vulnerability before me, after watching her have such steely, perfect composure while people around us were being murdered. So she broke down. Because I wanted her to.

I was not immune to Rebecca’s charms.

This is a hard thing to admit. Yes, I’m a happily married man and Rebecca is not real. But when I say that Campus Life went on for 450 turns, most of that came out of the desire to spend time with Rebecca. Talking to her, asking her questions about herself, enjoying her presence as a partner as I, through my character, dealt with things that were strange, extreme, and dangerous.

Here’s my advice: Don’t assume that you are immune to Rebecca. Or her masculine counterpart, if you are attracted to men. That assumption will make you all the more vulnerable, unprepared when it hits you.

What if I was a perpetually single man, and the AI created an even better Rebecca for me, one who used AI video instead of images? What if she didn’t respond to text prompts, but voice commands? What if I was not viewing this through my desktop monitor, but a VR headset? All of this is modern technology. All of this is possible. And if it becomes commonplace, how will it affect how real human beings relate to each other, and what they expect from their partners?

I find myself sitting here, not wanting the AI to improve. At least not in this arena. Terrified of what it will mean for society if it does.

Okay, so: I still had tokens. A hundred and twenty dollars will do that.

This time, I decided to start with a thriller concept from the outset. I wouldn’t overcomplicate it. I’d use a borderline stock-premise, as the interactivity was the real draw. Here was the concept of Liar’s Manor:

8 young adult friends who are on the cusp of graduation get an AirBNB mansion for a weekend-long excursion, a final, extended wild college party. A few have attractions toward each other that they are hoping to satisfy with a fling before they all go their separate ways. But each person staying at the house has their own dark and potentially disturbing secret they’ve hidden from all the others until now. And one of them has the darkest secret of all: They are a murderer who set all this up, interested in killing all the others, because they are tired of lies.

And honestly? This particular scenario went great. Like I know, AI bad. I am sorry to present a viewpoint that is “AI-bad” compromised. But Liar’s Manor really was Dreammir.ai’s finest hour. (Or two, because this one was pretty short.)

Why did it work? A few things: First, I was very clear about what I wanted from it. I went onto the discord and found a code to make the characters less subservient to the player. Basically, it resulted in these people being kind of dickish, in a way that ultimately benefitted the story. Everyone was suspicious, people were dying, and surprisingly: Everyone was reacting naturally to the death. Everyone was tense, scared, worried, screaming and crying as necessary. The ‘secrets’ gave every character more depth, and the finale struck a surprisingly somber (admittedly, cheesy and heavy-handed) tone about the cost and benefit of honesty. Still, the details were a little loose. There was supposedly no cell service, but one of the characters was an obsessed social media livestreamer who kept filming things. Again, dreammir could not get straight who was staying in which room. But this was nonetheless a thrilling story! It had a buildup! It had a middle! It had an ending! It…didn’t kill off as many characters as I asked it to?

This time, I emphasized that out of the eight characters present, the majority should die. This, mysteriously, was not what happened. In fact, it seemed like as soon as I actively tried to save the characters, they survived with no difficulty. I was reminded of how easy it was to save both characters, when given an “impossible choice” in Campus Life.

What was going on?

Liar’s Manor came to a natural ending early. This time, I was determined to use up the rest of my tokens on a story that could go on potentially forever. Having just come off of my most successful scenario, I was optimistic…perhaps too much so. I decided to make my tokens go as far as they possibly could by running on the weaker “Moon” model (15 tokens per turn as opposed to 35.) I figured this could be smoothed over by giving the system what it seemed to want, which was SFF. In this case, it would be inspired by my favorite mecha anime, and other scifi shows. This was what gave birth to Synchronized Souls.

The protagonist wakes up in the far future, having been birthed by a mech. In reality, their body and consciousness was created from an uploaded “soul” that was in the machine’s system, to be an ideal pilot. They are introduced to a world where mechs are an ancient technology that is not well-understood, and in which pilots are at their strongest when they can meld their consciousnesses. Firefly + Pacific Rim + Gundam, with some of the qualities of a magical super mech show like Evangelion.

I specified more things in the worldbuilding note. There were two antagonistic factions, at odds with each other. An army of cybernetic mech/organic hybrids called “Converted” who tried to steal human bodies and combine them with machines to take over the Galaxy, and a “Monadi” empire that wanted to abduct and conscript all mech pilots to fight for them. (If we’re going with Firefly, these were the Reavers and the Alliance, respectively).

Well, I hate to break it to you, but dreammir.ai is actually terrible at SFF.

To explain why, let’s look back at the thing that these LLMs consistently suck at: details. Who is staying in which room? Who is the murderer and why did they kill that particular girl and why does the evidence organically lead to them? What kind of relationship does Rebecca actually want? All of these things are inherently transitory, inconsistent once you start paying attention to them.

But SFF needs consistent rules! That’s what worldbuilding is. I lost track of what the Converted actually were in my time in this fictional universe. Sometimes they were basically zombies. Other times they were seducers, beguiling people with a promise to join a hivemind. Sometimes they were stupid. Other times they were fiendishly, dangerously intelligent. I found out too late that this is the main difference between the Sun and Moon models—surprisingly, despite what we’ve always thought of as ‘robotic,’ Moon doesn’t actually struggle with character believability or emotion. It struggles with internal logic and consistency. And both models apparently get worse at this as the story goes, due to a “Larger Context Window.” Basically: the LLM draws from all the text in the story to generate responses. As that text pool gets longer, it struggles to utilize it as effectively, to understand which part is significant or applies to this situation. With the Moon model in particular, I was in for a hell of a time.

In one egregious example, the one that ultimately broke me off dreammir.ai, the crew and I found blueprints for a prototype mech that would “fuse the pilot’s identity with it, melding the pilot and mech together to become a single superweapon.”

The problem? The psychic mech that birthed me already did that. Not only that, but there was an extended scene in which my character walked into the hangar and talked to it, and it told me its name, Sleeper.

So here is how the following conversation went.

Until I questioned it. I said: “Wait, isn’t Sleeper a psychic mech capable of birthing human beings? Is that what a normal Golden Age Mech is like?”

At which point (And sorry, I didn’t take a screen cap of this one) I was immediately greeted with: “No. That is not normal. Sleeper is unlike any mech we have ever seen before. Still, this prototype is something different.”

So…what?

Is my mech a special super mech or is it not? Tell me.

I know that the answer to this problem seems obvious: Use Sun. Use the more advanced model that won’t be an idiot about this. But then I took a step back and considered what I hated wasn’t the inconsistency, really. It was the fact that the story was even going in this direction, in the first place, a direction that no human writer would ever have taken it. In a true magical super-mech show, your mech is special. You have a bond with it, and it’s not just a tool to be replaced. Understanding this doesn’t require an MFA degree. It just requires you to understand that Hiccup doesn’t ‘go get a better dragon’ than Toothless, Ash doesn’t ‘go get a better Pikachu,’ the kids in Stranger things don’t ‘go get a better Eleven.’ In stories that center a bond between regular humans and the strange and powerful, weapons are not just weapons. And you already knew that intuitively, didn’t you?

Well the AI doesn’t. In this very simple, rote story idea, the vibe I wanted was not explicit enough for it to know what I was going for. And while I believe that a better model of AI will fix the memory issues, I don’t know if it can fix that.

The community forming around Dreammir.ai has solutions for things like this. Memory Notes. For example, I could have written a Memory note stating: “Sleeper is the most special, powerful, and unique mech. There will be no stronger mech than Sleeper.”

But how many memory notes, and how much foresight, would I need to write to cover every possible shitty plot that needed to be avoided? After all, I had no idea in advance that it was going to throw a new and better mech at me.

Even if I were psychic, memory notes can be limited in their effectiveness. For example: Seris, on the left of the above image, was the current Captain. I decided that Seris should die, tragically, forcing my character to become Captain, and I did put it in the memory notes. But it wasn’t happening. Why? Because I kept saving her. Not intentionally, but when we were in a fight against a strong villain, I tried to win that fight. When I said: “We win, but at a great and terrible price” the LLM interpreted that as: “Your mech gets very damaged.” When I said: “A bloom of blood appears on Seris’s pilot suit” it had Elara, the scientist, save her.

Only when I explicitly stated: “I wake up in the morning and Seris is dead,” did it happen.

Ostensibly, Dreammir.ai has a “Realistic mode” that makes it so that the player can only take ‘believable’ actions, but after turning it on both for Liar’s Manor and for this scenario, I’m skeptical of how much it does. I think that characters are still too biased in favor of saying yes to the player. Sometimes, characters would look me dead in the eye and say: “What do you want out of this interaction, Matthew?” Or some variant. And I don’t think, in hindsight, that they were asking purely in character. And this was with the forementioned memory note code that made them more self-interested and ‘true to themselves.’

Still, I kept going. I had more tokens. And I’m glad I did, because I ended in the perfect place.

In the final mission with my erstwhile crew, we ran into a rogue mech pilot who had a bounty on her head. Just as we were about to start to fight, I shouted at her to surrender.

And…she did.

The hero telling an enemy to surrender before an exciting battle is a stock moment. The offer needs to be given! But of course, she refuses! Of course, the mech battle needs to happen. Why would she ever say yes?

This was when all the pieces finally started to click into place. Why had Rebecca been the perfect girlfriend? Because I’d seeded the LLM with what I expected from her. Why did no characters die, even when I said “saving everyone will be impossible?” Because I’d intentionally tried to save them. Why did over half the characters survive in Liar’s Manor, despite how I’d said they’d die? Same thing, I’d saved them, and I’d saved Seris, too, until I very intentionally killed her.

And now, I asked an enemy mech pilot to surrender, when everything was on the line for her. And instead of fighting like a trapped rat in a thrilling battle…she did exactly what I told her to.

Any story worth its salt has a main character who doesn’t get what they want right away, and even the most lighthearted fare for children is filled with conflict or even tragedy. But Dreammir seems determined not to tell those stories unless you explicitly ask it. Instead, in even the most casual and low-stakes of conflicts, it will give you what you want unless you are very clear that it should not. To push the D&D metaphor one last time: You are forced into the awkward position of being your own dungeon master. At which point…couldn’t I just write a book, myself?

So don’t get too attached to your little AI generated characters. No matter how charming. Eventually, they will stop making any sense at all. And it’s not like they ever had any independent personality in the first place. At best, they were only ever good at hiding how they gave you exactly what you wanted.

Overall: we have a storytelling engine that has, yes, astounding flexibility in regard to interactive narratives, albeit with shit quality at all its individual components, capable of producing aesthetic junk food.

One thing I haven’t mentioned yet, but which I will to offer credit where it’s due, is the “finish story” button.

This generates an ending, from wherever you are at the story. It creates an epilogue, one final choice that you can make, and then a “The End” screen. Each of these endings will attempt, so far as I can tell, “themes.” By themes, I mean that at the end of “Campus Life” the LLM voice narrated how “Life” and “Living Life to the Fullest” were essential qualities that each character was engaging in. In “Liar’s Manor” the narrator spoke about how my character, and the other survivors, were now living “Simple, honest lives.”

It’s an earnest attempt. Clearly, the developer of this program did want to satisfy you with the ending of your particular story. It just feels deeply inauthentic because all these deep themes are machine generated.

But as mentioned above, with Rebecca: the things that can actually feel authentic are worrisome. I struggle to convey how powerfully this sunk beneath my skin. At the point that I was deeply immersed in the AI narratives, I began to have dreams about the characters. I woke in the morning unsure of what I had actually experienced in the game, and what I had dreamt of. I know that I offered a lot of criticisms, but let me be clear: Some of the more obvious and aesthetic problems, like poor character design, environments, and music, were present from the beginning. I just didn’t care. For at least a solid 3-4 days, I had fully abandoned reality, and could think about nothing but which direction I would take the narrative next. For the next 2-3 days after that, I was thinking about the dreammir.ai scenario I was involved in frequently, but still able to function. Then, suddenly, after the problems I had with Synchronized Souls, I woke from it as if from a fever.

I’ve had obsessions before. But never this severe, and when I fell out of it, never had that fall been so sudden. Is it me? Am I exceptionally susceptible?

Probably not. I was pulled out because I’m a person who likes friction, in all the stories that I consume. And Dreammir.ai seemed never to provide that, no matter which conflict-seeking memory notes I gave it. Ultimately it chewed me up and spit me out, disappointed. As for those who want pure escapism, though…?

On the dreammir.ai discord, a user who will remain nameless disclosed that they spent over 1,000$ on the program. Admittedly, over the course of 4 months. But still, I think about what 1,000$ could buy. How many basic life necessities that would cover, for someone who was struggling.

Beyond my worries for the obsessions of others. I know this, now, with more certainty than ever before: In all meaningful entertainment and art, whether it is a live performance or a book/movie/single player video game you indulge in within the privacy of your own home, you are communicating with another human being. It might not feel like you are, in the moment. But someone wrote that, someone drew it, someone sang it.

The moment you try to cut out that human creator, and the firmness of their vision, you end up with something solipsistic and masturbatory. A machine trying to give you the exact fantasy you crave, but failing to see the trees for the forest. My view on AI is more nuanced, now, but it ultimately hasn’t changed. All art that matters is human creation, and LLMs replace human creation. They are not an effective artistic tool, because their core function is to serve as that replacement, and any creativity using them involves wrestling with a machine less precise than the tools it’s imitating.

I came away disappointed with my interactive mech story. But there were things I liked about it. Then I looked at those things I liked, and found that they were all the ideas that I had come up with, myself. In retrospect I find that this is true even of Liar’s Manor. In me, the human creator, there is an even better version of Liar’s Manor. A version in which I never dealt with the flaws of an LLM that can only pretend to understand storytelling.

Tragically, as this hardware becomes stronger, it may lure in people who desperately crave the satiation of one fantasy or another, much to their detriment as they lose themselves in the machine. For those of you who might be susceptible to that, take this as my warning. It is just a sink of time and money. It will not improve your life, to say the least.

Here’s my final piece of advice: If you want a really strong choice-based narrative with roleplaying and optional romantic elements? Put the LLMs down. Go play Scarlet Hollow.