I think of this legitimately fun game as an AI game. It's only really possible I think with AI and it did have a big viral moment when it launched: neal.fun/infinite-craft
But ultimately, I think the explanations you suggest are correct. The business model doesn't make much sense for a game, a lot of game players are ideologically opposed to the technology, and AI just isn't that creative or funny.
I also wonder if some of this is because the technology is still relatively new and so people are mostly using it in kind of kitschy, parlor-tricky ways. Maybe the actual answer is to blend deterministic and stochastic together, but that might be something that we have to evolve into over time as AI becomes more "normal" to us.
I'm glad I am not the only one who has been thinking about this a lot. I briefly succumbed to the peer pressure and tried "vibe-coding" some of my numerous strange, abandoned game ideas. It just wasn't as stimulating as it is to learn a game engine, and the results were uncannily inert.
My conclusion differs from yours in one way-- I think that fun **does** come from other people. I believe that the trace of enjoyment left upon the game by its creator is what makes it fun. By contrast, AI-generated games are a desolate wasteland lacking intent, thought or pathos.
Another thought: When we play with a ball or a stick, it is not the lifeless object which provides the fun, it is our own imagination. Perhaps the question is: Can AI leave room for our imagination?
I anticipate this will sound a bit cheeky but don’t mean for it to land like that BUT - I think this is maybe a bit of a category error. Looking for “LLM Games” by trying to identify traditional (deterministic) games that are driven by LLMs feels like a trying to find an apple by looking at oranges. The cases you bring up imply a sort of legibility or signaling of “game-ness” based on proximity to some imagined non-LLM version of the game.
But I think this is maybe missing the point here that what LLMs drive for play won’t necessarily mean they augment existing types of games but instead create new categories. Though also maybe you’re looking just for LLM games, and I think your assessment is right, but I do think there is more here:
The most simple example of this is honestly…. the chat interface. The primary mode of interaction most people have with LLMs through ChatGPT/etc. is a sort of playful Socratic dialogue with a seemingly omniscient intelligence. People ask it silly questions, get it do fun things, etc. More structured things exist in the same vein though as you have people running their own pseudo TTRPG games inside closed ChatGPT sessions with no visibility on steam (obviously). The you have all the romantic/companion roleplaying, etc.
None of this we would really think of as “games” but I think they do really embody a sense of “digital play” that otherwise hasn’t been possible before.
1. I started out with exactly that assumption, that this new kind of gameplay would look nothing like the conventional structures of existing “clockwork” games, and…
2. I have spent a lot of time in open-ended playful exploration with these models, looking for promising new directions to explore.
You could make the case that plain chat, AI boyfriends, and similar activities are “game like” but I think there are important ways in which they are not like games, and the ways they are not like games are the ways that especially matter to me.
I think this is all true while it is also the case that they are STAGGERINGLY like a lot of early text adventure games. I don’t know where that trajectory goes though and if we just reinvent every game breakthrough but do it without determinism.
I think latency is also a bit of a blocker to thinking about design and LLMs - the turn around time from relying on a web request for core logic feels too long to really drive an experience.
To go back to the post though, I do also think “cultural poison” is real and has disincentivized a lot of people from trying stuff for fear of getting obliterated. So stuff happens privately - I’m in some AI gamedev centric discords, which is nice in some ways because you can see people really pushing these systems, but it also feels like a lot of people are starting with it as their first game ever and lack the guiding taste and experience from making games otherwise that would standout as some of the tentpoles you’re looking for. Maybe in a few years though!
Your conclusion 3) is why I've always bristled at the idea. AI-centric games remind me of the kinds of games I wanted to make when I was young, pre-LLM: small interactions with large, inscrutable systems - but in those days those systems were complex simulations. Somehow god games can get away with this under certain circumstances, but for the most part, I think these designs fail because the black box evades learning. And, I have to say it, I think Raph was at least partly very right with the "fun is learning" thing. If I can't feel my mindteeth sink into something and gain some traction, then at best my enjoyment is probably coming more from observation than from direct engagement. That can be something, but it's weirdly so much narrower than most of the pleasure games can offer.
Hi Alex!!! Just driving by and I saw your name on a comment!
But also: using Claude Code to program things feels weirdly like playing a game! The goal of the game is to get the model to accomplish a task, with text-based controls like Zork. And it has that roguelike quality of being different every playthrough, since the task is always different, but the basic controls and mechanics are the same.
I know it's not a game in the traditional sense, but it lights up many of the same parts of my brain. And "playing" it seems to result in me spending less time playing actual games.
I think part of the reason why Infinite Craft works for me as an AI game is that (my understanding is that) the AI only generates a novel response once and then that novel response is saved into a database and always re-used in the future when a user tries to craft something using that specific recipe. That enables more creative and fun gameplay from the human players because there’s something to learn and strategize against, it’s not just completely random each time.
I have once played Death by AI as a party game -- it was available on Discord by default. It was pretty bad, since the AI seemed to decide we would always die no matter what, until I once tried something like "I don't try defending and just accept my fate in a Zen fashion" and it liked that, which in retrospect seems exactly like what a corporate-guardrailed AI would like. The play had the dynamic of exploring the AI itself, and once we understood roughly how it tended to behave, the game became boring.
I think AI games aren't much fun for the same reason playing with AI image generators isn't much fun after a short time: you have infinite freedom but nothing you create means anything, or builds on itself in a complex structure. Then you're soon left with your own lack of creativity.
I think unbounded power isn't conducive to creativity, while constraints help. I'm starting out in an improv comedy group; it's *hard* to do an open scene and (unintuitively) much easier to have there's somebody on the outside pausing the scene and giving you challenges / asking the audience for things to do.
Consider instead something like Minecraft, with unintuitive mechanics and a rigid structure in which you can find something like a challenge. Once there is challenge, things in the world gain meaning -- you might build a huge mob farm, feel like a supervillain building torture farms, then do something with that feeling (embrace it, or instead go build a garden...).
Regarding GOFAI: There's lots of this stuff in Civ-type games, but it tends to be bad in the sense of stupid. Sometimes finding ways to sneak around the AI's limitations can be fun, but as I age it gets much less interesting and I'd rather just play with real humans. There was somebody trying to do something interesting with this concept in AI War: Fleet Command, which embraced the concept of overwhelmingly-powerful-but-exploitable AI and built the entire gameplay around this asymmetry. I haven't played it, though, so can't tell if it succeeded. Here's the design document for that: https://arcengames.com/designing-emergent-ai-part-1-an-introduction/
I like your insight about the relevant distinction between deterministic/mechanical logic vs AI's inherently chaotic elements. Within the domain of "AI opponents" or "Coaches" eg in Chess.com and I'm sure other apps--I'm not sure how much LLM usage is at play here (I've read that LLMs are quite bad at Chess, not at all the same scenario as AlphaGo) but here you could also imagine a better usage of LLMs combined with the more deterministic gameplay AI to liven up coaching sessions, game banter, and help better guide pre-baked coaching options. It seems like the inherent squishiness of AI seems best layered on top of the deterministic elements that are usually at the heart of good games.
This is true for software in general! When products try to rely too heavily on AI, it gets messy and annoying. Sometimes you honestly just need some good old fashioned regex and if/then logic.
Extreme nondeterminism is less fun than we expect it to be because we can't adequately come to understand it, which is a big part of the fun (and leads to the fun of putting a mechanic to use to achieve a goal). Extreme nondeterminism is too much like real life: Sloppy and unfair and unreliable. I think more than anything, we often go to games because they resemble familiar bits of real life, but in a way we can get our brains and arms around, a way that can be made cogent and fair and rewarding in a way life isn't.
When DALL-E first dropped, I was coincidentally reading one of the Love & Rockets omnibi, which has a big 10x10 grid of pictures of every single character in the book, and it occurred to me that you could probably make a game out of it: playing "Guess Who," but with AI-generated portraits in lieu of "does your character have a mustache"-type questions. You could call the game "JAIM-E," ha ha.
Insofar as that would be fun, the fun would come from watching the AI be BAD at generating cartoons, and having to work around its suckiness. (You know the Brian Eno quote.) I am skeptical that this would be fun now that those programs are, like... actually good at generating coherent images.
Yeah, by the same token (har) the game I designed to play with Alexa (Hey Robot) was a lot more fun than any of the similar things I tried playing with ChatGPT.
With most AI being around pattern seeking and regurgitation the next most common possibility, how would this translate to creating realities with rules, goals and obstacles?
I think drastic change is required to get to the next level and it will take a long time still.
#4. LLMs turn out to be surprisingly unreliable at producing structured outputs. This is something e.g. Eddie Hu is contending with in his AI character game, and Mike Treanor's group at American U is researching for games, etc. Problem being encountered all over the place. Supercell just came by and their AI chief pointed at it as a huge block
But really this is just a specific case of #3, it is the same "soft logic" problem at the production stage.
The successors to AI Dungeon now power a lot of the "chatbots," which are probably the most commercially successful entertainment category for LLMS? They are basically the same thing as AI Dungeon (user enters a bunch of context for background -> unpredictable output -> gradually decoheres) but RLHF'd into call and response format, and then with some more contemporary context tricks. But there are also still several companies that are making sustainable niche-audience money on the pure AI Dungeon text completion stuff.
Another way to look at it: using any generative model, image or text or music or whatever, is basically gambling where the payout is "meaningfulness," and what is more valuable to a human than meaning? This is why people keep going back over and over even when they're losing (or lose a whole "AI spouse" to decoherence or model version change) -- the horrible pseudo-grief is just the exciting stomach-drop of a big gambling loss, and then you get back into the game again without having to pay your gambling debts, since it's a subscription model. This also explains why people lose the buzz of trying new models, develop quasi-superstitious beliefs about prompting practices, have exuberance that it can be a little hard for others to understand when they hit a personal "meaning jackpot," etc. It's all meaning-gambling. Also why people overestimate their productivity / time savings by 20-40% when doing some kinds of AI-assisted coding, they get into a variable-reward reinforcement flow state and think they can estimate "amount of work done per hour" like a sprint burndown chart, instead of an expected value of meaning produced over time that also accounts for losses/failures/spaghetti
As much as I hate to say it, meaning-gambling is probably the easiest point to push on for finding more gameplay, and it's already fueling the existing "native" (nearly undesigned) entertainment forms with LLMs
I've been watching Kenta Cho (@abagames) go through a similar reckoning from a different angle (LLMs for generating small, simple, rules-bound browser games). I get the impression that while it's easy to undermine the specificity of your own skill, as designer and player, the preeminence of individuated experience, history, and taste towards the execution of a process, no matter how seemingly straightforward, can't be overstated .
The majority of the playfulness in LLMs seems the same as jack-in-the-boxs, ownership over the set-up is intrinsic to the punchline's value.
This weekend I was poking around a design for an AI-based game and came to a very similar realization. A pattern that kept coming up was that I would start with some hazy mechanic for the interaction and as I tried to beat it into a coherent shape, I would also beat the AI entirely out of the design.
Every time I attempted to answer the question "What would the AI do here that's interesting with this input?", I would have to wrap enough rules around the AI's behavior that I'd end up with a straightforward algorithm - something I could have easily run on an Apple II, no SkyNet required.
My sense is that aside from my lack of design skills, the reason this kept on happening is that a magic box that takes fuzzy inputs and spits out unbounded and unexpected results may be a great toy, but is hopelessly inscrutable. In order for it to be the engine of fun it has to fit inside of a player's head and that player needs to be able to predict its actions. AI fails badly at both of those conditions.
Anyways, thanks for the timely post and helping me make sense of my windmill tilting.
Hi Frank, great to see this piece. I'm Yuqian Sun. I've been building an LLM-based game since 2020 (starting with GPT-2), and have published several papers on this topic. My game 1001 Nights is a story-crafting game where players need to survive an AI king by telling stories. I was fortunate to present the system behind it at GDC last year, and we're currently working toward a full Steam release.
I think this framing might be part of the problem. Many large companies are integrating AI modules into existing games (e.g. NetEase), which researchers and industry insiders often dismiss as "not groundbreaking enough." But some smaller teams have been doing genuinely breakthrough work — it's just that the scale doesn't match expectations.
Take Krafton's Mimesis https://store.steampowered.com/app/2827200/_/ — it's not even LLM-driven, but it's absolutely an “AI-native game”. And from what I've observed, it tends to be small teams willing to take these risks, precisely because established studios have no incentive to gamble when the existing game industry pipeline works fine.
2. On player resistance to AI: This is real for visual AI, but more nuanced than it seems. Language models and non-visual AI tend to be much more accepted by players. For example, AI2U https://store.steampowered.com/app/2880730/_AI2U___AI/ — a room-escape-style game released around the same time as Suck Up — also reached millions of players, and the reception toward its use of LLMs was quite positive.
The nuance here is that when I try to discuss what kinds of AI are "acceptable," everyone has a different opinion. Some believe all generative AI is inherently wrong; others draw the line at visual models because they "steal" from artists, while arguing that language models are fine because they can “explain their sources”. The reality is that public perception of generative AI is still in a chaotic state. In that chaos, the loudest voices tend to express the strongest resistance. But within specific domains — at least between visual and language AI — the distinction in acceptance is actually quite significant.
3. On whether fun AI games exist: I'd argue they do, just not at the scale people expect. Take Infinite Craft, mentioned in other comments here. Millions of people have played it. It does something no previous game could: a generative Doodle God where players can explore endlessly. Isn't that a breakthrough?
The issue isn't that AI can't be fun — it's that we're measuring "breakthrough" against the wrong benchmark. As someone working in this space, I see hidden gems emerging all the time. But due to their scale, their subject matter, or simply a lack of visibility, they end up being judged against a subjective standard of what "breakthrough" means.
I'd say genAI-driven games are becoming something like a subgenre within indie games. Rather than asking for more time, I'd rather stay open-minded and hope more people are willing to experiment. What concerns me is that the current climate of AI resistance is making many creators afraid to share their work. At Gamescom, multiple indie developers told me privately that they were experimenting with LLM-based games but didn't dare publish or even talk about them publicly.
I'm not sure whether this will get better or worse over time. But I keep finding surprises in this space, and I intend to keep looking.
I really hope my comment doesn't derail the discussion because you opened a lot of interesting threads, but I'd really love to open one more.
I'm a big fan of your work, and if I'm not mistaken you have some knowledge of the Chinese tech scene. Do you think that on these questions there is a divide between the Chinese speaking world and English speaking world?
From my limited observation the AI backlash is a very western thing (for instance I see games like Doki Doki AI Interrogation or Deduce Together from Japan getting moderate success).
I'd love to know more about interesting games, institutions, creators, artists, etc... operating in China because it's a bit opaque for non-Chinese speakers apart from big AAA releases.
Do feel free to delete this if it derails the conversation but I think the contrast between different cultures might be enlightening to the wider discussion.
great article! But to be fair it's not very mysterious, in the middle of a massive witch hunt against AI in the west, all the talented creatives that could have used AI for interesting things just don't. The subscription based pricing is the nail in the coffin.
It's a shame because it's a really powerful and potent technology, I hope that someday the West politics will evolve, or that we get cool games out of China.
I think of this legitimately fun game as an AI game. It's only really possible I think with AI and it did have a big viral moment when it launched: neal.fun/infinite-craft
But ultimately, I think the explanations you suggest are correct. The business model doesn't make much sense for a game, a lot of game players are ideologically opposed to the technology, and AI just isn't that creative or funny.
I also wonder if some of this is because the technology is still relatively new and so people are mostly using it in kind of kitschy, parlor-tricky ways. Maybe the actual answer is to blend deterministic and stochastic together, but that might be something that we have to evolve into over time as AI becomes more "normal" to us.
I'm glad I am not the only one who has been thinking about this a lot. I briefly succumbed to the peer pressure and tried "vibe-coding" some of my numerous strange, abandoned game ideas. It just wasn't as stimulating as it is to learn a game engine, and the results were uncannily inert.
My conclusion differs from yours in one way-- I think that fun **does** come from other people. I believe that the trace of enjoyment left upon the game by its creator is what makes it fun. By contrast, AI-generated games are a desolate wasteland lacking intent, thought or pathos.
Another thought: When we play with a ball or a stick, it is not the lifeless object which provides the fun, it is our own imagination. Perhaps the question is: Can AI leave room for our imagination?
I anticipate this will sound a bit cheeky but don’t mean for it to land like that BUT - I think this is maybe a bit of a category error. Looking for “LLM Games” by trying to identify traditional (deterministic) games that are driven by LLMs feels like a trying to find an apple by looking at oranges. The cases you bring up imply a sort of legibility or signaling of “game-ness” based on proximity to some imagined non-LLM version of the game.
But I think this is maybe missing the point here that what LLMs drive for play won’t necessarily mean they augment existing types of games but instead create new categories. Though also maybe you’re looking just for LLM games, and I think your assessment is right, but I do think there is more here:
The most simple example of this is honestly…. the chat interface. The primary mode of interaction most people have with LLMs through ChatGPT/etc. is a sort of playful Socratic dialogue with a seemingly omniscient intelligence. People ask it silly questions, get it do fun things, etc. More structured things exist in the same vein though as you have people running their own pseudo TTRPG games inside closed ChatGPT sessions with no visibility on steam (obviously). The you have all the romantic/companion roleplaying, etc.
None of this we would really think of as “games” but I think they do really embody a sense of “digital play” that otherwise hasn’t been possible before.
I know what you mean, but keep in mind:
1. I started out with exactly that assumption, that this new kind of gameplay would look nothing like the conventional structures of existing “clockwork” games, and…
2. I have spent a lot of time in open-ended playful exploration with these models, looking for promising new directions to explore.
You could make the case that plain chat, AI boyfriends, and similar activities are “game like” but I think there are important ways in which they are not like games, and the ways they are not like games are the ways that especially matter to me.
I think this is all true while it is also the case that they are STAGGERINGLY like a lot of early text adventure games. I don’t know where that trajectory goes though and if we just reinvent every game breakthrough but do it without determinism.
I think latency is also a bit of a blocker to thinking about design and LLMs - the turn around time from relying on a web request for core logic feels too long to really drive an experience.
To go back to the post though, I do also think “cultural poison” is real and has disincentivized a lot of people from trying stuff for fear of getting obliterated. So stuff happens privately - I’m in some AI gamedev centric discords, which is nice in some ways because you can see people really pushing these systems, but it also feels like a lot of people are starting with it as their first game ever and lack the guiding taste and experience from making games otherwise that would standout as some of the tentpoles you’re looking for. Maybe in a few years though!
Your conclusion 3) is why I've always bristled at the idea. AI-centric games remind me of the kinds of games I wanted to make when I was young, pre-LLM: small interactions with large, inscrutable systems - but in those days those systems were complex simulations. Somehow god games can get away with this under certain circumstances, but for the most part, I think these designs fail because the black box evades learning. And, I have to say it, I think Raph was at least partly very right with the "fun is learning" thing. If I can't feel my mindteeth sink into something and gain some traction, then at best my enjoyment is probably coming more from observation than from direct engagement. That can be something, but it's weirdly so much narrower than most of the pleasure games can offer.
Hi Alex!!! Just driving by and I saw your name on a comment!
But also: using Claude Code to program things feels weirdly like playing a game! The goal of the game is to get the model to accomplish a task, with text-based controls like Zork. And it has that roguelike quality of being different every playthrough, since the task is always different, but the basic controls and mechanics are the same.
I know it's not a game in the traditional sense, but it lights up many of the same parts of my brain. And "playing" it seems to result in me spending less time playing actual games.
I think part of the reason why Infinite Craft works for me as an AI game is that (my understanding is that) the AI only generates a novel response once and then that novel response is saved into a database and always re-used in the future when a user tries to craft something using that specific recipe. That enables more creative and fun gameplay from the human players because there’s something to learn and strategize against, it’s not just completely random each time.
I have once played Death by AI as a party game -- it was available on Discord by default. It was pretty bad, since the AI seemed to decide we would always die no matter what, until I once tried something like "I don't try defending and just accept my fate in a Zen fashion" and it liked that, which in retrospect seems exactly like what a corporate-guardrailed AI would like. The play had the dynamic of exploring the AI itself, and once we understood roughly how it tended to behave, the game became boring.
I think AI games aren't much fun for the same reason playing with AI image generators isn't much fun after a short time: you have infinite freedom but nothing you create means anything, or builds on itself in a complex structure. Then you're soon left with your own lack of creativity.
I think unbounded power isn't conducive to creativity, while constraints help. I'm starting out in an improv comedy group; it's *hard* to do an open scene and (unintuitively) much easier to have there's somebody on the outside pausing the scene and giving you challenges / asking the audience for things to do.
Consider instead something like Minecraft, with unintuitive mechanics and a rigid structure in which you can find something like a challenge. Once there is challenge, things in the world gain meaning -- you might build a huge mob farm, feel like a supervillain building torture farms, then do something with that feeling (embrace it, or instead go build a garden...).
Regarding GOFAI: There's lots of this stuff in Civ-type games, but it tends to be bad in the sense of stupid. Sometimes finding ways to sneak around the AI's limitations can be fun, but as I age it gets much less interesting and I'd rather just play with real humans. There was somebody trying to do something interesting with this concept in AI War: Fleet Command, which embraced the concept of overwhelmingly-powerful-but-exploitable AI and built the entire gameplay around this asymmetry. I haven't played it, though, so can't tell if it succeeded. Here's the design document for that: https://arcengames.com/designing-emergent-ai-part-1-an-introduction/
I like your insight about the relevant distinction between deterministic/mechanical logic vs AI's inherently chaotic elements. Within the domain of "AI opponents" or "Coaches" eg in Chess.com and I'm sure other apps--I'm not sure how much LLM usage is at play here (I've read that LLMs are quite bad at Chess, not at all the same scenario as AlphaGo) but here you could also imagine a better usage of LLMs combined with the more deterministic gameplay AI to liven up coaching sessions, game banter, and help better guide pre-baked coaching options. It seems like the inherent squishiness of AI seems best layered on top of the deterministic elements that are usually at the heart of good games.
This is true for software in general! When products try to rely too heavily on AI, it gets messy and annoying. Sometimes you honestly just need some good old fashioned regex and if/then logic.
Perhaps apprehending the LLM itself is the game.
Extreme nondeterminism is less fun than we expect it to be because we can't adequately come to understand it, which is a big part of the fun (and leads to the fun of putting a mechanic to use to achieve a goal). Extreme nondeterminism is too much like real life: Sloppy and unfair and unreliable. I think more than anything, we often go to games because they resemble familiar bits of real life, but in a way we can get our brains and arms around, a way that can be made cogent and fair and rewarding in a way life isn't.
When DALL-E first dropped, I was coincidentally reading one of the Love & Rockets omnibi, which has a big 10x10 grid of pictures of every single character in the book, and it occurred to me that you could probably make a game out of it: playing "Guess Who," but with AI-generated portraits in lieu of "does your character have a mustache"-type questions. You could call the game "JAIM-E," ha ha.
Insofar as that would be fun, the fun would come from watching the AI be BAD at generating cartoons, and having to work around its suckiness. (You know the Brian Eno quote.) I am skeptical that this would be fun now that those programs are, like... actually good at generating coherent images.
Yeah, by the same token (har) the game I designed to play with Alexa (Hey Robot) was a lot more fun than any of the similar things I tried playing with ChatGPT.
Completely agree and agree on your reasoning in the "You know what's fun?" paragraph. Well said!
With most AI being around pattern seeking and regurgitation the next most common possibility, how would this translate to creating realities with rules, goals and obstacles?
I think drastic change is required to get to the next level and it will take a long time still.
#4. LLMs turn out to be surprisingly unreliable at producing structured outputs. This is something e.g. Eddie Hu is contending with in his AI character game, and Mike Treanor's group at American U is researching for games, etc. Problem being encountered all over the place. Supercell just came by and their AI chief pointed at it as a huge block
But really this is just a specific case of #3, it is the same "soft logic" problem at the production stage.
The successors to AI Dungeon now power a lot of the "chatbots," which are probably the most commercially successful entertainment category for LLMS? They are basically the same thing as AI Dungeon (user enters a bunch of context for background -> unpredictable output -> gradually decoheres) but RLHF'd into call and response format, and then with some more contemporary context tricks. But there are also still several companies that are making sustainable niche-audience money on the pure AI Dungeon text completion stuff.
Another way to look at it: using any generative model, image or text or music or whatever, is basically gambling where the payout is "meaningfulness," and what is more valuable to a human than meaning? This is why people keep going back over and over even when they're losing (or lose a whole "AI spouse" to decoherence or model version change) -- the horrible pseudo-grief is just the exciting stomach-drop of a big gambling loss, and then you get back into the game again without having to pay your gambling debts, since it's a subscription model. This also explains why people lose the buzz of trying new models, develop quasi-superstitious beliefs about prompting practices, have exuberance that it can be a little hard for others to understand when they hit a personal "meaning jackpot," etc. It's all meaning-gambling. Also why people overestimate their productivity / time savings by 20-40% when doing some kinds of AI-assisted coding, they get into a variable-reward reinforcement flow state and think they can estimate "amount of work done per hour" like a sprint burndown chart, instead of an expected value of meaning produced over time that also accounts for losses/failures/spaghetti
As much as I hate to say it, meaning-gambling is probably the easiest point to push on for finding more gameplay, and it's already fueling the existing "native" (nearly undesigned) entertainment forms with LLMs
I've been watching Kenta Cho (@abagames) go through a similar reckoning from a different angle (LLMs for generating small, simple, rules-bound browser games). I get the impression that while it's easy to undermine the specificity of your own skill, as designer and player, the preeminence of individuated experience, history, and taste towards the execution of a process, no matter how seemingly straightforward, can't be overstated .
The majority of the playfulness in LLMs seems the same as jack-in-the-boxs, ownership over the set-up is intrinsic to the punchline's value.
“ownership over the set-up is intrinsic to the punchline's value” ! 💡
This weekend I was poking around a design for an AI-based game and came to a very similar realization. A pattern that kept coming up was that I would start with some hazy mechanic for the interaction and as I tried to beat it into a coherent shape, I would also beat the AI entirely out of the design.
Every time I attempted to answer the question "What would the AI do here that's interesting with this input?", I would have to wrap enough rules around the AI's behavior that I'd end up with a straightforward algorithm - something I could have easily run on an Apple II, no SkyNet required.
My sense is that aside from my lack of design skills, the reason this kept on happening is that a magic box that takes fuzzy inputs and spits out unbounded and unexpected results may be a great toy, but is hopelessly inscrutable. In order for it to be the engine of fun it has to fit inside of a player's head and that player needs to be able to predict its actions. AI fails badly at both of those conditions.
Anyways, thanks for the timely post and helping me make sense of my windmill tilting.
Hi Frank, great to see this piece. I'm Yuqian Sun. I've been building an LLM-based game since 2020 (starting with GPT-2), and have published several papers on this topic. My game 1001 Nights is a story-crafting game where players need to survive an AI king by telling stories. I was fortunate to present the system behind it at GDC last year, and we're currently working toward a full Steam release.
Here's a playable demo if you're curious: https://store.steampowered.com/app/2782660/1001_Nights_Demo/
Here is our latest paper: https://dl.acm.org/doi/full/10.1145/3706599.3721163
I'd love to share a few thoughts and questions
1. What counts as "groundbreaking"?
I think this framing might be part of the problem. Many large companies are integrating AI modules into existing games (e.g. NetEase), which researchers and industry insiders often dismiss as "not groundbreaking enough." But some smaller teams have been doing genuinely breakthrough work — it's just that the scale doesn't match expectations.
Take Krafton's Mimesis https://store.steampowered.com/app/2827200/_/ — it's not even LLM-driven, but it's absolutely an “AI-native game”. And from what I've observed, it tends to be small teams willing to take these risks, precisely because established studios have no incentive to gamble when the existing game industry pipeline works fine.
2. On player resistance to AI: This is real for visual AI, but more nuanced than it seems. Language models and non-visual AI tend to be much more accepted by players. For example, AI2U https://store.steampowered.com/app/2880730/_AI2U___AI/ — a room-escape-style game released around the same time as Suck Up — also reached millions of players, and the reception toward its use of LLMs was quite positive.
The nuance here is that when I try to discuss what kinds of AI are "acceptable," everyone has a different opinion. Some believe all generative AI is inherently wrong; others draw the line at visual models because they "steal" from artists, while arguing that language models are fine because they can “explain their sources”. The reality is that public perception of generative AI is still in a chaotic state. In that chaos, the loudest voices tend to express the strongest resistance. But within specific domains — at least between visual and language AI — the distinction in acceptance is actually quite significant.
3. On whether fun AI games exist: I'd argue they do, just not at the scale people expect. Take Infinite Craft, mentioned in other comments here. Millions of people have played it. It does something no previous game could: a generative Doodle God where players can explore endlessly. Isn't that a breakthrough?
The issue isn't that AI can't be fun — it's that we're measuring "breakthrough" against the wrong benchmark. As someone working in this space, I see hidden gems emerging all the time. But due to their scale, their subject matter, or simply a lack of visibility, they end up being judged against a subjective standard of what "breakthrough" means.
I'd say genAI-driven games are becoming something like a subgenre within indie games. Rather than asking for more time, I'd rather stay open-minded and hope more people are willing to experiment. What concerns me is that the current climate of AI resistance is making many creators afraid to share their work. At Gamescom, multiple indie developers told me privately that they were experimenting with LLM-based games but didn't dare publish or even talk about them publicly.
I'm not sure whether this will get better or worse over time. But I keep finding surprises in this space, and I intend to keep looking.
Would love to discuss further.
I really hope my comment doesn't derail the discussion because you opened a lot of interesting threads, but I'd really love to open one more.
I'm a big fan of your work, and if I'm not mistaken you have some knowledge of the Chinese tech scene. Do you think that on these questions there is a divide between the Chinese speaking world and English speaking world?
From my limited observation the AI backlash is a very western thing (for instance I see games like Doki Doki AI Interrogation or Deduce Together from Japan getting moderate success).
I'd love to know more about interesting games, institutions, creators, artists, etc... operating in China because it's a bit opaque for non-Chinese speakers apart from big AAA releases.
Do feel free to delete this if it derails the conversation but I think the contrast between different cultures might be enlightening to the wider discussion.
great article! But to be fair it's not very mysterious, in the middle of a massive witch hunt against AI in the west, all the talented creatives that could have used AI for interesting things just don't. The subscription based pricing is the nail in the coffin.
It's a shame because it's a really powerful and potent technology, I hope that someday the West politics will evolve, or that we get cool games out of China.
AI lacks context coherency. What game lacks context?