Why No AI Games?

> You know what’s fun? A stick. A stick is fun. A ball is fun.

Having a body is fun. I think that's one reason why VR has such quick hype/death cycles--it doesn't do a good enough job of fooling your body. Conventional games induce more like a dissociative or hypnotic state where you temporarily forget your body. That can range from very, VERY abstract (like Pong or Pac-Man, or BABA IS YOU), or built on an attempt to simulate the real world as convincingly as possible through high-end graphics and physics engines.

One of the things that made Untitled Goose Game so much fun for me was that playing it made me _feel like a goose_. It made me want to run around doing goose things for goose reasons. You can spread your wings and honk, regardless of whether it advances the game. A similar game that came out called Little Kitty, Big City offers the promise of the same idea but as a cat instead of a goose. I tried that game but never felt like a cat playing it, instead it felt like being a person controlling a cat. These are such subtle shades of gameplay and storytelling that I have a hard time imagining LLMs being useful in the design.

Only because it is something i find fascinating: > There are those Orcs in that one Lord of the Rings game who hold grudges against you.

Is referring to the nemesis system in Middle-Earth: Shadow of Mordor and Shadow of War, and its an amazing set of interlocking procedural systems that do genuinely feel like its AI, but is really AI in the sense its always been used by games (the rules the games follow to govern NPCs+world) and not AI in the sense of modern LLMs or even other generative systems. This video is a great look at what it is and why its great IMO https://www.youtube.com/watch?v=Lm_AzK27mZY

I think a system like this could really work well with some modern LLM stuff, but it certainly feels magic without it.

I think AI could reinvent text based games in a fresh way that would be fun. Text based games were always so constrained by what the developer could guess that you would type but with AI it really could read more like an interactive choose your own adventure novel that is easy to navigate.

Patrick McKenzie, aka patio11, used a LLM to generate and run a DnD campaign for a gaming convention/conference, He discusses the process and results as part of this podcast on the con.

https://www.complexsystemspodcast.com/episodes/narrative-mas...

I've spent way too many years chipping away at genai gaming, here's my 2c.

First, the intersection between someone who understands and cares about genai (and keeping up with the SOTA), with someone who wants to put in the work and compromise to make great games, is slim.

Think about the intersection between a procedural generation/shader programmer, someone focused on making a really fun and compelling game tech be damned. The skills don't really overlap. One requires deep knowledge, technical mastery, geometric thinking, algorithmic optimization, and time in a dark room, and the other requires empathy for humans, brutal iteration, artistic taste, and a knack for storytelling (even if just to sell the player on your game). You basically need both for your "ai game" to even qualify as an ai game. And in an industry that's so anti-ai it's hard to find teams where these two personalities meet and vibe enough to work on a game together.

Secondly, is model performance: quality, speed, cost. I don't think the industry has yet crossed the threshold where generation is high enough quality, fast enough for players, and cheap enough that it amortizes compared to traditional gamedev. You see tons of AI used in development but deploying this stuff to the front lines doesn't really napkin math. That will change though.

Third, game engines are NOT designed for genai. They are decades of optimization for a completely different pipeline, both technical and human: pitch -> publisher -> game engine -> art team + coders -> trailer -> marketing -> release -> support. Adding even simple very, very basic genai requirements to this (say, dynamic texture generation and loading over the network, streaming character dialogue) completely breaks decades of assumptions of how games are built and shipped (e.g. everything mega-super-duper-compressed and loaded to statically allocated gpu buffers). So doing anything new means throwing away large parts of the game engine. And if you pitch not using 70% of the game engine, or making your own, to a 20 year game veteran/studio boss calling the shots in this industry: well, your game will not be getting funded.

This will all get fixed though. The technology jsut came faster than the industry could adapt.

Looking forward to GDC this year!

They're are plenty of mods that inject AI into games, like Skyrim. Though in order for a game to have LLMs and scale to a shipping audience without a subscription, you would want to run them locally.

The real technical blocker is performance and voice synthesis. If voice synthesis was at parity with human actors, it would be worth it to battle the performance aspect for major studios. In text based games especially, taking time to generate enough text is just too slow to be convenient

> Maybe there are some cool AI-centric games that I’m overlooking?

The enemy locomotion in Arc Raiders was almost entirely created with reinforcement learning. It’s a very impressive modern example (the game has been out for less than 6 months).

Here’s the documentary explaining how the RL locomotion works. Skip to 10 min 36 sec. https://www.youtube.com/watch?v=DRlhpzc7ImA&t=636s

It’s really weird that this article walks past the 2 likely explanations and then declares them both insufficient in favor of an assertion that maybe it’s not any fun.

Speculative generation is expensive and time consuming and in most cases will just be the game company writing a check to a provider every time the player does something. It’s very difficult to imagine a revenue model which allows that to make sense. Even if you did get it to make sense, you then have to worry about their being a market risk to people associating your game with AI. I don’t think you need more than those two explanations to understand why in two or three years we have not seen games that the author describes come out.

Is it a possibility that talking to a chat bot is not any fun from a games per perspective? Yes it might even be very likely. But we’re not gonna have a real answer to that question tested out by real game developers until it becomes pragmatically possible to ship a game using these tools.

The cost of generation will come down and people will find clever uses for it and one way or another opinions about AI will change. Then maybe we’ll see whether or not these are any fun.

This has been discussed ad nauseum in other places here over the years, but basically I don't think the conclusions in the post are correct.

It does not make a whole lot of sense to let an AI determine story, whether it be dialogue, changing the game state, or whatever. The reason basically being a few things - one, the fundamental aspect of a game is that there are rules and boundaries these rules need to remain consistent (and testable). It is entirely jarring if an AI NPC that says something that's not consistent with the game state, or changes the game state in a way that violates the constraint of the understood rules/boundaries of the game - this isn't fun to the user at all, even if it sounds cool. If you do not believe me, you can test this out trying to play DND with these things and see how annoying it can become. If you decide ok, I'm going to try to tightly bind what the AI can do or not with branching and rules, you are basically engineering the same way that games have engineered already, so why even use AI? It's a solution looking for a problem here.

The second big issue is determinism - LLM's are fundamentally non-deterministic. In most games, you would expect an action to have the same or at least very similar reaction most of the time - RNG comes into play naturally already in a lot of games in a predictable way (dice rolls, chance to hit, etc.) LLM's bring nothing new to the table here.

For world gen, we already have games like No Man's Sky deterministically generate quadrillions of worlds. What help is an LLM here? We already have the technology.

One area that would be interesting, is agentic bot players, but that sort of leads down the same path as the above arguments - there already are extremely sophisticated bots that play a huge variety of games. What do LLM's bring here?

The costs of interactive AI have interesting effects as the author points out. Much like the lack of variety in music models, 3D asset generation via AI has a long way to go, particularly as studios have no incentive to share their data. But I think AI assistance could at least make some marginal improvements. Take a procedural game like No Man's Sky. There are billions of possible worlds in the game. But, presumably because of team size and/or agile philosophy, they drop these rigid, nearly identical, items and buildings throughout those billions of planets. AI could at least conveniently empower some added incremental complexity (via Claude Code etc.) to their existing asset generation, which lends more believability to these planets. If studios were willing to partner and share data, incredible world generation models, which could be used exclusively as asset generators rather than real time world renderers initially, could be built and dramatically empower designers.

In analogy to suspension of disbelief I'd like to postulate suspension of genericity:

when I sink countless hours skimming text content in some The Elder Scrolls I'm not really convinced that much of it is a story worth consuming (my opinion of TES is lower than my number of gameplay hours might suggest), but I'm at least open to the idea that it might be. Thousands of players have paid good money for consuming those stories, do who knows, I might actually find something meaningful, perhaps something I'll one day connect with a real person about, so it's not necessarily a given that it's all just filler.

But when it's an LLM generating the story for me and only for me, there's no other possibility. It's a hamster wheel time sink and I can't even pretend that it's not.

For games, interacting with an LLM is an improv experience. It’s a fractal that can be explored in any direction.

Video games have discrete, static goals that let a player focus on an objective. Compared to LLMs, it’s a passive experience.

People play games for all sorts of reasons (to relax, competition, to build something of their own, solve challenges).

I think this is a fundamentally different experience than what an LLM can offer.

That’s not to say LLMs can’t become a fun experience, but it’s going to take decades to develop a way to procure that experience. Look at how long it took dungeons and dragons, or any video game genre, to get to the level of polish it’s at today

It could be about cost: big LLMs don’t work locally like most games, and even with server hosting, the context for a well-detailed and non-textual game would grow very fast, so processing it would become slow and/or expensive. These are the current limits for truly generative games.

As for generated content in regular games, I don’t see an issue if the content is high quality and free of errors. People don’t like low-quality content regardless of who generated it - human or LLM. It’s just that there’s currently more bad content coming from LLMs, that’s all

One thing I've learned is, there's a ton of interesting stuff in the world. Super talented people. But, what surface and gets widespread recognition and adoption is often very much determined by luck, and these days by algorithms. It's very possible there's some really cool AI games out there already, and they just haven't been discovered.

The other aspect is that a lot of people in the AI community are completely hooked on the reward-seeking of 'new thing'. Every hour there's some new tool to try out that 'changes everything'. But, without grinding it out you might never create something interesting by constantly jumping to the next new thing. Because, technological capabilities are just one part of the story.

Third, what makes something 'fun' often requires tons of iteration and a lot of LLM-type tools still rely on snapshots of applications. They're not fully there yet in terms of what happens in between moments. There are lots of principles they can draw on from and get a good prototype quickly, but it again comes back to really just slogging through lots of iterations until you find something fun.

Finally, I think with every new medium you get a first wave that really just copies the medium that came before it. Early radio was a lot like theater. Early TV was a lot like radio. I think eventually we'll see people – probably younger people – who are fully into AI and will find interesting things about that medium to express in ideas, and we'll get some really new and interesting stuff as a result.

I did a conferencr talk maybe three years ago on using AI for games.

Since the talk, code, audio,and imagery have progressed leaps and bounds, almost all of the points on active AI use haven't changed much.

You have more capable smaller models now that means you could at least run local models in games, but you still don't get to see what it is going to do before the user does. Developers are acutely aware of the climate. if a game has a fault because the AI does something unexpected, it's not just a bug, It's a news story.

It still cannot be realistically used for multiplayer games. The models are prone to adversarial attacks, which can harm a multiplayer experience.

I did point out at the time that solo sandbox games could benefit. An active AI can be empowering for the player, unchecked empowerment can be great in a solo game. It destroys a multiplayer game.

Then there's just a lack of training in some of the specific areas of games. You pretty much have to filter game state down to text for a llm, I don't yet know of any gameplay embedding model.

I have played around with a little tower defence to see how much an AI can work on gameplay, It essentially has a REPL interface to the game logic where it can take actions and advance the game a tick at a time. It does an ok job, but there's still not much understanding of time and urgency to work with an active environment

Having recently considered playing Dwarf Fortress(and quickly rejecting it for UI issues and overall tediousness), it seemed like a game that would benefit from an LLM backend. I'm curious what sorts of world history an LLM could come up with. It likely would be a lot less efficient than the current system, but interesting nonetheless. I think it would be easy for devs/modders to add a way to supply an API key to an LLM and experiment there.

Also having recently played a run of Civ5(before burning out because the late game is tedious and the overall game so unrealistic), I thought that good AI would help 4x games remain epic while letting the user choose their level of tedium. It would be nice, for instance, to manually go through the early game when management is easy and exploration is fun, only to turn that over to an AI at some point(and you generally do with autoexploration). Same thing with other eras though. By the time I'm in the late game and have a large empire, I'd rather focus on diplomacy and moving armies, not city management or worker management. I don't want to commit to fully giving up on those aspects of the game though. It would be nice to, say, automate war mostly but be able to jump in if I want to micromanage for a few turns.

I'm making one as a hobby project alongside my day job.

It's a debate chat game where two humans debate on a topic and a biased AI judge picks a winner. Over time players compete for ELO rankings like in chess or similar head-to-head competitive games.

I'll probably make a blog post someday about what's been involved but I think the points in OP are spot on. I've been struggling to make the format feel fun and lively. Two things really helped compared to my early versions - real-time streaming of the other person's arguments so you're not just staring at a timer while they're writing and very very short rounds so you get feedback and results quickly.

The costs are a bit high but not terrible for a hobby project that no one really plays. My unit cost is around $0.008 per game if an NPC is involved and around $0.004 if it's between two humans. I could probably cover that with really intrusive ads but I'm hoping to get enough users paying for cosmetics (speech bubbles, profile pictures, etc.) to help subsidize the rest of the platform.

I've also tried a version with AI NPC opponents and it's much less fun than playing against real humans for reasons I can't entirely articulate. The AIs just don't make for very interesting debaters and the wins against them feel a bit shallow somehow. Right now, I've got NPCs on the game which you can opt into debating if you spend more than 30 seconds in the matchmaking queue but I hope once I get proper player counts I can disable them entirely.

I'm super biased, but proud of how the game has been coming along. Will make a show hn post someday but if anyone's curious my alpha version is live here: https://letsarg.com

I think this is a very interesting question.

But I think getting good AI features into a game requires a very very high engineering effort that is easily underestimated.

Current game engines are very good at creating large worlds with interactable assets: "Pull a lever and the door opens", "enter a building and a preprogrammed event starts", etc.

They are also good at physics simulations.

But ultimately, most of the content is hand-crafted/"hardcoded". And the systems in systemic games are very simple. (With a few exceptions like Dwarf Fortress).

In order to to - for example - let an AI introduce an ad-hoc questline, a lot would need to be formalized.

Imagine the following scenario: When I was playing Cyberpunk 2077 I hoped that I could roleplay as a corpo, slowly going up the ranks, backstabbing, etc. But the sheer amount of social and economic simulation this kind of openness would require is huge.

And it does not help that the work of writers is undervalued in the industry. And also that most creatives are in the anti-AI camp.

I think there are few plausible reasons not mentioned.

We are still early in the tech. 2021 LLMs were not fit for any real purpose outside zany Madlibs. The tech has only really gotten fit for purpose since GPT4 and coherent models only got cheap and fast enough since llama 2 (July 2023). Even then smart models only got affordable around Deepseek v2 (May 2024) which was the first gpt4 level model to consistently serve subdollar per million tokens. You need at least a GPT4 level model to make a really interesting game. That's like a blink of time in game making time, indie games have 2-3 year cycles at best and AAA are like 5+. Even now fast, tool native interaction is only just coming online and there are no cheap models for that. Really fun game playing AI needs something like that to be magical.

There are AI games they just don't look like AI games neal.fun's Infinite Craft made a boat load of money and was very popular and it was powered by a Llama backend. Character ai and it's dozen copycats are like online storytelling/roleplay things. These are fairly popular. That could be a game or like a game platform or maybe it's more like fanfic. It has no fail state but people have made variants with fairly complex rules and states. Skryim and a few other games have pretty popular AI mods that let you add in NPC interactions that can talk, see, and even change their interactions with AI. The Skryim one Mantella has more than ~100k downloads.

We don't really have a northstar yet. Making a good game is really hard and usually takes someone doing something weird and clever. Minecraft is obvious in hindsight but many games on the way failed to make 'legos on the computer' fun. Incorporating AI breaks all the 'rules' of gaming. It operates slowly, it has high potential to break the rules and it has nearly unrecoverable failure states. Indie devs haven't figured out the best way to handle this and I suspect most big companies have just washed their hands of the whole thing until the AI gets faster, cheaper, and more stable.

At Rec Room, we made tons of AI games. The problem is that they weren't economical. You need to make a ton of money if your players are burning tokens: https://www.tyleo.com/professional-work-history/rec-room-ugc...

I think folks are having a lot better luck on the tools side. Some interesting things are happening in the AI Game Dev Org Discord: https://www.tyleo.com/blog/the-excommunicated-devs-making-ga...

I think the biggest issue with any sort of AI games is the fact most of them are so open ended, the buck ends with your own creativity. When you are the one in control of so much, you quickly get tired when you run out of the story you want to tell.

I have a bit of hope that some sort of proper setup will fix this issue, especially with multiple agents able to pass json to and from the "game" side and the writer side. One of my ideas is an open-ended roguelike card game where the player runs through a bunch of player-created scenarios and builds their deck by taking things from that world to the next, so people can compare how their build influenced their version of the scenario.

Creator of AI Game Master - talk to me

(iOS - https://apps.apple.com/us/app/ai-game-master-dungeon-rpg/id6...) (Android - https://play.google.com/store/apps/details?id=com.aisuccess....)

This is roughly my 5th startup, going through Computer Vision -> Special Needs -> Cannatech -> Hacking RNA -> AI -> Gaming with AI

AMA :)

Compelled to point out there was a wave of "AI games" before proper generative models too: the Kinect used decision forests for pose estimation [1], but in an example of the "AI effect", the Kinect games are mostly remembered as cumbersome and largely dancing-themed, not specifically AI/ML games

[1] https://www.microsoft.com/en-us/research/blog/kinect-body-tr...

A few years ago when the first diffusion models for image generation came out, we built a Dixit-like party game where you were supposed to guess what was the prompt that generated the image. Back then the image generation models were still quite crappy and that was the main source of fun in the game. You never knew how many limbs you were going to get.

I was pretty surprised to see that the service we set up still works! https://impromptu.fun/ If you'd like to check it out.

And to separate my thoughts from the info blob:

i think the culture war point is also super true of the game design industry, not just the consumers, where the already ultra competitive nature of the work means that the creatives and the industry as a whole have taken a veeeery strong stance against genai. Thats a reckon, and i dont know if its good or bad.

It does feel a little counter to the march of progress, but in a medium where high effort can be enjoyed by many, im personally cool with artisinal handmade games.

I agree that AI is not as good as people at being entertaining. And everybody looked at putting LLMs in their games as GDC circa 2022 or so. It seemed like a good idea until you realized the server cost or the performance hit. So this is right.

But.. as someone who works in the business of games.. how sure are you that AI hasn't been there all along in ways you may not realize? Companies are very disincentivized to let you know when they are using it.

I think the biggest potential at the moment for AI in games is by introducing new mechanisms for interacting with them. I'm not 100% if 'Before your eyes', a game centred around the players blinking, uses AI but similar technologies for eye-tracking, emotion detection, real-time audio transcription and of course semantic analysis are ripe with interesting gameplay mechanics.

Because AI is not good or interesting.

I've said this before: In Mass Effect, characters are interesting because they are unique, principled, not always right, and are tested. Garrus for example is a former law enforcement official who got upsetty spaghetti about how sometimes the law fails to catch a criminal so he decides to start murdering them instead. In the course of several missions, you as the player personally experience this dynamic, hearing his story, seeing how the failure of real justice impacts people, talk with Garrus about his beliefs and how they changed, you can critique his position and point out ideals or whatever, and then after an hour of directly chewing on and mulling over the finer points of this textbook moral quandary, you are put in a position to make a snap decision whether you should kill a bad guy or let the courts handle it, including having your own motivations and incentives.

AI doesn't do that. There's no internal consistency, and certainly not consistency with the surrounding game. It won't build a character to explore a point, and also build levels and game and even background events to tie in with it. Do you remember any characters from playing AI dungeon a few years ago? People remember Garrus for decades.

Go read an AI generated book, they aren't very good. They aren't trying to say anything. They don't have a unique perspective or set of ideals they want to explore. They don't have a purpose in their content creation.

If you burn immense effort tuning a procedural generation system, it will occasionally spit out content that seems to have a story, but 95% of it will just be unconnected boring mush. Game paste.

If you want to know what a LLM produced story looks like go play a recent Call of Duty Campaign. It's just a list of 5 second Set Pieces and the entire point of the game is just to be dragged through set piece after set piece with barely the fragrance of a narrative in there. It's utterly boring and desensitizing.

There are no (big) AI games yet, but I would bet money that the next big wave of adventure games (GTA6, TES6, HLX) is planning to make a splash with this tech. But things are changing so fast so it's probably a very challenging planning task.

No one wants to move first with such a polarizing, unproven and rapidly changing tech.

Why pay for a game that's a ChatGPT wrapper, when you can just describe the game to ChatGPT and get the same experience?

It's the same problem as a bunch of startups that made products that were wrappers around OpenAIs APIs. They weren't providing much value.

I tried to write a different "convince an AI" game about a year ago, but it was hard to work with, hard to figure out a business-model for, and more importantly-- it just wasn't very fun to play. Maybe there's a different scenario than the one I chose.

Sheesh. This man doesn't know what he is talking about.

"Here’s a puzzle — back in 2021, when the new AI era was just kicking off, it seemed obvious (to me at least) that it would lead to new kinds of games and radically new forms of gameplay."

Wrong premise. Next.

There are AI games. Check out Weekend (formerly Volley) and play a realtime AI voice game on your smart TV. https://www.weekend.com/

there are some

I built a little mafia game a few years ago, but it was never very popular. http://gptmafia.io/

and someone i know made one too: https://talktomehuman.com

maybe the games market is just hard in general, but i dont know that ai games are super popoular even though people are building them. playing mafia against the ai feels somehow tiring, even if it is also occasionally fun

I have had a great amount of fun prompting Claude with: “Lets play a choose your own adventure game. Give me a completely random scenario. I’ll be a xxxx.”

Surprised nobody has mentioned Pax Historia yet. Very fun game and I believe the most popular startup by traffic in the most recent YC batch.

A YC startup is building one

https://www.paxhistoria.co

If you're reliant on AI you're reliant on some Yankee company's API.

AI romantic partner, doesn't that qualify?

> The video game audience has decided, almost unanimously, that they don’t want generative AI in their games.

Nah, it's just a very vocal minority making lots of noise. We're slowly starting to see saner minds prevail, with steam backtracking on the ban of aigen and so on. People don't care if the end result is good. And the "but muh art" is just virtue signalling. I've used this example before, for me "art" is the fact that I can have a horse as my ruler in EU, not the fact that the horse might be generated by a diffusion model. Art is always in the eye of the beholder, or something.

On a more general note, like with the question "where are all the good open source projects", things will come. We're still in the learning phase of this new shiny toy that we got. It's only been 3 years since the tech is readily available, and maybe 2 years since people can take offline models and run with them. Give it time.

In the meantime, I see a lot of help from agentic coding in game dev. I've been following a space rouguelite thingy on itch, the developer has been pumping updates every other day, and the game is pretty much ready in ~1month. Sure, some updates broke stuff, some things didn't make much sense, but the speed with which the dev could take the feedback and pump out an update is remarkable. Small projects are doing it today. Larger projects will probably do it tomorrow.

My guess is that there's a big chance that the next "big thing" in gameplay w/ AI being somehow used will come from a smaller team / single dev type of thing, and not the big labs. Something genre-defining like minecraft or tarkov.

People are not intrinsically fun? What on earth is this author saying??

As for why no AI games, it's pretty simple really. Planet-destroying plagiarism machines built at a steep loss by oligarchs who delight in putting artists out of business are not fun!