How will Generative AI change video games

The Game Room
,
1

Open world games are likely to be changed dramatically by generative AI; detailed environments at every scale, more complex and more variety of NPCs. It will make content generation so much easier that, as I’ve noted previously, I think it will be a historical curiosity that open world games preceded generative AI by so many years.

But that’s as far as my imagination has gone. What other ways might games change? Could there be entirely new genres?

And (for the more tech-inclined) is it likely to need new hardware? The next gen of consoles, and an AI specific card in PCs?
(I’m aware that GPUs are doing the job of AI today, but there are a few start ups and projects from the bigger players on making something more dedicated to AI. And even if GPUs remain the standard, that doesn’t mean there won’t be consumer boards with GPUs but explicitly for AI only, like the datacenter cards)

2

Graphics cards that you might reasonably buy already have DLSS and “tensor cores” for better graphics:

These cores do certain types of computation; I do not think you should necessarily call them “explicitly for AI only”

3

Ah that’s cool. But yes I was trying to be more specific than this. Cards mostly, or entirely, for AI rather than a graphics card with some AI-driven enhancement on the side.

Of course this may never happen; it might just be that graphics and AI will be integrated such that the same GPUs on the same board do both.

4

There are two, somewhat independent, responses whether the GenAI is performed by the game developer or by the gamer. And ultimately this comes down to whether or not the AI content is customized to the gamer.

Not Customized Content:
If the content is not customized to the gamer’s state, choices, settings then it will be more efficient and produce better results if it is done by the developer and deployed: either as DLC or on-demand (e.g. “when the gamer runs out of quests, download more”). This is basically the development model used today, but with GenAI doing part of the work.

Even for a vast open world game, it would make sense for the developer new areas, quests, items one time on their systems. It also let’s them fix issues and improve the generation so that gamers are using the best version of the new content.

Customized Content:
If the content is customized to the gamer, then from a cost perspective it makes sense to do it on the gamer’s computer. Generating customized dialog, voice “acting”, and translations are pretty lightweight and don’t necessarily need to use the gamer’s GPU; they would perform well with the CPU or NPU. This is especially true if they can be pre-generated in the background.

Some immediate, lightweight ideas:

  • Customize the NPC responses based on the gamer’s current state within the entire narrative.
  • Use specific names, genders, identifiers in dialog. This is often done in text, but not always. I’m playing a game now where I noticed the dialog uses awkward phrasing to avoid whether my character is a male or female.
  • Building on the previous point, I’ve noticed that many games are taking a modern approach to gender. GenAI could let the gamer dial-in exactly how they want this to be handled.
  • Use Swahili for the text and dialog.
  • Let the gamer provide open prompt instructions. “Do not swear”, “Use the term Hobbit, not Halfling”, “Don’t be woke”, “Make all of the enemies types of dinosaurs”, etc.

Some immediate, medium-weight ideas:

  • Generate gamer customized levels, areas, quests, etc. I think these are often created internally with scripts so fine-tuning an LLM to generate custom scripts (as code) is possible. Currently this would probably need to be done on-demand or on an event – not dynamically as the character is walking around.
  • Tweak or mod the ruleset. Add, change, remove recipes. Adjust balance.
  • Replace textures or images with gamer prompted versions.
  • Generate new cut-scenes
  • I think the end point for this medium term would be to replace a game’s modding interface with a higher-level AI-powered interface so that the gamer doesn’t have to learn how to mod in order to mod.

Many of the customization ideas hinge on how well they integrate in with voice acting. It is easy to generate new dialog, it’s easy to speak the dialog, but more work is needed to act the dialog well. Games will continue to use human voice actors so customized content would need to blend with those voices.

Longer-term we might see games move toward generating content in real-time on the gamer’s computer, but I think we are a ways away from that being the most efficient solution.

Finally, my response was really only considering the high-end, marque games. As you move down in budget, where expectations are different / lower, we will see GenAI being used more aggressively. A smaller shop or individual developer might use GenAI now to add translations or voice acting where before they would have skipped it.

5

Right now, the way it’s changing games is that it’s the kiss of death for any game that gets caught using it.

https://www.polygon.com/party-animals-ai-video-contest-review-bombing/

https://www.washingtonpost.com/technology/2026/01/26/gamer-protests-ai-slop-backlash/

6

I feel like I’m suffering from a lack of creativity here. As noted, an obvious choice would be things like NPC dialogue. But I feel like you’d need a small LLM model to be safe so your barmaid doesn’t start talking about Zionism or how to make bombs as players bang against the guard rails. At that point, it feels safer to just use AI to assist in making 10,000 vetted lines of dialogue (potentially even voiced) and put those in the game rather than allow the NPCs to free-style it. Likewise for things like image assets or procedural map generation. A definite useful tool for the development phase but not necessarily something you want to set free on the players.

7

Does anyone remember AI Dungeon? A text-based, storyteller based on GPT-2?

This was my first introduction to LLMs and how good (and bad) they can be.

8

Yeah, I suspect that we’ll see AI first making an impact on the development side of things before we see it directly creating content.

Like we’ll see AI-generated imagery, models, voice work, etc… well before we see real time AI generated gameplay elements.

9

Not really a prediction, but for years I’ve been thinking how creepy it would be if a horror game took images from the player’s webcam (of the player, passers-by, etc.) and incorporated them into the game.

Better (or indeed, worse) than that would be to have AI scrape the player’s social media feeds and have all the NPCs in a game looking uncannily familiar.

10

You don’t need (what’s currently called) “generative AI” for that. Games have been doing that for decades. It’s usually called “procedurally-generated worlds”. I rather doubt that there’s any open-world game in the past 20 years that hasn’t.

11

Sure; I’ve even worked on a game like that.
But I meant more detailed and/or interactive content than is feasible with conventional procedural algorithms. It’s very hard to write a bespoke procedural algorithm that can fill every house of a virtual city with realistic objects. It’s more feasible (not yet in real time) to train a modern AI to do it.
As with many terms, “generative AI” can be misleading because of course various flavors of this have existed as long as computers have, but it has a more specific meaning now.

12

Great response btw @CaveMike and it is indeed interesting to think how this will play out over time, as the algorithms, hardware and toolchain develop over time.

13

I am amazed I am not seeing AI used to program game AI to be a better opponent. Maybe they are doing that but I have not seen it. AI opponents are notoriously bad and almost always cheat in some fashion to be competitive.

I’d much rather have an AI, playing by the same rules I am, fight me.

14

Not only do I remember it, I still have it on my phone.

It looks very different than it did when I first downloaded it, back when it was brand new. When it was new, you’d pick a vague category (sci-fi, fantasy, post apocalypse) and then the game would wing it. Now it’s a bunch of very specific scenarios, and half of them look overtly sexual in premise (including PG-rated hentai-looking cover art). Ugh.

15

What about those libraries like Infinigen?

16

This is where it’s going. The boundary between procedural algorithms and generative systems is going to become very blurry. Studios will have both humans and custom-trained AI use their in-house development tools to generate game assets and content.

17

I recall a game dev discussing that on the old City of Heroes forum in response to complaints about enemy AI; they point out that to a large degree the actual issue is balance. They listed off a bunch of easy ways they could improve the AI, but all of them would render the enemy somewhere between “extremely annoying” to “invincible”.

Since a computer game program and a person playing it are very different things interacting with the game world in very different ways, it’s hard to make the intersection between the two a “fair fight”. You’re matching a sapient being against a non-sapient program with in-game omniscience, superhuman reflexes and arbitrary accuracy that can change the in-game rules as desired. So it’s easy to make a program that will win every time or lose every time, but harder to make one that just puts up a good fight.

18

Oh wow, you weren’t kidding. That’s too bad.

Maybe it’s a gap in the tools available. Generating graphics, sounds, dialog, etc. are tasks that are generic – not specific to gaming. But improving AI opponents needs to be customized.

FWIW, the PvE games I’ve played recently have decent AI opponents (at least at their tactical level). The quality seems to scale with budget though.

PvP games solved this with human opponents. Considering the size of the slice of the pie, it might have stalled AI opponent technology.

19

A game like City of Heroes, that’s fundamentally PVE, is very different from, say, Starcraft, or Quake. In PVE games, a single player is expected to be able to overcome hundreds, thousands, or more enemies, so the enemies must have inherent weaknesses to make that possible, and poor AI is a common choice of weakness. Look at Mario, for instance: 3/4 of the angles at which Mario can meet a koopa, the koopa wins, and there’s a ton of them. What makes the game winnable is that all a koopa can do is move in one direction at a constant speed.

By contrast, in Starcraft, the basic gameplay design is PVP, where both players start off with the same knowledge, the same resources, and the same capability. There’s not an expectation that you should always be able to beat your opponent: Logically, it’ll only happen half the time. An AI player is meant to replace a human player, for times when you’re offline or just don’t feel like dealing with other humans.

You can make a Starcraft AI by giving it in-game omniscience and letting it change the rules as desired. And that, by and large, is what the computer game companies do, at the higher difficulty levels. Which is what people complain about when they complain about bad AI, because they don’t want an opponent that wins by cheating; they want an opponent that wins by just playing really well.