Blinky chased you. Pinky tried to cut you off. Inky was moody and hard to read. Clyde just wandered off to do his own thing. Four ghosts, four distinct behaviours, and an entire generation of players who learned to fear the pink one. That was Pac-Man. That was 1980.
I bring it up because the games industry has been shipping AI longer than most of the people now arguing about AI have been alive. The generative wave is new in one specific way — now the AI can also write the dialogue, build the level, and talk back to the player. But the craft of making any of it feel good runs in a continuous line back to those four ghosts. People who forget that tend to ship games that are technically impressive and no fun at all.
Three things are getting genuinely exciting. Then I want to talk about the part that's making me nervous, because it's not the part you'd guess.
Companions that actually remember you. The oldest sin in RPGs: you meet a beloved character the second time and they greet you with the same thirty canned lines as the first. That wall is finally coming down. Models grounded in a game's own lore can produce companions who recall what you said last session and react to choices you actually made. The implementations that work are hybrids — human writers and human actors still carry the story and the big emotional beats, and the model fills the seams: the small talk on the road, the reaction to the weird thing you just did. Modest use of a powerful tool. It works precisely because it's modest.
Procedural content that isn't boring. Procedural generation is ancient, and most of it had a tell — infinite corridors of the same idea, technically endless and experientially deadening. Diffusion models and LLMs are shifting that. Terrain that genuinely varies. Side-quests with twists that read as authored. Dungeons whose layout has narrative logic, not just spatial logic. The old Dwarf Fortress school was brilliant-but-inscrutable; the new school is brilliant-and-readable. And crucially, it doesn't replace level designers — it hands them faster iteration. "Give me twenty takes on this shrine in this tone" went from a week to an afternoon, and the designer still picks the one that's right.
Accessibility that ships on every budget. This one I'm proud to watch happen. Real-time voice conversion lets players with speech differences join voice chat without that being the first thing anyone clocks. AI live captions make competitive games playable for Deaf players. Generative helpers let players with cognitive disabilities work through complex systems at their own pace. None of this replaces dedicated accessibility design — the work coming out of studios like Naughty Dog is still the bar — but it puts features that once needed a AAA budget within reach of a four-person team.
So what's making me nervous
Not the robots-take-our-jobs headline, exactly. Two less obvious things.
The first is the squeeze on creative labour. Some studios are reaching for generative AI to hire fewer writers, concept artists and voice actors. That's not only an economic worry, it's a craft one. Games are barely fifty years old as an art form, and the people inventing its grammar deserve to keep eating. The optimistic version — where AI makes each creative more prolific and the industry grows to match — is entirely possible. It is also a choice, being made right now, and it can just as easily go the other way.
The second is personalisation pointed at the wrong target. Tuning difficulty and pacing to an individual player can be wonderful; the game meets you where you are. But the same machinery can tune spending to each individual — a slot machine recalibrated for every gambler at the table. Regulators are starting to notice, and they should.
If a studio asked me how to use this stuff well, I'd say four things and stop. Amplify your creatives, don't subtract them; the brush is not the painter. Playtest generated content with real humans, because it fails in fascinating ways and your schedule needs room for that. Be open about where AI touched the voice, the writing, the art — players are sharp about this and trust is hard to win back. And keep the humans in the credits, especially anyone whose work trained a model that later generated your assets.
Games, more than almost any medium, are about play and joy. It would be a genuine waste to take the most thrilling AI moment in their history and use it to build something that feels like a spreadsheet with a soundtrack.
