When the first Skyrim mod with Chat GPT hooked in came out, I was completely stoked. I couldn’t believe it. Characters saying something that isn’t scripted. Characters rambling about their made-up personal lives and making bad jokes. For a moment, it almost felt like I was glimpsing the future of gaming. But after spending three hours of my Sunday night talking to AI-powered NPCs, cracks started to appear. Conversations often veered into nonsense. Characters forgot key details moments later. And after a while, it wasn’t fun anymore. Something that initially suggested we’re on the verge of an infinitely dynamic world left a bitter taste in my mouth: There’s a real risk that generative AI could make games worse, not better.
Generative NPCs And Where They Fail
NPCs can make or break immersion, and with generative AI’s, NPCs suddenly get huge potential. We all know traditional games rely on painstakingly crafted scripts. Take Red Dead Redemption 2 with over 500,000 lines of dialogue, each recorded by voice actors and tied to specific triggers. NPCs follow rigid schedules, repeating the same routines unless the player intervenes. It’s a realistic illusion of life, but one that crumbles under scrutiny, and the hard work of many hundreds of people. What if there were an easier way to do it? What if something like generative AI delivers something very similar, or even much better? Instead of dialogue trees, players could, theoretically, ask anything, and get a unique response. Inworld’s demos show detectives interrogating suspects in real time, while Ubisoft experiments with AI-generated side characters who remember your choices. So how does it work?
A generative NPC is really three nested systems. At the bottom sits the giant language model, pretrained on internet text and then fine-tuned with studio lore. A middle “character sheet” layer adds memories and personality sliders. Confident vs. shy, gullible vs. cynical and such. Also the world knowledge so a medieval smith doesn’t reference Apple Pay. A thin orchestration layer finally turns raw text into lip-synced dialogue, branching quests, and on-screen animation. It’s not just plug and play. There’s still work to do with each character. While it’s still not as much work as writing 500,000 lines of dialogue by hand, you need to take into account ridiculous hardware and resource requirements for a game with hundreds of generative AI NPCs.
However, studios already exploit portions of that stack. Ubisoft’s Ghostwriter churns out bar-room chatter, freeing human writers from filling thousands of empty dialogue bubbles. Electronic Arts claims more than half its pipeline could soon be AI-assisted. Research sandboxes like Smallville show twenty-five AI villagers planning a Valentine’s Day party with no designer watching.
So why do generative AI NPCs fail? Say you greet an AI bartender. Your microphone captures the question, speech recognition turns it to text, the text joins the bartender’s memories, the LLM predicts an answer token by token, a voice model reads it aloud, and animation syncs lips to phonemes. Beneath the magic, generative AI in games is still a patchwork of clever tricks. Most systems don’t run raw ChatGPT. Instead, developers fine-tune models to stay somewhat on-brand. Inworld’s studio, for example, lets designers define a character’s personality, speech patterns, and knowledge limits (no Taylor Swift references in medieval RPGs, ideally). Narrative guardrails try to prevent lore-breaking tangents, while voice synthesis tools like ElevenLabs add realistic speech. Currently, the latency is a killer. Waiting seconds for an NPC to generate a response destroys pacing. Consistency is another issue. An AI might brilliantly role-play a grizzled detective in one conversation, then spout nonsense in the next. And without strict controls, characters can hallucinate, inventing facts or contradicting established lore. Even when it works, there’s a deeper question: Is this fun? Game designer Mike Cook points out that endless, unscripted dialogue can feel aimless. “If you can’t think of interesting things to say, you’re basically reading your own bad creative fiction,” he says. Players might enjoy a few novel interactions, but without structure, the experience risks becoming a tech demo rather than a compelling game.
Generative NPCs And Where They Succeed
It would be unfair to say generative AI completely fails in gaming. There are places where it shines, and almost invents whole genres. AI dungeon masters? Brilliant. In text-based roleplaying games like AI Dungeon, generative models allow players to wander through open-ended adventures with no script. The model invents story beats on the fly, responding to outlandish player choices with improv flair. This format doesn’t suit every audience, most players still prefer structured progression, but it does scratch a specific itch. AI companions fall into a similar space. Chatbots with digital personalities aren’t meant to replace rich game narratives. They serve players who want connection or just someone to talk to at 3AM. The AI doesn’t have to be perfect. In fact, a slightly awkward or overly eager companion can be charming. When generative AI can sustain that illusion, it creates moments that feel personal, even if they’re just probabilistic strings from a neural net.
A Nod to Procedural Generation
Investors love recurring revenue, and an “infinite” game that never loses novelty looks like software with zero churn. But investors and players measure success differently. A16Z can celebrate Inworld’s $500 million valuation but the player just wants a memorable gaming session. If that magic is better delivered by 500,000 hand-written lines, as RDR2 already proved, then labor, not algorithms, remains the secret sauce. Some of the loudest claims, like AI-driven games that never repeat, ignore player psychology. No Man’s Sky famously promised infinite planets, yet many found its algorithmically generated worlds shallow (which, to be fair, was somewhat fixed in the No Man’s Sky 2). Same with Minecraft. If you just play it for world exploration, it gets boring pretty fast. Infinite doesn’t mean meaningful. Looking back, generative AI feels like procedural generation’s spiritual successor. But do players really want a million forgettable things, or just a few unforgettable ones? In both cases, studios quickly learned that pure generation isn’t enough. You still need curation. You still need human intent. That’s why Spelunky and Hades succeeded where others fizzled. They used procedural logic inside clearly defined constraints, layered with hand-designed pacing, emotion, and reward. Generative AI will need to follow the same path. Without structure, it risks becoming just another content mill.
A More Grounded Future
So what’s the solution? Does generative AI have a place in video gaming? I would say yes, and I would say it’s not going anywhere. But the approach will be adjusted. The most realistic path forward isn’t AI-dominated games, but AI-enhanced ones. Hidden Door, a narrative startup, uses LLMs to generate dynamic text adventures, but keeps sessions short (20-40 minutes) to avoid meandering. Bitpart’s approach blends AI-generated “Lego bricks” of dialogue, curated by human designers, ensuring coherence. Ubisoft may forever use Ghostwriter for their virtual crowds while reserving main quests for human pens. The frog twitches, but it isn’t singing opera.
Even Frank Lantz, a designer enthralled by AI’s potential, admits the best use might be weird, experimental experiences, not AAA blockbusters. Imagine games where NPCs rewrite the story as you play, or strategy titles where the AI co-creates the rules. But these ideas are far from mainstream.
Generative AI in games works, just not the way keynote slides imply. It succeeds as a tool framed by human creativity and the hard ceilings of latency, cost, and storycraft. It is overhyped because endless novelty is a myth and because improv without direction devolves into noise. There’s also the oldest design problem, “make something fun,” cannot be outsourced to statistics. The next decade will showcase smarter NPCs and leaner pipelines. They will not, however, flip the medium upside down. Skeptics should remember that technology headlines often misread milestones as finish lines. The printing press did not kill storytelling, film did not erase theater, and generative AI will not render authored design obsolete. It will, at best, stretch our toolbox.
