Two completely different things are called AI in games, and conflating them is why most coverage of this topic is confusing.
There is the AI that has been in games since the 1980s — the systems making enemies flank you and racers keep pace. And there is generative AI, the thing currently being announced in press releases. They share a name and almost nothing else.
The AI already in your games is not machine learning
Game AI is largely hand-authored: state machines, behaviour trees, utility scoring, navigation meshes, planners. It is deterministic, debuggable, and tuned by designers who can explain every decision it makes.
This matters because the classic examples of "smart" enemies are triumphs of authoring rather than learning. F.E.A.R.'s squads, still cited twenty years on, ran a goal-oriented planner — the soldiers were not learning, they were selecting from carefully written options and, crucially, announcing them out loud. Half of that legendary intelligence was audio design telling you what was happening.
Alien: Isolation's Xenomorph is the sharper lesson. It is famously two systems — one that always knows where you are, and one that only receives hints — and the second is deliberately kept slightly stupid so the hunt stays survivable. Its brilliance is in being handicapped on purpose.
That is the recurring finding across decades of game AI: the goal is never maximum intelligence. A genuinely optimal opponent is miserable to play against. Perfect aim is trivial to write and gets deliberately degraded. What designers actually build is an opponent that is legible, beatable, and fun to lose to — none of which are things an optimiser produces on its own.
What generative AI could plausibly change
Set the hype aside and a few applications look genuinely useful:
- Volume in the unimportant middle. Barks, ambient chatter, filler dialogue for the fiftieth guard. Nobody's favourite line is a guard bark, and there are never enough of them.
- Animation and asset workflows. Motion matching and ML-assisted cleanup are already in shipping engines and mostly invisible, which is the sign of a tool working.
- Localisation drafts for human editing, especially in the long tail of languages that otherwise get skipped.
- Playtesting at scale — agents pathing through a level to find geometry holes and softlocks, a genuinely good fit because you want breadth rather than judgement.
Notice these are all production tools. The player-facing case is far weaker than the pitch suggests.
The infinite-NPC-conversation problem
The recurring demo is an NPC you can say anything to. It is compelling for about ninety seconds, and then the problem becomes obvious: an NPC who can discuss anything cannot be relied on to say anything.
Good game dialogue is load-bearing. It tells you where to go, foreshadows, characterises through specific word choice, and lands a joke on a rhythm someone tuned. A generated line is plausible-sounding and unaccountable — it cannot be trusted to convey the quest-critical detail, and it will happily confirm something that is not true about the world.
Then there is voice. Text-only NPCs feel like a regression after two decades of full performance, and synthesised delivery is currently the difference between a line landing and merely occurring. Baldur's Gate 3's companions work because specific actors made specific choices about specific lines — the least automatable part of the whole enterprise.
The likely shape of this is unglamorous: generated filler in the gaps, hand-written and performed dialogue anywhere it matters. Which is roughly how games already treat barks.
The parts nobody wants to discuss
Three real problems, stated plainly.
Training data. Voice models built on performances whose contracts never contemplated this are the reason performers went on strike over it. This is unresolved, not settled.
Cost per player. Inference is not free. A game running model calls for every conversation has an ongoing per-player cost that no current business model absorbs comfortably — which is a bigger constraint on shipping this than any technical limit.
The QA problem. You cannot fully test a system whose output space is unbounded. Studios ship deterministic content partly because they can verify it. A generative NPC is a permanent open risk of saying something wrong, offensive, or lore-breaking to a customer.
The realistic near future
Generative AI will mostly show up where you cannot see it — asset pipelines, localisation drafts, automated playtesting, filler barks. Not as the NPC you have a real conversation with.
And the games that feel smartest will keep being the ones with the best-authored systems, because the hard problem was never producing intelligence. It was producing intelligence a player can read, predict, and enjoy outwitting. That remains a design problem, and there is no model that solves it for you.