Games with Artificial Intelligence

Games with artificial intelligence are older than the current wave of demos. Long before language models wrote quest text, designers built enemies that chased, fled, flanked, and cheated just enough to feel alive. “AI” in a game credit could mean a state machine, a pathfinder, a utility scorer, a planner, or a pile of special cases dressed as personality. Players rarely asked which. They asked whether the fight was readable and whether the companion was useful.

This article keeps that practical frame. It maps the main ways artificial intelligence shows up in games, how endless score chases should use it, and how the phrase differs from the newer label “super intelligence.” The companion piece on games with super intelligence covers the rename. The piece on AI-generated infinite games covers authorship pipelines. Here the subject is the whole family of machine decision-making inside play, from classic NPC brains to modern learning systems, under the name people still search most often: artificial intelligence.

What players mean when they say AI

In ordinary player language, AI is whatever made the other side act. A guard that hears a footstep has AI. A racing ghost that replays your best lap has AI. A director that spawns a tank when you are doing too well has AI. A tool that filled a forest with trees before ship has AI only in the production sense; players may never say the word.

That looseness is not a bug in conversation, but it is a hazard in design docs. A team that says “improve the AI” without naming the technique will argue for weeks. Improve the perception graph? The navigation mesh? The aim assist? The dialogue model? The difficulty scaler? Artificial intelligence in games is a family of jobs, not a single ingredient you sprinkle until the trailer glows.

The useful habit is to pair the umbrella with a noun: combat AI, navigation AI, narrative AI, generative AI, analytics AI. Infinite Game’s articles do the same when they talk about generation versus rivalry versus tools. Clarity is part of fairness. Players can forgive a simple brain that telegraphs. They struggle with a mysterious brain that seems to rewrite the rules.

The classic stack still carries most shipped games

Most commercial titles still run on techniques that predate deep learning.

Perception systems decide what an agent knows: line of sight, sound radii, last known positions. Navigation finds a path on a mesh or grid. Decision systems choose among actions with state machines, behavior trees, or utility scores. Animation and aiming layers turn those choices into motion that does not jitter. Special cases paper over the moments where the general system looks stupid.

This stack is artificial intelligence in the games-industry sense even when no neural network is involved. It is also controllable. Designers can tune a hearing radius. They can forbid an enemy from using a grenade in a tutorial room. They can guarantee a telegraph. For a score chase, controllability matters more than fashion. A maximum score wants enemies and hazards whose legal moves can be listed.

Learning systems enter when the team wants policies that would be painful to hand-author: a bot that plays a complex multiplayer title at a high level, a ranking model that predicts churn, a vision model that flags broken collision in a screenshot. Those tools are powerful. They are not automatically better at teaching a first-time player why they died.

Where AI sits in an endless run

An endless game is a loop that keeps asking for another legal moment. Artificial intelligence can sit in several seats around that loop.

As a generator, it proposes the next stretch. That seat needs a grammar and a veto, or the loop becomes a slot machine of unreadable deaths.

As a director, it chooses intensity. After a hard phrase, it can insist on a rest beat. After a long safe stretch, it can raise density inside a budget. Directors are easy to over-trust. If the director can mint new verbs, it is no longer directing; it is rewriting the sport.

As a rival, it races the same course. Ghosts, time-trial ghosts, and learned policies can push a player without changing the level. Shared seeds keep the rivalry honest.

As a coach, it watches failures and suggests a tip. Coaching is safest when it points at a rule the game already taught. A coach that invents a technique the physics does not support becomes another liar.

As a companion, it fills silence with talk. In a pure score chase, companions are optional flavor. Flavor must never block the next jump.

The image is a rivalry picture on purpose. Artificial intelligence in games is often imagined as an enemy mind. In endless design, the better first question is whether you need a mind at all, or whether you need a fair course and a number. Minds are expensive. Courses are the sport.

Generative AI is one branch, not the whole tree

The last few years collapsed “AI” into “generative model” in public talk. Games should resist that collapse. A pathfinding crow does not become smarter because a diffusion model can draw a crow. A behavior tree companion does not become wiser because a language model can improvise a monologue.

Generative systems earn their place when variety is the bottleneck and semantics can be constrained. They are weak when the bottleneck is clarity. Endless games starve for clarity more often than they starve for novelty. A hundred new silhouettes that lie about collision are worse than twelve honest tiles in a fresh order. The AI generation article is the deeper dive on that branch. This article’s point is taxonomic: generative AI is a branch of games with artificial intelligence, not a replacement for the tree.

Adaptation without moving the goalposts

Dynamic difficulty adjustment is one of the oldest AI-adjacent ideas in games. Slow the chase if the player fails often. Raise the spawn rate if they never miss. In a narrative adventure, soft adaptation can keep a story moving. In a public score chase, silent adaptation can poison the board.

The repair is to separate practice modes from record modes. Let a coach mode adapt freely. Let a ranked or “official run” mode freeze the rules, publish the seed, and refuse mid-run rewrites of enemy budgets. Players who want a personal trainer can have one. Players who want a comparable maximum score need a fixed sport.

The same split applies to personalization powered by player models. Knowing that someone loves precision platforming is useful for recommending a mode. It is dangerous if the live course quietly becomes a different game for each account while sharing one leaderboard. Artificial intelligence that profiles players belongs in menus, matchmaking, and practice tools before it belongs inside the scored loop.

Ethics and authorship under the ordinary name

Even without the “super” prefix, AI in games raises authorship questions. Who created the quest line the model drafted? Who owns a creature the image model painted? What happens when a dialogue model quotes a famous line or insults a player? Games with artificial intelligence need filters, provenance, and a human-readable credit path for anything that ships.

They also need honesty about imitation. A model trained on many games will suggest many games. Genre fluency is fine. Specific character theft is not. Infinite Game’s image briefs already refuse known mascots and readable logos for that reason. The playable layer should refuse them too, whether the author was a person, a classical generator, or a model.

How this differs from “super intelligence” talk

Artificial intelligence is the established search term and the honest umbrella. Super intelligence is either a research threshold or, more often now, a marketing rename for advanced AI. If you arrived here from that newer phrase, the practical advice is identical: name the subsystem, gate anything that can change a score, keep failures soft, and measure difficulty with numbers. The super intelligence article exists so the rename has a landing page that translates instead of inflating.

Infinite Game uses “artificial intelligence” when it means machine decision-making or generative tools in general. It uses “super intelligence” when discussing the public rename and the stronger claim. It uses neither as a substitute for “chunk grammar,” “seed,” or “fairness checker.”

What the hero run uses today

The playable hero is an endless two-face platform run. Its course extends from a designed grammar. That is procedural content in the classical sense, authored by people, not a live learning agent. Deaths return you to a checkpoint. The score is distance. No model is choosing your next gap in real time.

That choice is intentional. Before AI sits closer to the score, the site wants the design notes public: what an infinite game can mean, how generation pipelines should be gated, how difficulty should be measured, and how the two AI vocabularies — ordinary and “super” — should be translated for players.

The workshop image is the target culture. Artificial intelligence is welcome at the table. The rulebook stays on the table. The seed stays on the table. The stage that reaches the player is the only object allowed to change the number in the corner.

A practical checklist for games with artificial intelligence

Write the player-facing verbs and failure conditions before you pick an AI technique. Name each AI job with a concrete noun. Keep scored modes deterministic enough to replay when records matter. Put generative systems behind grammars and checkers. Let directors pace intensity inside budgets, not invent new buttons. Separate adaptive practice from official runs. Fail soft when a model or a network is unavailable. Filter player-facing text and images for contradiction, abuse, and specific copies of existing characters. Explain deaths with readable telegraphs. Credit humans and tools without hiding which one decided collision.

When that list is true, artificial intelligence is a craft amplifier. When it is false, AI is a fog machine. Endless games punish fog. The maximum score should stay a measure of timing and reading under published rules, whether the other side of the canyon is a script, a planner, a learned policy, or a generating model with a fashionable new name.