Artificial intelligence in games works on two levels. It helps studios make games (code, art, audio, testing and now production planning), and in a smaller but growing set of titles it runs during play, generating dialogue, controlling characters and judging what the player types or says. Most surveyed developers already use it in some form. A Google Cloud survey of 615 developers in the United States, South Korea and three Nordic countries found 90% integrating AI into their workflows in 2025, while the Game Developers Conference’s 2026 State of the Game Industry survey found 36% of professionals actively using AI tools in their own work. Generative AI is also broader than chatbots: it covers models that produce images, audio, animation and 3D assets as well as text. A 2026 study of indie development reports that 60% of Steam’s AI disclosures concern visual asset generation.
Behaviour Trees and Language Models Are Different Kinds of Game AI
Unreal Engine’s documentation describes Behavior Trees as the asset used to build AI for non-player characters. StateTree is a hierarchical state machine combining behaviour-tree selectors with states and transitions, and the Environment Query System gathers data about the world so an AI can decide what to do next.
The 2013 paper “Playing Atari with Deep Reinforcement Learning” trained one network, with the same architecture and hyperparameters, on seven Atari 2600 games using only screen pixels, rewards and the set of legal actions. It beat all previous reinforcement learning methods on six of those games and an expert human on three.
The AI Native Games survey draws the line explicitly. In its definition, enemy AI, pathfinding, finite-state behaviour, utility-based decisions and adaptive difficulty control are not generative AI. The distinction is not pedantry. A behaviour tree costs next to nothing per decision, while every call to a generative model carries a marginal cost tied to tokens, model size and inference time, so the two belong in different parts of a game’s architecture.
Which AI Tools Studios Use to Build Games
Unity and Unreal Engine both document AI tools for their engines. Unity’s documentation describes an in-editor assistant that writes code, troubleshoots, automates tasks and creates sprites, textures, sounds, animations and materials from text prompts or reference inputs. Unity’s Sentis runs trained machine learning models on end-user devices or in the editor, and Unity says running models on the player’s CPU or GPU reduces server costs. Unreal Engine documents its own neural network engine for running models.
In Unity’s 2024 Gaming Report, 62% of surveyed studios said they used AI, mainly for fast prototyping, concepting, asset creation and worldbuilding, and 71% of those studios said it had improved their delivery and operations. Google Cloud’s 2025 survey put playtesting and balancing (47%), localisation and translation (45%) and code generation and scripting support (44%) at the top of the list.
Why the Adoption Figures Disagree
Google Cloud’s survey counts developers using “some form of AI” in their workflows. GDC’s 2026 survey of more than 2,300 professionals, with a stated ±3% margin of error, counts those actively using AI tools in their own work. The GDC figure measures active personal use and the Google figure any use in workflows, so the two answer different questions.
Writing Aids and Automated Testing
Ubisoft’s Ghostwriter is an assistive tool for scriptwriters. It drafts barks (the short lines NPCs say when an event triggers) so scriptwriters can spend their time on the main narrative. The writer defines the character and the interaction, picks between pairs of generated variations and edits freely, and the tool learns from those choices.
EA’s SEED group added an experimental reinforcement learning system to an existing scripted-bot testing setup to increase test coverage on AAA games including Battlefield 2042 and Dead Space (2023). That is the sensible shape for test automation: learned agents explore states nobody thought to script, and deterministic bots keep the regression checks repeatable.
How Generative AI Is Used in Video Game Development Today
Generative methods in games have mostly been applied before play, to create assets, levels, rules and code. Procedural content generation (PCG) is the algorithmic generation of levels, visuals or whole games, and unlike free-form storytelling it must satisfy constraints such as playability and aesthetic quality. MarioGPT, for example, trained a GPT-2 model on fewer than 20 Super Mario Bros levels and used an evolutionary loop to produce a large, diverse set of playable ones.
Design assistance comes in degrees. The 2024 survey “Large Language Models and Games” separates conceptual assistance, where the output is not game-ready and the designer adapts it, from procedural assistance, where the AI takes part in producing increasingly final versions of an asset; in both cases the designer stays responsible for curating the result. Many creatives use LLMs and image generators this way for brainstorming and concept development.
On Steam, generative AI in shipped content must now be disclosed. One in five 2025 releases disclosed generative AI use, an eightfold rise in disclosed titles over eighteen months. Valve’s form separates pre-generated content, which follows the same rules as any other content, from live-generated content, where developers must describe the guardrails that stop their AI producing illegal material. Behind-the-scenes efficiency tools fall outside the requirement.
Ownership remains uncertain. The 2024 survey of LLMs and games reads past US legal consensus as indicating that content an LLM produces automatically cannot be copyrighted, while work where the model acts as an assistive tool to a designer likely can be. In Google Cloud’s survey, 32% of developers named uncertainty over who owns AI-generated content as a challenge.
How Do AI NPCs and Adaptive Gameplay Actually Work?
LLMs can generate context-appropriate replies to whatever the player says, which cuts repetitive dialogue. Foreground NPCs tied to the plot need tightly constrained output that tracks the narrative and the playthrough, while background NPCs making small talk can be looser. An LLM-driven character is where conversational AI and generative AI overlap, and the engineering problem is keeping the generative half inside the game’s rules.
The Generative Agents research (Park et al., 2023) gave each agent a complete natural-language record of its experiences, condensed those memories into higher-level reflections over time and retrieved them to plan behaviour. In a sandbox town of 25 agents, a single seeded idea that one agent wanted to throw a Valentine’s Day party led the others to spread invitations, ask each other out and arrive together on time.
In Ubisoft’s NEO NPC prototype, built with Inworld’s LLM and NVIDIA’s Audio2Face, a writer creates each character’s personality, backstory and conversation style and keeps tuning it once the model starts improvising. A data scientist shapes the model’s behaviour through guardrails, analysis of player input and textual instructions, and filters catch toxic player input, after which the character stops cooperating. Ubisoft describes NEO NPC as a prototype with some way to go before it reaches a game.
A Working Architecture: The PUBG Ally Case
NVIDIA’s account says its ACE technology powers KRAFTON’s PUBG Ally, a co-playable character. A behaviour tree handles movement and combat at game tick rate, and a language model handles reasoning, coordination with the player and speech. The model is a 2B-parameter Mistral-NeMo-Minitron small language model running locally on the player’s GPU, which removes network latency and fits in the VRAM left over after the game itself. KRAFTON narrowed its world to one map and mode and distilled game-specific knowledge from a larger teacher model into the on-device model, with structured long-term memory carrying preferences and match history across sessions. Validation combined automated checks, live A/B playtests and feedback from more than a thousand players.
What Breaks: Hallucination, Latency and Cost
Hallucination, latency and cost constrain AI NPCs. NPCs can hallucinate quests that do not exist in the game. When model inference is slow or unpredictable, the turn rhythm breaks and play starts to feel like waiting on a remote service. Inference also has a per-use cost, so practical systems handle frequent, low-risk behaviour with small local models or traditional AI, save larger models for moments that matter, and cache results so generated content becomes stable game state. Because AI-native games generate content during play, after certification, safety becomes a runtime design problem rather than a pre-release review.
AI-Native Versus AI-Augmented Games
The AI Native Games survey classes a game as AI-native when removing the generative component would make the core loop impossible or fundamentally different. An AI-augmented game uses runtime generation for a non-core layer, so the game survives if authored assets replace the model. Of the 53 AI-native games and prototypes the survey analysed, 24 were narrative adventures, the format where language naturally sits in the core loop. Few successful ones rely on unrestricted generation: most combine models with some mix of authored goals, hidden state, safety filters, cached outputs and engine validation.
Matching the Kind of Game AI to the Job
| Job in the game or pipeline | Typical technique | Where it runs | Main risk |
|---|---|---|---|
| NPC movement, combat, routine decisions | Behaviour trees, StateTree, EQS | Every game tick, on the player’s machine | Predictable, scripted-feeling behaviour |
| Free-form NPC conversation | LLM with writer-defined persona and guardrails | Runtime, local small model or hosted model | Hallucinated quests, latency, per-call cost |
| Sprites, textures, sounds, animations | Image, audio and animation generators in the editor | Before ship (pre-generated) | Ownership and disclosure questions |
| Levels and puzzles | Procedural generation, e.g. MarioGPT | Before ship | Unplayable or unbalanced output |
| Test coverage | Scripted bots plus RL agents | Existing automated game-testing solution | RL system still experimental |
What AI Is Doing to Studio Costs, Workflows and Jobs
One measured cost effect is in production planning. A 2026 study of an agentic planning tool found it generated a structured plan of about 16 epics and 59 stories in roughly five minutes for $0.27 to $0.58, against a traditional baseline of $2,400 to $4,800 and one to two weeks of producer time. The author co-founded the platform studied and states that the comparison covers cost and time only, not whether the cheap plan performs as well as the expensive one. Adoption is also uneven by layer: GDC’s 2026 survey found 58% use at publishers, support and marketing firms against 30% among studio production staff.
Jobs are the harder question. The same study cites outside estimates totalling roughly 44,000 game industry positions lost from 2022 to 2025, and GDC’s 2026 survey found 28% of respondents laid off in the previous two years. Whether AI caused those cuts is not established: companies’ own stated reasons, as reported in GDC’s 2026 survey, cite automation only 6% of the time, and both the displacement and the exculpatory accounts are confounded by the post-pandemic correction. AI’s share of AAA layoffs is unmeasured, so a page claiming either “AI is causing the layoffs” or “AI has nothing to do with them” outruns the evidence.
Cheaper production is also meeting a crowded market. Steam releases rose from 9,654 in 2020 to more than 20,000 in 2025, and about 300 titles (1.5%) grossed above $1 million.
Why Performers and Artists Push Back on AI in Games
The pushback is strongest among the people closest to the craft. In GDC’s 2026 survey, 52% of professionals said generative AI was having a negative impact on the industry, up from 30% the year before and 18% the year before that, with visual and technical artists (64%), designers and narrative staff (63%) and programmers (59%) the most negative.
For performers the issue is consent over their voices and likenesses. SAG-AFTRA’s video game performers struck after nearly two years of negotiations stalled over AI protections, arguing that voices and likenesses could be replicated by AI without consent or fair pay. The agreement covered more than 2,500 voice, motion capture, stunt and other performers, and the union said studios refused to extend AI protections to on-camera performers. Performers also warned that AI could take the small background NPC roles where newer actors usually start. The 2025 Interactive Media Agreement that members ratified set informed-consent requirements across AI uses and let performers suspend consent for digital replica use during a strike, alongside an immediate 15.17% pay increase.
Players and award bodies are drawing their own lines. Embark Studios’ Arc Raiders used AI text-to-speech built from licensed actor recordings for location-ping lines, which some players found unacceptable. Clair Obscur: Expedition 33 had its Indie Game Awards withdrawn two days after winning over its generative AI use, which the awards make ineligible; it launched with generated posters later patched out. On the studio side, an OECD.AI incident record reports that Activision Blizzard pushed artists to use Midjourney and Stable Diffusion for Call of Duty: Modern Warfare 3 concept art and cosmetics while cutting about 1,900 jobs.
Can AI Make a Whole Game Without a Team?
No, on current evidence. Keywords Studios ran a six-month project, Project Ava, that evaluated more than 400 generative AI tools, and ended up drawing on staff from seven of its studios because the tooling was “unable to replace talent”. Its CEO’s conclusion was that generative AI can simplify or speed up certain processes, but the quality needed comes from experts using it as a tool.
The planning study found a similar limit in generated plans. Users described generated tasks as “good but generic”, and students could produce a professionally formatted plan on day one without being able to tell a good plan from a plausible one.
For a studio deciding whether to avoid AI altogether, the narrower question is which jobs to give it. Surveyed developers report AI use in playtesting, localisation, code and prototyping, while performers and award bodies have contested its use for voices, likenesses and shipped content. Ubisoft’s NEO NPC team, building a generative character system, still held that an expressive character needs an expressive actor.
The open engineering question is runtime, not production. Whether a generated character can stay in character, answer within a frame-time budget and fit beside the renderer on a player’s GPU is still being settled one constrained design at a time, and the games that get it right will look less like a chatbot inside a level and more like a behaviour tree that occasionally thinks.