Generative AI in video games is rarely used today for player-facing features. Teams use it in the content pipeline for visual assets, audio, narrative text, marketing material and code. The GDC 2026 State of the Game Industry survey found that 19 percent of all respondents use generative AI tools for asset generation, 10 percent for procedural generation and only 5 percent for player-facing features.
What using generative AI in video games means in 2026
This page looks at three places a studio can put a generative model. The first is the development pipeline, where content is created before launch. The second is the NPC layer, where a language model shapes dialogue or reasoning while the game runs. The third is runtime content generation, where the game produces textures, maps or story content while it runs.
The procedural content generation literature distinguishes two timings: online generation happens while the game is running, and offline generation happens during development. Steam’s live-generated category covers content produced while the game runs, and it requires developers to describe their guardrails against illegal content.
A Totally Human Media analysis of Steam’s AI disclosures found visual asset generation in about 60 percent of them. For background on the model families behind these outputs, our explainer on how generative AI models work covers the mechanics. This page sticks to what happens when those models meet a game build.
How does procedural content generation interact with generative AI in Unity and Unreal?
Procedural content generation (PCG) and generative AI are not the same thing. PCG is the automatic creation of game content using algorithms. That definition covers rule systems, noise functions, grammars and search, none of which need a trained model.
A 2024 survey of the field counted 13 LLM papers in 2023 alone, plus 2 more combining LLMs with other methods, a sharp rise on previous years. The same survey notes that LLMs are mostly used for narratives while GANs are considered better for images. It also points to a role that gets less attention: reinforcement learning and search-based algorithms work well for repairing content that other generators produced, and the authors suggest that repairing flawed content may be a better research target than generating whole items or levels.
Unity Sentis is a neural network inference library that imports trained models into Unity and runs them in real time on the target device’s CPU or GPU. That lets a studio run a model on the player’s own CPU or GPU. The PANGeA research system takes a different route: a Unity plug-in talks to a REST server that can be hosted locally and shipped with the game, or hosted in the cloud, and works with local models served through tools such as llama.cpp or with private APIs such as GPT-4.
How do major engines like Unreal Engine 5 expose procedural content generation tooling, and where does generative AI plug into that PCG workflow versus replace it?
Unreal Engine’s Procedural Content Generation Framework is a node-graph toolset for building procedural tools and content, from asset utilities like buildings or biomes up to entire worlds. Spatial data flows from a PCG Component in the level into the Procedural Node Graph, which generates points; nodes filter and modify those points, and the results spawn assets, with changes updating in real time in the editor viewport. Each point carries its transform, bounds, colour, density, steepness and seed. Epic’s UE 5.8 documentation describes the framework as built for extensibility and for integration into existing world-building pipelines.
With Unreal Engine 5.7, Epic considers the PCG framework suitable for final projects, and the release adds a dedicated PCG Editor Mode and GPU compute optimisations. The Procedural Vegetation Editor that ships alongside it is still Experimental, and Epic recommends verifying runtime performance and scalability before adopting it in production.
So where does a generative model fit? In our reading, it plugs into a PCG workflow in three places and replaces it in almost none:
- Upstream of the graph, producing meshes, textures or variants that become the asset pool the graph spawns from, reviewed before they enter the library.
- In front of the graph, turning a text request into structured parameters. Nyric works like this: an API breaks a player’s text prompt into world attributes, which generate biomes from a pre-loaded cache of assets at runtime.
- Inside custom nodes, since the PCG graph can execute user Blueprints derived from PCGBlueprintElement, so a team can wrap a model call as one step among deterministic ones.
Swapping the seeded, rule-driven graph for a model that emits layouts directly gives up the main reason designers use PCG: a seed that reproduces the same world when someone files a bug.
Where do AI NPCs work, and where do they break?
AI NPCs work for dialogue within a defined persona and knowledge boundary, on hardware the player already owns, when the latency is masked. The authors of a 2025 study on fixed-persona small language models put the constraints plainly: LLMs remain limited for game dialogue because of hardware requirements, latency and the need to keep clear knowledge boundaries.
What the consumer-GPU measurements show
That study fine-tuned small language models (SLMs) with LoRA to encode each NPC’s persona, and paired them with swappable ChromaDB memory stores for conversation history and world knowledge, so several NPCs can share one model while keeping separate histories. All training and evaluation ran on a Windows 11 desktop with an Intel Core i7-8700K, 32 GB of RAM and an NVIDIA RTX 2070 Super with 8 GB of VRAM.
| Model variant (SLM study) | Mean VRAM | Mean latency to full response | Practical fit |
|---|---|---|---|
| Jack variants | ~130 MB | ~0.8 s | Low-resource, interactive |
| Casper (TinyLlama) | ~807 MB | ~1.7–1.9 s | Balanced quality and speed |
| OliverS (Mistral-7B-Instruct) | ~4.2 GB | 5.49 s | Important NPCs, with latency masking |
| OliverQ (quantised) | ~4.2 GB (reported for Oliver variants together) | 34.58 s | Significant latency from quantisation overhead |
Two figures from that study matter more than the headline latency. OliverS reaches its first token in 0.1145 seconds, which is why streaming text or real-time text-to-speech can hide a five-second full response. Swapping an NPC’s memory database took between 0.012 and 0.027 seconds, quick enough to switch NPC instances during play without a visible stall. The trade-off still shows in quality: the balanced CasperS (TinyLlama) variant scored 55 percent on factuality and 63.3 percent on context retention.
KRAFTON describes PUBG Ally running a quantised 2B-parameter Mistral-NeMo-Minitron model locally on the player’s GPU, sized to fit the VRAM headroom left after the game itself. The team tested cloud-hosted LLMs first and found that network latency plus inference latency often made responses too slow for live squad communication, while the local model gave far more predictable response times. For pipeline-side work where the model does not compete with the renderer, our sibling piece on running open-weight LLMs like DeepSeek on Intel hardware for game pipelines covers the hardware set-up.
Where NPCs break
NPCs break when free-form input pushes the model past the narrative’s scope. The PANGeA authors found that free-form text can prompt an LLM to generate content outside the intended narrative. In their tests the baseline Llama-3 (8B) handled most categories of out-of-scope input poorly. Their validation layer improved results and extended what smaller local models could do. AI Roguelite players report the unaided failure in its rawest form: the AI sometimes decides you take damage and die instantly just for talking to a friendly NPC.
The generative agents research shows what memory, reflection and planning can achieve in a sandbox: from one seeded idea of a Valentine’s Day party, 25 agents spread invitations over two in-game days and coordinated to arrive together. That is believable social simulation, but it is not a game loop with frame budgets, combat and win conditions. In PUBG Ally, a behaviour tree handles movement and combat at game tick rate, and the language model handles deliberate reasoning, coordination and speech. We read that split as the honest current boundary: models narrate and coordinate, while classic game AI keeps control of moment-to-moment behaviour and long-running plans. For a small hands-on case of a fully bounded opponent, see our walkthrough of building a chess game with a browser AI opponent.
Which popular games actually ship generative AI features?
The big-name games that ship runtime generative AI keep it tightly scoped. The table separates what runs in the shipped game from what was used in production, and its scope column is the test against marketing.
| Game | What is generated | When | Scope limit |
|---|---|---|---|
| PUBG: BATTLEGROUNDS (PUBG Ally) | Teammate reasoning and speech | Runtime, on-device 2B SLM | One map (Sanhok), one mode (AI Duo), fixed item taxonomy |
| inZOI (Smart Zoi) | Character actions and thoughts | Runtime, on-device 0.5B SLM | GeForce RTX GPUs only |
| NARAKA: BLADEPOINT Mobile PC Version | AI teammate assistance | Runtime | Free Training mode |
| Arc Raiders | Location-ping voice lines via licensed text-to-speech | In-game voice lines at launch | Many lines later replaced with human recordings |
| Fortnite | AI-powered Darth Vader voice | In the live game | One character (Darth Vader) |
| AI Roguelite | Locations, NPCs, enemies, items, mechanics | During play | Everything AI-determined, with visible instability |
The Totally Human Media count found 7,818 Steam titles disclosing generative AI use, about 7 percent of roughly 114,126 titles. That figure is a floor: VGC notes Steam AI disclosures are voluntary, so the real number is likely higher.
Which pipeline patterns let a studio integrate generative AI without breaking determinism and QA?
The patterns that hold up all keep generative output behind something a designer can inspect, replay or reject. The last two items in the checklist below are our engineering judgement.
- Generate offline and freeze. Offline content is created during development, before it ships. Even the NPC study’s training data went through manual review of synthetic prompt-response pairs before fine-tuning.
- Draw from a bounded asset library. Nyric and Bitmagic both have the LLM interpret the request but build from pre-loaded or pre-defined assets. The model chooses; it does not create geometry the art team never approved.
- Put a validation layer between player input and the model. PANGeA generates gameplay rules from the designer’s criteria, asks the model whether an input breaks a rule, and on a “yes” produces an in-character correction. Its authors are explicit that this protects earnest players from derailing the story and is not anti-cheat. Its threat list (prompt leaking, NPC hacking, unauthorised skills) overlaps with broader generative AI security risks.
- Close the world. KRAFTON reports that limiting PUBG Ally to one map, one mode and a fixed item taxonomy made everything downstream far more tractable.
- Ground facts in engine state through tools. Ally fetches its status, teammate, items and combat situation through observation tools, and the system prompt treats those results as the only ground truth for the match.
- Test in layers. KRAFTON combined automated checks on protocol and tool use, live playtests and A/B tests, and large-scale playtests with over a thousand players.
- Keep PCG seeds deterministic and log model outputs as data. Unreal’s PCG points already carry a seed. We would pin model versions and sampling settings, and record every runtime generation alongside the game state, so a bug report can be replayed from the log rather than regenerated.
- Treat a model upgrade as a content change. A new checkpoint changes every output it touches, so we would re-run the validation suite before it ships, as we would for a rebalance patch.
What does the legal picture look like for shipping a game built with generative AI?
Three questions decide it: whether you own what you ship, whether you can prove where it came from, and whether you disclosed it the way each storefront and territory requires.
Ownership and provenance
The US Copyright Office’s 2025 report concludes that generative AI output is protectable only where a human author has determined sufficient expressive elements, through perceptible human work or creative arrangement or modification, and not through prompts alone. Precise prompting does not change this; what counts is control over the expressive elements. For a studio, the narrative, level design logic and creative assembly of assets may be protected, while individual AI-generated art, environments and audio are not. Selection and arrangement can give thin protection to the assembly itself.
Promise Legal’s analysis of Thaler v. Perlmutter (D.C. Cir. 2025) argues that when AI is the sole generative force there is no author, and so a contractor’s purely AI-generated deliverable may carry no copyright to assign under a work-for-hire clause. The same analysis recommends written notice of AI use before delivery and contractor indemnity for training-data claims, pointing to Getty Images v. Stability AI as the litigation backdrop. Tool choice affects exposure as well: Promise Legal identifies Adobe Firefly as the only major image platform offering IP indemnification to paid subscribers, and notes that GitHub Copilot’s defence obligations for commercial users apply only with duplicate detection set to Block. Publisher due diligence checks IP ownership, and AI ownership issues found there lead to delays, renegotiation and, at worst, abortive deals.
Storefront and platform rules
Steam has required AI disclosure since January 2024 and rewrote its rules in January 2026 to focus on player-facing content, separating pre-generated from live-generated material, carving out development tools and adding an in-game way to report unlawful live-generated content. In practice, Copilot-written code needs no disclosure, while Midjourney art or Suno music that ships in the game does. Live-generated content brings an extra duty: developers must describe the guardrails that prevent illegal output, as a real explanation of their moderation approach rather than a checkbox. Inaccurate answers can block a release or lead to removal after launch. Epic Games Store, by contrast, imposes no mandatory AI disclosure.
EU obligations
Under the EU AI Act, players must be told when they are talking to an AI system, so a conversational character must say it is artificial unless context makes that obvious, and Reed Smith notes that the supporting guidelines read that exception narrowly. Synthetic voices or a resurrected performer can count as a deepfake that must be disclosed even with full consent, with fines of up to €15 million or 3 percent of worldwide turnover. Article 50’s machine-readable marking obligations for AI-generated audio, images, video and text apply from 2 August 2026. The engineering side starts with an inventory of AI-generated assets by tool and licence. The legal conclusions belong with counsel.
Where is the controversy over AI in video games landing?
The controversy over AI in games spans labour, IP and content moderation. The GDC 2026 survey of over 2,300 professionals found 36 percent using generative AI at work, yet 52 percent believe it is hurting the industry. Opposition runs highest among visual and technical artists (64 percent), game design and narrative staff (63 percent) and programmers (59 percent). Studio workers report using the tools far less (30 percent) than staff at publishers, support teams and marketing firms (58 percent).
Voice performance drew formal union action. SAG-AFTRA filed an unfair labour practice charge over Fortnite’s AI-powered James Earl Jones Darth Vader voice, arguing it displaced bargaining-unit work without negotiation. The union’s members ratified the 2025 Interactive Media Agreement by 95.04 percent to 4.96 percent, ending an 11-month strike with what both sides described as AI protections. Arc Raiders shows how player sentiment works on top of contracts: Embark had paid actors to license their voices for text-to-speech location pings, still faced backlash, and has since replaced many of those lines with human recordings its CEO calls better. The trade-offs in our piece on the benefits of generative AI for text-to-speech look different once a union contract and a player base are involved.
On IP, cloning has become cheaper and faster. Reed Smith reports that AI coding assistants can turn a short gameplay clip into a playable copy in hours, with clones of indie prototypes appearing on storefronts within days. Unprotected AI assets make that worse, because competitors are free to copy them. On moderation, the burden has moved onto developers through Steam’s guardrail descriptions and player reporting for live content.
The decision a studio has to make before choosing any model
Before picking a model, a studio has to decide which content must be reproducible. We would put anything a designer balances, QA signs off or a lawyer has to vouch for in the offline pipeline, generated, reviewed, registered and frozen. Anything left to runtime has to justify its VRAM, its latency budget, its Steam guardrail statement and its EU disclosure. What the evidence does not yet show is a reliable way to replay and regression-test runtime generative content at scale.
