AI NPCs: how they work, the tools to build them and games that use them

An AI NPC is a non-player character whose lines, and sometimes its decisions, are generated while you play by a language model, instead of being picked from dialogue a writer scripted in advance. In practice it's a pipeline: the player's speech is turned into text, a model writes a reply in character, and that reply is spoken aloud and animated. As of October 2026, a handful of released games use them, including inZOI and Where Winds Meet. Krafton has run a public beta of an AI squadmate in PUBG: BATTLEGROUNDS, and several SDKs let small teams try the same thing in Unity, Unreal and Godot.

AI NPCs vs. traditional game AI

"Game AI" is much older than large language models, and the two terms get mixed up.

Classic game AI: behavior trees and state machines

Most NPCs run on hand-built logic. A finite state machine switches a guard between states such as patrol, chase and attack. A behavior tree is a branching set of tasks; Epic's Unreal Engine documentation describes one that picks which branch of logic to run, paired with a "Blackboard" that stores the facts it decides on. This kind of AI is fast, predictable and cheap, but it doesn't write new sentences.

Scripted dialogue trees

Most RPG conversations are dialogue trees: every line and player choice is written ahead of time. Writers control everything, but players can only say what's on the menu.

LLM-driven NPCs

An LLM NPC lets the player type or speak freely and generates a reply on the spot. The writer's job moves from writing every line to writing the character: who they are, what they know, what they want and what they must never say. Ubisoft made the same point about its NEO NPC prototype in a March 19, 2024 press release, saying each character's personality, history, goals and emotions had to be worked out in detail and built into the model.

Shipped systems often combine both. Krafton's PUBG Ally keeps a behavior tree for reflexes that must run at game speed and hands reasoning and speech to a small language model, according to an NVIDIA developer blog post on how the character was built.

How an AI NPC works

Most tools break the job into the same stages.

Character definition

A character is a structured prompt: personality, backstory, speaking style, goals and knowledge. Hosted tools offer editors for this; with local libraries you write the system prompt yourself.

Dialogue generation

The language model takes the character definition, recent conversation and current game state, and writes the next line. Cloud tools call a hosted model. On-device tools run a small model on the player's GPU or CPU, often in the GGUF format used by llama.cpp.

Memory and knowledge

Models forget anything outside their context window, so tools add conversation history, long-term memory and retrieval (RAG), which searches your lore and inserts relevant facts into the prompt. Convai's docs list a "Knowledge Bank" and long-term memory. NVIDIA says PUBG Ally pulls live engine state through tool calls rather than relying on what it remembers, which keeps its answers about the match accurate.

Guardrails

Guardrails keep the character on-script and safe: rules in the prompt, content filters, topic limits and checks on the output. Inworld's docs say its current runtime includes safety checks. Virginie Mosser, narrative director on Ubisoft's later Teammates project, said she still writes the story and personalities, then sets "fences" so the characters can improvise without leaving the game's lore.

Speech-to-text and text-to-speech

Voice NPCs need speech recognition on the way in and a synthetic voice on the way out. NVIDIA's Mecha BREAK demo in August 2024 ran OpenAI's Whisper for recognition on the player's PC but used ElevenLabs in the cloud for the voice. Krafton paired NVIDIA's speech recognition with its own in-house TTS model for PUBG Ally.

Animation and lip sync

A talking 3D character needs its face to match the audio. NVIDIA's Audio2Face turns audio into facial animation and was used for Ubisoft's NEO NPCs. Convai has lip sync for Unity and a "Face Sync" component for Unreal.

Actions

Through function calling, the model picks from a list of game actions, such as follow or change a loadout, and the game carries them out. In NVIDIA's Mecha BREAK demo, players could ask the mechanic to change their mech's paint job. Convai calls this feature "actions," and LLMUnity supports function calling by using a grammar to limit the model's output to the function names you allow.

Tools and SDKs for AI NPCs

Here's what each tool says it does, based on its own docs and pages as of October 2026. AIGamer hasn't tested these tools.

Inworld has retired its older Character Studio in favor of its Agent Runtime SDKs. Player2's Godot plugin, announced on June 20, 2025, needs players to install and run the Player2 app.

We left out Charisma.ai, which used to offer Unity and Unreal kits. Its homepage now pitches conversational AI for training and brand campaigns and doesn't mention games, and its Unity and Unreal SDK repositories on GitHub haven't been updated since September 2025.

If you're choosing an engine first, our AI game engines guide compares the options. For building the rest of the game with AI help, see vibe coding games and the best AI game makers.

Games and demos that use AI NPCs

inZOI (Krafton)

inZOI's "Smart Zoi" characters arrived with the game's early access on March 28, 2025, as an optional experimental setting. NVIDIA says they run on a 0.5-billion-parameter Mistral NeMo Minitron model on the player's GeForce RTX GPU. A Smart Zoi reacts to what's around it according to its personality and reworks its daily schedule based on what happened to it, and players can watch its thoughts and redirect it with typed text.

PUBG Ally (Krafton)

PUBG Ally is an AI squadmate for PUBG: BATTLEGROUNDS that you talk to by voice; Krafton says it can discuss strategy, fetch items you ask for, help revive teammates and fight on its own. Krafton announced it on October 30, 2025, with support for English, Korean and Chinese. Its first public test was Ally Duo, a two-week global beta in PUBG's Arcade mode on the Sanhok map that began on June 17, 2026 (Krafton gave July 1 as the end date; NVIDIA said June 30). NVIDIA's developer blog says Ally uses a quantized 2-billion-parameter model and the game runs on GPUs with as little as 8GB of total VRAM. As of October 2026, Krafton hasn't announced a full release; it said it will keep refining the technology using feedback from the beta.

NARAKA: BLADEPOINT (NetEase)

NVIDIA says an AI teammate came to NARAKA: BLADEPOINT MOBILE PC VERSION on March 27, 2025, running locally on the player's PC. It can join your party in Free Training, find items you ask for, swap gear and suggest skills to unlock.

Mecha BREAK (Amazing Seasun Games)

At Gamescom in August 2024, NVIDIA showed a voice-controlled mechanic in Mecha BREAK running NVIDIA's Nemotron-4 4B model on-device. Players could ask about objectives and which mech suited a job. That was a demo, and we couldn't confirm that the feature made it into the finished game.

Ubisoft NEO NPCs and Teammates

Ubisoft showed NEO NPCs at GDC in March 2024, a prototype built with Inworld's character engine and NVIDIA's Audio2Face. On November 21, 2025, it followed up with Teammates, a playable first-person shooter experiment with voice-controlled squadmates, tested in a closed playtest by a few hundred players. Neither is a released game. For Ubisoft's wider generative AI plans, see our report on its Motherbrain unit.

Where Winds Meet (Everstone Studio, published by NetEase Games)

This wuxia open-world game, released worldwide on November 14, 2025, has NPCs that players can chat with; its Steam page says the AI-driven NPC text and voice chat responds to players in real time. Kotaku reported on December 3, 2025, that players were already gaming the system: one got a quest to end by turning everything an NPC said back into a question. PCGamesN reported that the chat NPCs are flagged in the game and players can skip them entirely.

Suck Up! (Proxima)

In Suck Up!, you play a vampire who has to talk neighbors into letting you inside by speaking into a microphone. The Steam page lists a release date of October 1, 2025, says the game connects to OpenAI's ChatGPT to generate content, and lists a broadband connection as a requirement.

Practical problems

Latency

Recognition, the model's reply, speech synthesis and animation each add delay. Krafton says running PUBG Ally entirely on-device keeps its latency very low, and NVIDIA's blog says cutting out the trip to a server made response times more predictable. The blog also describes keeping the stable part of the prompt the same from turn to turn so the model can reuse cached work.

Cost per conversation

Cloud NPCs are billed per use. Inworld's pricing page charges for speech by the character, recognition by the hour and the model by the token, so a long voice chat costs more than a short typed one, and NPCs talking to each other cost money even when no player is listening. On-device models move the cost to the player's hardware.

On-device vs. cloud

Local models are cheaper to run at scale, work offline and keep player speech on the device, but they must be small. NVIDIA says its 2B Minitron model fits in as little as 1.5GB of VRAM, and that memory competes with your game's graphics. Cloud models are stronger but need a connection and a budget per player.

Keeping characters on-script

Players will try to break your NPCs, as Where Winds Meet showed. Don't let the model alone decide whether a quest is complete; check it in game code.

Content moderation

Free-text input means players can steer characters toward offensive or sexual content. Kotaku's December 2025 report included a Where Winds Meet player who shared an erotic roleplay with one of the game's NPCs. Plan filters, player reporting and your age rating before launch, and log conversations so you can see what went wrong.

Voice actors and SAG-AFTRA

On May 19, 2025, SAG-AFTRA filed an unfair labor practice charge against Llama Productions, the Epic Games subsidiary behind Fortnite, over an AI-generated Darth Vader voice, the Associated Press reported. The union said the company swapped in AI voices for work its members would otherwise do, without notifying or bargaining with the union first. SAG-AFTRA members ratified a new video game contract on July 9, 2025, with 95.04% voting yes, ending a 320-day strike, TheWrap reported. Under the deal, companies must explain how AI will use a performer's voice, likeness or movement and get written consent. It also sets pay for "real-time generation," such as a voice-replica chatbot in a game, at 7.5 times minimum scale or more.

Design tips

• Constrain the world. Krafton limited PUBG Ally's beta to one map and one mode, which NVIDIA says made the problem far easier.
• Give the NPC a job. A mechanic or squadmate with a clear goal stays in character better than an open-ended chatbot.
• Ground replies in game state. Pass in facts from the engine instead of trusting the model's memory.
• Make actions a fixed menu. Let the model choose from functions you define and validate each call.
• Test with lots of players early. NVIDIA says Krafton tuned Ally with feedback from more than 1,000 players.
• Have a fallback. If the model times out or the player is offline, fall back to scripted lines.

FAQ

What is an AI NPC?

It's a non-player character whose dialogue, and sometimes its actions, is generated during play by a language model instead of being fully scripted. Many also let players talk to them out loud.

What games have AI NPCs?

Released examples include inZOI (in early access since March 2025), NARAKA: BLADEPOINT MOBILE PC VERSION, Where Winds Meet and Suck Up!. PUBG Ally ran as a public beta in June 2026.

How do I add AI NPCs to Unity, Unreal or Godot?

For Unity, options include Convai, Inworld's Unity Runtime SDK, Player2 and the open-source LLMUnity. For Unreal, NVIDIA ACE, Inworld and Convai have plugins. For Godot, there's Player2's plugin and the open-source NobodyWho. AIGamer hasn't tested any of them.

Can AI NPCs run offline?

Yes, if you use a small local model. LLMUnity and NobodyWho run llama.cpp models on the player's machine, and NVIDIA ACE targets on-device inference on RTX GPUs. Cloud services such as Inworld and Convai need a connection.

Do AI NPCs replace voice actors?

Not under union rules without consent. SAG-AFTRA's 2025 video game contract requires consent and disclosure for AI replicas of a performer's voice and sets higher minimum pay for real-time generated dialogue.