AI NPCs: The Road from Scripted Characters to Open-World AGI

Non-player characters (NPCs) exist in nearly every game, yet players are increasingly unsatisfied with how they behave. Recent demonstrations by NVIDIA at CES 2024 and Microsoft's partnership with Inworld AI show that large language models are making NPCs capable of open-ended, real-time conversation. This shift has pulled the gaming industry, and many Chinese AI companies, toward NPCs as one of the most practical arenas for AI deployment.

Most NPCs are driven by state machines or behavior trees, which follow predefined rules. Surveys by Inworld AI cited by Microsoft Xbox found that 84% of players consider NPCs important and 79% interact with them, yet 52% complain that NPCs simply repeat the same lines. Rule-based systems are also exploitable: once players identify a designer's pattern, they can game the logic, as seen in the well-known strategy of boss 'foot-trapping' in Sekiro: Shadows Die Twice.

Richly written NPCs demand enormous resources. Red Dead Redemption 2 features over 1,000 NPCs across six chapters and more than 100 missions, with development taking nearly eight years and costing close to $500 million. Ambitious designs such as Ken Levine's 'Narrative Legos' concept show how hard it is to build flexible storytelling from scratch, and delays are common across the industry. Rising player expectations collide with rising production costs, which is why generative AI looks attractive.

NVIDIA's Avatar Cloud Engine (ACE), which combines Audio2Face and Riva with LLMs, is designed to deliver real-time, lip-synced NPC responses. Microsoft Xbox has partnered with Inworld AI, whose Character Engine integrates multiple models to simulate a human-like cognitive system. In China, Baichuan released Baichuan-NPC, its first vertical application model, while MiAO (founded by former Giant Network CEO Wu Meng), RPGGO, and Kunlun Wanwei's Play for Fun studio are building AI-driven games and NPCs. Baichuan's Richard notes that games are often the first landing ground for new technology and a proven monetization path; China's game market reached 302.96 billion yuan in 2023.

NVIDIA's pipeline converts player speech to text, feeds it to a large model to generate a response, synthesizes speech, creates lip-synced animation, and renders the character back into the scene. Inworld builds a multimodal engine and a no-code studio where developers describe characters in natural language. MiAO's LARP (Language Agent for Role Play) framework uses a set of smaller, domain-tuned language models coordinated in a modular architecture for memory, decision-making, and learning from interaction.

Making NPCs convincing requires strong base abilities in general intelligence, character knowledge, dialogue, plot performance, and logical reasoning, best strengthened through targeted high-quality pretraining data. A central problem is out-of-character (OOC) behavior, where a model slips back into a generic assistant persona or says things inconsistent with its role. Baichuan-NPC addresses this with chain-of-thought data construction and chain-of-thought alignment, so that both reasoning and resulting actions resemble human behavior. MiAO balances personality and world rules through an external database of shared world knowledge plus a 'collective consciousness' that propagates content generated during runtime.

Integrating LLMs into mature development pipelines is difficult. Game assets such as models, animations, and world states are non-textual and must be translated into semantic representations. Open worlds are unbounded, so NPCs must adapt to unfamiliar environments and accumulate long-term memory that supports continuous growth; current long-context understanding remains limited. High-quality model and API costs also exclude smaller studios. Industry observers, including MiAO's AI lead, note that AI researchers and game developers often misunderstand each other, and the field lacks complete system-level solutions rather than just stronger models. Designers should also avoid making NPCs too human at the expense of playability.

Games have long served as AI's experimental field. They offer clear win conditions and metrics for testing and iteration, cheap virtual data for supervised and reinforcement learning, and diverse environments for studying cooperation and competition among agents. Tencent AI Lab argues that mastering long-term strategic planning and collaboration in complex strategy games could represent the frontier of multi-agent decision-making. Game performance also serves as a practical showcase of a company's application capabilities.

Beyond NPC dialogue, the ultimate direction is the open world. RPGGO is building a generative engine that turns a single text prompt into a playable, multiplayer role-playing world, a concept it calls text-to-openworld. Kunlun Wanwei's Club Koala includes a zero-code editor so ordinary players can build their own games and NPCs with self-awareness, with a beta planned for March. Investors at Monolith Capital argue that AI expands the supply of gameplay, potentially shifting games from professionally generated content (PGC) to user-generated content (UGC).

Richard of Baichuan proposes four levels of AI development: L1, the model, with basic generalization; L2, the agent, with tool use, planning, and long-term memory; L3, the human-like agent, with persistent state and emotionally varied responses to events; and L4, the social human-like agent, which forms relationships and organizations with other agents. Most companies remain at the model stage. Sam Altman has suggested that future GPT generations will handle a growing share of human tasks and that AGI may arrive in the near future, though the route remains uncertain.

AI NPCs are becoming the focal point of AI application in gaming because they answer pressing needs: more engaging characters, lower development costs, and a testbed for general intelligence. Realizing truly open worlds would mean AI that reasons, acts, and cooperates much like humans, a projection of AGI into virtual environments. Technological progress is only the beginning; many questions about direction and responsibility still require human judgment, and the destination, as RPGGO's Li Jiaying puts it, is to push 'Magic' to its limits through technology.