- Type
- Application domain
- Used for
- NPC behaviour, content generation, QA, matchmaking, prototyping
- Key techniques
- Behaviour trees, Monte Carlo tree search, reinforcement learning, small language models
- Emerged
- Pathfinding in the 1960s; generative AI from 2023
- Related
- Reinforcement Learning, Generative AI, World Models
- Type
- Application domain
- Used for
- NPC behaviour, content generation, QA, matchmaking, prototyping
- Key techniques
- Behaviour trees, Monte Carlo tree search, reinforcement learning, small language models
- Emerged
- Pathfinding in the 1960s; generative AI from 2023
- Related
- Reinforcement Learning, Generative AI, World Models
Background
Game AI predates modern machine learning. A-star pathfinding dates to 1968, and finite state machines, steering behaviours and behaviour trees became the standard way to script enemy and NPC reactions. DeepMind's AlphaGo victory in 2016 brought Monte Carlo tree search combined with neural networks to mainstream attention, while procedural generation — most famously No Man's Sky's mathematically constructed planets — showed content could be grown rather than authored.
The generative wave arrived in 2023-2025. The Game Developer Conference's 2025 State of the Game Industry survey reported that more than half of development companies were using generative AI in production for content, testing and design[2], and around 4,000 games carrying AI disclosures appeared on Steam during 2025, with storefront disclosure now mandatory.[1]
Key Concepts
- NPC dialogue and behaviour. Small language models running on-device give non-player characters unscripted conversation; NVIDIA's ACE suite provides the pipeline for conversational NPCs, demonstrated with on-device models in the Dead Meat project at GDC 2025, while Ubisoft's experimental Teammates initiative targets companions that react to speech, mood and gameplay.[3]
- Prototyping and world models. Microsoft's Muse, a World and Human Action Model trained on gameplay footage, generates playable action-and-response sequences for ideation, followed by a real-time playable demonstration, WHAMM.[4]
- Production support. Concept art, textures, dialogue drafts, localisation and code assistance, alongside automated regression-test bots that play builds to catch defects.
- Live operations. Matchmaking, dynamic difficulty, anti-cheat anomaly detection and player-support chat.
>See Also
🇲🇾 Relevance to Malaysia: Malaysia hosts a games industry built on art production, co-development and outsourcing — studios such as Lemon Sky and Kuala Lumpur's Metronomik (No Straight Roads) — plus an animation and games ecosystem supported by MDEC under Malaysia Digital incentives. AI asset pipelines raise the competitiveness of Malaysian outsourcing houses, while international publishers increasingly contractually require disclosure and originality warranties for AI-assisted work, making documentation a commercial requirement rather than a courtesy. Game-development and animation programmes across Malaysian institutions feed the talent pipeline that MDEC and HRD Corp upskilling aim to extend. Studios running online services and telemetry must treat player data as personal data under the Personal Data Protection Act, including when it is used to train matchmaking or anti-cheat models.
References
- ↑AI and Games. (2026). 10 Predictions for AI in Games for 2026. https://www.aiandgames.com/p/10-predictions-for-ai-in-games-for
- ↑Business Insider. (2025). AI is speeding up video game development. https://www.businessinsider.com/ai-video-games-development-faster-production-new-improved-gameplay-2025-8
- ↑NVIDIA Developer. ACE for Games. https://developer.nvidia.com/ace-for-games
- ↑Microsoft Research. WHAM (World and Human Action Model). https://www.microsoft.com/en-us/research/project/wham/