A grounded survey of where LLMs actually stand in game development as of 2026. The biggest real-world impact is in the development pipeline — ideation, prototyping, and QA — not in gameplay itself. Dynamic LLM-powered NPCs show promise but face serious scaling, latency, and cost barriers. Automated QA agents using LLMs offer the clearest value today by exploring games in non-human ways. Asset generation remains immature for production use. The emerging pattern is redistribution of effort rather than replacement, with on-device inference flagged as a potential future inflection point that remains speculative.
Table of contents
Adoption is real—but unevenDynamic NPCs: promise vs realityThe real impact: the development pipelineAutomated QA: where AI already worksStructural limits (still very real)Where the industry is headingThe next potential shift: on-device inferenceSo where is the real impact today?TL;DRSort: