AI tools remove many coding, art, and modeling barriers, but making a good game still depends on aesthetics, cohesion, playtesting, and taste.
2
Runtime LLMs can give NPCs personalities and make decisions during play, creating games that change from one session to the next.
3
AI changes game development from a linear process into parallel iteration, while introducing difficult problems around debugging, scalability, tokens, and content safety.
Summary
Danielle An and David Hoe describe how AI is changing game creation at Meta. Basic games such as platformers and Tetris can now be generated quickly, which makes the initial novelty wear off when everyone produces similar results. AI can help people who lack coding, art, or 3D modeling skills, but polished output still requires aesthetic judgment, a coherent world, and feedback from real players. The talk then moves to runtime LLMs, where NPCs make decisions during play based on their personalities and goals. This can create multiplayer games that are different each time and can adjust the experience for individual players. AI also lets teams work across disciplines in parallel and make changes in hours or days rather than months. An is honest about the limits: runtime LLM design is still experimental, and systems built around changing models and prompts are hard to debug. Meta also has to address platform scale, token economics, and safety for generated content.
AI lets people make basic games quickly, but the results become repetitive
An says that by 2026 many people, including their children, friends, or family members, will have built a small game with Gemini. A couple of prompts can produce something recognizable, such as a platformer or Tetris, and sometimes it works immediately. The novelty fades when many people ask for the same kind of game, such as a platformer that looks like Mario. The generated games tend to look alike, and they do not always work. The new problem is therefore what comes after the first successful prompt, when everyone can make a weekend game.
AI removes skill barriers without removing the work of making a good game
An describes how AI can help people work around missing skills. A coder who cannot draw or make 3D models can use tools such as Meta Banana or Meshy to create assets. An artist who lacks coding skills can also start building a game without waiting for an engineer. That access does not automatically produce a good game. An and Hoe say games still need strong aesthetics, a consistent relationship between the user interface, story, and art, and playtests with real people who can say what works.
A single piece of key art can give AI-generated assets and gameplay a shared direction
Hoe presents a workflow from an art director named Dale. A creator can choose one key art image, such as the bear example shown in the talk, and use it as an anchor for other models. The image provides an art style that can guide the creation of game assets. It can also suggest gameplay and help an LLM keep information consistent across sessions. Hoe presents this as a way to shorten the distance between an automatically prompted game and a professionally made one.
Taste determines whether a polished game feels worth playing
Once AI can produce games that look reasonably polished, An says appearance alone will not decide which game succeeds. Taste means knowing whether people will enjoy the game, which group of people it is for, and why they would have fun with it. The creator has to make choices that give the game a specific feeling. This judgment remains necessary even when AI can produce the underlying code, art, interface, and other pieces quickly.
Runtime LLMs can make NPCs act according to personalities and goals
An introduces runtime LLMs as systems that change or direct a game while someone is playing it. Hoe describes a multiplayer game built in a few days, with four players competing while LLM-driven NPCs try to collect cubes. The creators give each NPC a personality, such as being a thief, honorable, or fast. During the game, NPCs independently decide whether to steal cubes, block another character, or kick someone. Those actions are not scripted in advance, so each session can develop differently.
AI can let teams iterate on game ideas in hours or days
An contrasts AI-supported development with the older waterfall process. Game work used to move from design to art, modeling, animation, and coding, which made upstream changes expensive after a decision had been committed. Teams working with AI can work in parallel and update a game in hours or days instead of months. That gives them more time for playtests and more chances to change the game idea before release.
A runtime game master could adjust difficulty and participation for individual players
An gives a personal example of how runtime LLMs might make games more welcoming. She says coordination is difficult for her, and in cooperative games she can feel that she is holding her team back. A game master powered by an LLM could adjust the experience so she can keep playing with her friends and have fun. The system could make a game more personal, although An says teams are still working out how to use runtime LLMs in ways that genuinely distinguish a game.
AI game platforms will have to handle changing behavior across the entire stack
Meta expects large numbers of gaming experiences to appear as creation becomes easier. An says agentic systems may operate from the user's prompt through runtime game decisions and the platform's work of serving, ranking, and delivering content. That creates less determinism throughout the system. Model upgrades and prompt changes can alter behavior, which makes debugging difficult for engineers used to testing a known codebase. She also names token economics and content safety as open problems, especially when runtime systems can generate images or other material for an audience.
"What it will separate a good game versus a bad game is still a lot of the typical stuff."05:06
Who should watch
You have made a simple AI-generated game and want to understand what would make the next version feel distinctive.
You work on game design, art, engineering, or playtesting and want examples of how AI changes collaboration and iteration.
You build game platforms or agentic systems and need to think about runtime behavior, debugging, token economics, or generated-content safety.
