AI Can Code a Complex Video Game, But Rarely Plays It Well 2026

Daniel Harrolds
AI Can Code a Complex Video Game, But Rarely Plays It Well
This page may contain affiliate links.

AI can code a complex video game from a simple prompt, but very few models can actually play it well. That's the paradox at the heart of recent research by Julian Togelius, director of the Game Innovation Lab at New York University and co-founder of testing company Modl.ai. In an interview, Togelius explained why large language models (LLMs) can generate a functional clone of a classic game like Asteroids yet fail to surpass the first level of their own creation.

The reason isn't that games are inherently difficult in the abstract. It's that the real world obeys the same physical laws everywhere, while video games each have their own unique rules and mechanics. This fundamental difference explains why AI excels at coding but struggles at playing.


Furniture with Meaning.

Find Original Home Decor sparking conversation at Grand Goldman Shop.


Why AI Excels at Coding Video Games

Togelius defines programming from a structural standpoint: it's a very well-designed game. Each line of code comes with a clear statement, a verifiable success criterion, and feedback on potential failures. The program indicates exactly where and why it failed.

ADVERTISEMENT

LLMs have been trained on massive amounts of code and fine-tuned through reinforcement learning to solve exactly this type of problem. In terms of task structure, programming is an exceptionally "well-posed" game, as Togelius puts it. That's why tools like Cursor or Claude can generate a working game clone from a single prompt.

The Paradox: Making vs. Playing

The paradox is striking: with tools like Cursor or Claude, a prompt generates a functional clone of a classic game like Asteroids. However, that same system wouldn't pass the first level of its own creation. This gap between making and playing highlights a fundamental limitation of current AI models.

Togelius has been investigating this phenomenon for months, and his findings shed light on why AI can't easily transfer its coding prowess to gameplay. The issue isn't about raw intelligence but about the nature of the tasks themselves.

Why Playing Is a Different Challenge

Playing a video game requires real-time decision-making, adaptation to dynamic environments, and understanding of game-specific rules that aren't always explicitly stated. Unlike coding, where success is clearly defined and feedback is immediate, gameplay often involves ambiguous goals and delayed rewards.

ADVERTISEMENT

Moreover, each game has its own physics, controls, and objectives. An AI that masters one game can't necessarily apply that knowledge to another. This lack of generalization is a major hurdle for AI in gaming. The background to this story is set out in Gemini 4 Neon: Google's New AI Model for Coding and Complex Tasks. A related thread runs through AI Report Predicts the Decline of Video Game Consoles. Earlier coverage in AI-Powered Video Game Mashups: The New Trend Explained shows how the argument developed.

What This Means for AI Development

This research has important implications for AI development. It suggests that while LLMs are powerful tools for structured tasks like coding, they still struggle with open-ended, real-time problems that require common sense and adaptability. A closer look at the same subject appears in Epic Games Embraces AI for Marvel and Star Wars Game Development. Anyone following the wider picture can start with Epic Games Layoffs Lead to AI-First Development for Disney Game. The same theme surfaces in Claude AI Powers PS1 Development Kit PSoXide for Indie Games, which takes a different angle.

For game developers, it means AI can assist with coding but isn't ready to replace human playtesters. For AI researchers, it points to the need for models that can learn and adapt to new environments quickly, much like humans do. Context on how this played out before sits in 18-Year-Old Developer's $200 Steam Game Earns $1.3 Million. A companion piece, USMNT Savoir Complex: Moving On Without Pulisic, approaches the same subject from another angle.

Key Takeaways

  • AI can generate functional video games from prompts but often fails to play them well.
  • Programming is a well-posed game with clear success criteria, making it ideal for LLMs.
  • Playing games requires real-time adaptation and understanding of unique game rules.
  • Each game has its own physics and objectives, limiting AI's ability to generalize.
  • This gap highlights the difference between structured coding tasks and open-ended gameplay.

Comparing AI Performance: Coding vs. Playing

Aspect Coding a Game Playing a Game
Task Structure Well-posed, clear success criteria Ill-posed, ambiguous goals
Feedback Immediate and specific Delayed and often unclear
Generalization High (code patterns transfer) Low (each game is unique)
Real-time Demands Low (can iterate) High (requires quick decisions)
Example Generating Asteroids clone Beating level 1 of Asteroids

FAQ

Can AI really code a complex video game?

Yes, AI models like Claude and Cursor can generate functional clones of classic games from simple prompts. However, they often struggle to play those games effectively.

Why can't AI play the games it creates?

Playing requires real-time decision-making and adaptation to unique game rules, which are not as well-defined as coding tasks. AI models are trained on structured problems like code, not dynamic gameplay.

What does Julian Togelius say about AI and games?

Togelius explains that programming is a well-posed game with clear success criteria, making it ideal for LLMs. Playing, however, is ill-posed and requires common sense and adaptability that current AI lacks.

Will AI ever be good at playing video games?

It's possible with advances in reinforcement learning and generalization, but current models are far from human-level gameplay. Research is ongoing to bridge this gap.

ADVERTISEMENT
ADVERTISEMENT
Daniel Harrolds

Author

Daniel Harrolds

With a career spanning four decades, Daniel is almost a library in the field of precious metals investing and Gold IRAs. His insightful strategies and pragmatic results-oriented approach make him a resource in safeguarding wealth, and financial foresight.


SPONSORED

Get Lifetime Access to the Lastest Movies, with Exclusive Offers & Free Express Order Delivery.

ADVERTISEMENT