Evolution of Dynamically Generated AI Games

One of the most exciting applications of artificial intelligence, especially for the general public, is generative AI. While for everyday users, it is most commonly associated with generating text and images, for game studios, it opens up wide possibilities: dynamically creating characters, dialogue, environments, maps, and entire levels.

In this article, we take you on a journey of a still largely unexplored genre of games built around open worlds, endlessly unique missions, and gameplay that constantly evolvesAI Dungeon (2019)

One of the first AI-generated games is a board-style game, AI Dungeon, a text-based adventure in which the game responds to your prompts in a chatbot style. Players can create any scenario they imagine, from epic quests to modern-day stories, and the AI dynamically adapts the narrative, making each playthrough unique.

Oasis (2024)

Oasis is a Minecraft clone powered entirely by AI, rather than a traditional game engine. Oasis generates each frame based on previous frames and the player’s inputs. Trained on millions of hours of gameplay, it can simulate actions like walking through a blocky world, interacting with objects, and breaking blocks in real-time. While it demonstrates the new concept very well, it can feel glitchy, blurry, and as soon as you turn around, it can forget the surroundings created beforehand, and create something totally else, or blur it out.

AI Doom (2024)

While Oasis is trained on recordings of human gameplay sessions, “ AI Doom ” by GameNGen was trained on gameplay videos of an AI agent rather than human players.The AI Doom generates gameplay that remains consistent over time, preserving key game states, such as ammo count, player health, doors, and enemy positions. In large part, however, this improvement is also due to the choice of game setting. Doom is a closed-environment game with fixed maps and predictable mechanics, which makes it easier for an AI system to model and maintain consistent game states.

Game Gen X (2024)

Trained on data from over 150 different games, it can generate worlds from text or video prompts, allowing users to specify scenarios like “an ice magician” or “an angel watching over a holy kingdom.” Game Gen X supports camera perspective control, dynamic actions such as driving or flying, and real-time instruction-based modifications to the environment. While interactive and visually impressive, its consistency is limited to the short term, as the AI struggles to maintain persistent world states beyond a few seconds. It also exhibits biases toward Cyberpunk style, probably because of the training sample. This stylistic lean may be very appealing for many gamers, since it’s a very popular theme among gamers, way back to the 80s. It’s very big in the online casino games as well, with gems like Cosmic Spins being popular in a big part because of its sci-fi thematics and visual identity.

Genie 2 (2024)

Genie 2 marked a major step forward in AI-generated games by allowing users to create small playable environments from a single image or text prompt. Developed by Google DeepMind, the model could generate interactive scenes where players move around using keyboard and mouse inputs. Compared with earlier systems, it introduced object interactions, simple non-player characters (NPCs), and, most importantly, the ability to remember parts of the world that were temporarily off-screen. This allowed the environment to remain consistent for up to about a minute, making the generated worlds feel more stable and believable than earlier AI game simulations.

Genie 3 (2025)

Building on the progress of Genie 2, Genie 3 significantly expanded the concept of AI-generated worlds. Instead of only generating short playable snippets, the system can simulate persistent environments with a longer memory of three to five minutes. With what researchers describe as “world memory”, the model stores the positions of objects and player actions so that changes in the environment remain when the player looks away and returns. This allows for consistent gameplay and the creation of explorable environments generated from images or prompts. Genie 3 was trained on petabytes of Google Street View imagery, providing it with a massive dataset of real-world environments across cities, forests, coastlines, and more, which helps make generated worlds feel spatially coherent and visually realistic. Genie 3 is available for free use, so many great tryouts can be found online, people testing it on unique ideas, and trying to recreate games like Fortnite, GTA V, and so on.

Conclusion

One YouTube user jokingly commented, AI might end up generating Grand Theft Auto VI before Rockstar Games releases it. While this sounds like an obvious exaggeration, the AI really advances not in steps, but in leaps, which is clear when we remember that not so long ago, we were impressed with how GPT was able to generate code for a simple 2D Snake game. It is safe to say that we’re up for an exciting future in game development and the AI space in general.