It’s late 2025, and the buzz around "AI-generated games" has reached a fever pitch, especially since models like Gemini started making waves. We’ve been diving deep, trying out everything from early players like Rosebud to newer entrants like Alibaba’s Lingguang AI. The promise is tantalizing: create a game with just a few words. And honestly, for certain types of games, it’s surprisingly close to reality.
My own initial forays were pretty straightforward. Tools like Gambo, Rosebud, and even Google's AI Studio, while not exclusively for games, showed glimpses of what's to come. The real eye-opener, though, was Alibaba's Lingguang AI. The concept of typing "I want to make a Zuma game" or "a Tetris clone" and having a playable prototype emerge, sometimes with as few as 20 prompts, is genuinely impressive. The workflow is there, and it’s pushing the boundaries of what we thought was possible.
But here’s where things get a bit complicated. The narrative often gets inflated. When you hear "one sentence to make a game," it’s easy to jump to the conclusion that game development as we know it is about to be replaced. That’s a conversation we’ve been having a lot lately, and it’s worth unpacking.
From an engineer's perspective, tools like Cursor, Claude Code, and Copilot are game-changers. Give them a 20-30 word prompt, and they can whip up a simple 2D game. They often build from scratch using JavaScript, creating a basic engine first, then layering the game framework. It’s fantastic for demos, and they can even integrate with engines like Phaser for 2D or Godot. The caveat? There are still plenty of bugs.
Then there are the "Material Creation Platforms" (MCPs), like Pixso Lab. These tools, often integrated with the coding assistants mentioned, can generate art assets using techniques like style transfer. For 2D games, this is a huge win for maintaining visual consistency. For 3D, however, the results are still a bit… rough.
Finally, we have the end-to-end game platforms: Gambo, Rosebud, and yes, Lingguang AI. Gambo leans towards 2D, while Rosebud claims to be a 3D generation platform. Both often use Phaser for 2D and Three.js for 3D. What you get are playable prototypes that feel like they’re from the early 2000s – functional, perhaps even fun, but definitely raw. The amazing part? You can often generate these in about two minutes, complete with assets and music.
I remember one instance where the prompt was something like: "You are a genius game designer who understands what makes a good game. I want to make a Bubble Bobble, but I have no other ideas. Make it industry-leading, AAA quality." And it just… did it. The first version was surprisingly functional, bug-free, and looked the part. It was a bit of a mind-bender, honestly, seeing a complex game concept emerge from such a seemingly casual prompt.
This capability is deeply tied to the advancements in models like Gemini 3 and Claude 4.5. Unlike older models that would give vague answers to vague requests, Gemini 3, for example, goes through a lengthy "thinking" process. It uses recursive thinking to expand your request into a vast search space, then translates that into code. It’s like a super-high-level programming language, converting your intentions into executable logic.
However, we also hit walls. When we tried to add a specific mechanic – like the bubbles in Bubble Bobble descending with each shot to increase pressure – the AI struggled. Even with a clear prompt, translating that strict rule into engine logic and simulating the physics accurately proved difficult. It seems that while AI can grasp linear logic well (e.g., "if X > 500, then collision"), it falters with more nuanced, visually-dependent concepts, especially in 2D or 3D where spatial understanding is key.
It’s not just about the complexity of the request, but also how we describe it. The AI might generate a "cool effect" when asked for something vague, often manifesting as fireworks on every elimination. This works because it’s easier for the AI to generate something visually impressive from a fuzzy concept. But when you have a very specific visual outcome in mind, or require assets that go beyond code-generated effects, its understanding weakens considerably.
So, what are the actual product forms emerging from this? We're seeing a few categories. First, games built directly by large language models. Second, all-in-one generation platforms that offer games as one of their interactive entertainment options. Third, platforms that directly output finished games.
In terms of output, there are two main types of experiences. One is the "information stream" style, like Aippy, where each interaction is short, under 10 seconds, and requires minimal iteration. This is very much AI-native generation. The second type is what we tried with Bubble Bobble or Flappy Bird – games with a sense of established genre, aiming for multiple levels and deeper engagement, like what Rosebud offers.
How do these stack up? At its core, a game is an interactive experience. If we place "interaction depth" on a spectrum, at the shallow end, you have passive viewing like TikTok. Slightly more interactive are longer videos on platforms like Bilibili, where comments and likes are forms of engagement. Then you get to things like TikTok filters, which blur the line between short video and game. Here, the presence of a real person performing adds significant value – we’re drawn to the performer, not just the filter’s mechanics.
When this shifts to abstract game mechanics, the creator’s value becomes less clear. Move further along the spectrum, and you hit hyper-casual games, then platformers like Flappy Bird, or puzzle games like "Yang Le Ge Yang." Further still are traditional game genres: idle games, casual games, match-three, MMOs, and finally, the highly interactive AAA action games. This is the spectrum of "interactive engagement."
We’re not expecting AI to churn out a "Souls-like" game anytime soon. Frankly, I’m a bit pessimistic about AI fully replicating the human interactive experience. AI is a tool for production. Humans consume content because we have lived experiences, emotions, and physical needs. AI doesn't have that. Games, at their heart, have always been about simulating or exploring aspects of life. And that deep-seated human need for experience is something AI, for now, can only mimic, not truly embody.
