"AI game maker" now describes at least three unrelated kinds of software. People try one, decide the whole category doesn't work, and walk away. Usually they just picked a tool built for a different job. Here's how to tell them apart before you spend a weekend.
Category 1. Assistants inside a professional engine
A panel bolted into Unity, Unreal or Godot that generates code, textures or animations while an experienced developer works.
Built for: people who already ship games and want to go faster.
Where it breaks for beginners: it doesn't remove the engine. You still face the import settings, the collision layers, the export configuration. If the engine was the obstacle, an assistant inside the engine doesn't move it.
Category 2. Browser generators
A text box on a website. Describe a game, get a playable link, share it. Fast, and good fun the first time.
Built for: a playable idea in an afternoon, and for sharing.
Where it breaks: the output is a web page. If your goal is a product you can sell, a web page is a long distance from a signed executable or a mobile app package. The other wall shows up around the fifth edit, when fixing one thing quietly breaks another.
Category 3. Desktop tools with an engine inside
A program you install. A real engine ships inside it, and you drive it by conversation rather than menus. This is the newest category, and the one we're building, so read the next paragraph with that in mind.
Built for: people whose goal is a finished, sellable build.
Where it breaks: it has to earn an install, it's heavier than a browser tab, and it lives or dies on how good the underlying templates feel. A desktop wrapper around mediocre game logic is just a slower browser generator.
The failure mode all three share
Every one of these tools is good at producing code that runs. None of them can tell whether the result is fun. There is no compiler error for boring.
This is why the honest question to ask any AI game tool isn't "can it make a game?". They all can. It's:
Who tuned how the jump feels, and when?
A jump that works takes gravity and a velocity. A jump that feels good needs coyote time, jump buffering, variable height depending on how long you hold, and a landing that doesn't kill your momentum. That's a week of a human's attention, and no model reliably invents it from a prompt.
So there are two possible answers. Either a person tuned those behaviors in advance and the AI picks and configures them, or nobody did and you get something technically correct and strangely lifeless. You can feel the difference within ten seconds of playing, which is why you should always play the demo rather than watch the video.
The second question worth asking
What happens on the fifth edit?
Demos always show the first prompt, because the first prompt is the easy one. Real work is the fifth change, and the sixth, each one landing on top of things you already like.
Tools that regenerate the whole program each time degrade fast. You fix the enemy speed and lose the boss. Tools that keep the game as structured data and change one value hold up, because there's nothing else to accidentally rewrite. When you try any of these, don't judge the first output. Ask for five changes in a row and see what survives.
A checklist for evaluating any of them
| Ask | Why it tells you something |
|---|---|
| Can I play it right now, not watch a video? | Feel can't be faked in a trailer. |
| What comes out at the end: a link, or a file? | Decides whether you can ever sell it. |
| Does the project open without the tool? | Tells you if you're locked in. |
| Can I undo a change from three steps ago? | Reveals whether state is structured or regenerated. |
| Who owns what it makes? | Answer should be short and say "you". |
| Does it check its own output before showing me? | Separates a generator from a tool. |
Where the category stands today
Being straight about it: this technology is very good at 2D games with clear rules, and nowhere near ready for large 3D projects, systems-heavy simulation, or anything where the pleasure comes from the depth of the systems. Anyone claiming otherwise is selling a demo.
Inside the space it does work, though, the change is real. Somebody with an idea and no code background can now get to a playable, sellable 2D game, which was not true a few years ago at any price.
If that's where you are, the practical route is in this post, and what you'll face on the store side is in this one.