The honest answer depends entirely on one question: does your game have a clear, checkable definition of "working"? Pong does. An open-world RPG with emergent systems does not. That single distinction predicts almost everything about whether AI can build it with minimal supervision.
Where "yes, mostly" is true today
Small, rule-bound games with an objective pass/fail condition, a platformer that needs to be completable, a puzzle game where a solution either exists or doesn't, a runner with a clear scoring system, are genuinely within reach of AI-driven generation with automated testing in the loop. The model writes code, an automated playthrough checks whether the level is beatable and the game doesn't crash, and it iterates until the check passes. This loop is exactly what makes it reliable: the computer, not a human, is verifying the output.
Where it's still "no, not unsupervised"
Anything whose correctness is a judgment call rather than a pass/fail check, is this boss fight fun, does this economy feel fair, is this level paced well, can't be automatically verified the same way. A model can generate a boss fight that technically works and is dull. No test catches "dull." That's still a human call, and it will be for a while.
The part that surprises people
The bottleneck usually isn't code generation, models are decent at writing a platformer's movement code. It's the surrounding infrastructure: rigging that code to actually run, catching the silent failures (a coin that's unreachable, a sound that never triggers), and fixing them without a human reading every line. That loop, generate, run, observe, fix, repeat, is the actual hard engineering problem, and it's the part most "AI makes games" demos skip by showing you only the happy path.
What this means if you're deciding whether to trust one
Ask what happens when it's wrong. If there's an automated check that catches the failure and retries, that's a real pipeline. If the answer is "it usually works," that's a demo, not a product you can build a business on.