A demo shows you the happy path by definition, that's what makes it a demo. Here's what actually goes wrong once you're past it.
Silent logic gaps
Not a crash, a quietly wrong rule: a coin that's visually there but doesn't add to the score, a door that opens but doesn't trigger the next event. These pass a "does it crash" test cleanly and only show up when a human actually plays through with attention, which is exactly why an automated playtest loop that checks specific outcomes, not just "no errors," matters so much more than it sounds like it would.
Context drift across edits
Ask for five changes in a row and something you approved on change two can quietly break by change five, because each edit is applied against the current state without perfect memory of every earlier decision's intent. This is the single most common complaint from people who've used these tools for more than a few minutes, and it's a real, current limitation, not user error.
Generic-feeling output at default settings
Ask for "a platformer" with no other detail and you'll get the average of every platformer in the training distribution, which is to say, something that resembles a hundred other generated platformers. This isn't a bug, it's what happens when a prompt under-specifies what makes an idea distinct, and the fix is the same one that fixes generic AI art: name the specific choices that matter to you instead of leaving them all open.
Difficulty and pacing that's technically fine and actually flat
Covered in more depth elsewhere on this blog, but worth repeating here: a generator can guarantee a level is completable. It has no reliable sense of whether the difficulty curve feels good, and this is consistently the single biggest gap between "this works" and "this is fun."
Asset mismatch under time pressure
Art and code generated in separate passes can drift apart, a sprite's hitbox doesn't quite match its silhouette, a sound cue fires a frame off from the animation it's tied to. Individually tiny, collectively they're what makes a generated game feel slightly "off" even when nothing is technically broken.
Why this list is worth publishing, not hiding
Every one of these is something a serious tool in this category is actively working against, with playtesting loops, better context handling, and tighter defaults. Pretending the failure modes don't exist doesn't make the tools better, it just means you find out about them yourself, later, with less patience for it.