I tried playing with the new Sora 2 model this week. I am not a huge fan of AI-generated art and videos (side note, see my blog’s AI manifesto), but I like to be aware of their capabilities.

My main “test” I try out with pretty much all AI image and video creating tools is to prompt them to render “a moose playing Go in a park while drinking boba.” Kind of like my own version of pelicans on a bicycle. It… never works.

It’ll get close, kind of, and I will say, Sora 2 was better than previous attempts with video. But, I will not show you the video results, because the results genuinely just kind of made me uncomfortable. I will show you images, but first, let me explain.

I think this prompt specifically has some challenges that AI has yet to overcome:

• Moose are kind of weird animals.
• The grid on a go board is 19x19 and counting is very hard for AI tools.
• Go pieces look an awful lot like tapioca balls in a boba cup.
• A small problem I can forgive, but a very real problem, is that a natural game of go has the same number of pieces on the board in each color, and some arrangements of pieces just don’t exist in a real game.

In every single attempt I’ve tried (I have tried this with pretty much every video and image generation tool you can think of), it has at least one of these problems, if not most of them:

• The moose isn’t a moose, or doesn’t stay a moose (Sora 2 transformed the moose into… some kind of scary hairy blob on several occasions)
• The moose ears and antlers aren’t in the right spot (did you know that antlers are like giant “hearing aids” for moose? They’re like giant parabolic dishes for sound. So cool.)
• The moose is just nearby while a random man plays go instead
• The boba straw is jank in some way (Sora 2 had the straw shrink as the moose drank from it at least 3 different times)
• The go pieces are not the same sizes on the board (in most video generations, the pieces pulse in size? Which is weirdly unsettling.)
• The actual gameplay is super wrong on the go board (incorrect number of pieces, non-sensical placements, pieces just “on the board” instead of in proper positions on the lines of the grid)
• The go board is a weird shape (in videos, it’s often concave like a bowl, and the grid shifts around)
• There’s no bowls of go stones on the side of the board (or anywhere)
• The moose has sunglasses on (?) and the reflection in the sunglasses doesn’t match the board
• There are go pieces in the cup of boba, or the boba ends up being the go stones
• The game isn’t actually go

I do massage the prompt, like sometimes I’ll give it some more details or iterate on it, but alas, these problems are still pretty consistent. Which I’m okay with! It’s a good test!

Here’s some examples of outputs I’ve gotten (first one being a snapshot of a Sora 2 video that almost looked good, until the moose turned into a nightmare creature, the straw floated around the go board, and the pieces moved themselves into a corner):

Before you say, “Now, Cassidy, you’re being a bit strict with these AI tools, these are pretty dang close. One might say, even, that they are okay.” Sure, sure. But, I counter: no human artist would ever make these mistakes. If I asked an artist to draw/paint/create a moose playing go in the park while drinking boba, the straw would be in the cup. The go board would be valid. A man would not be drinking the boba. The moose would be a moose.