AI can't understand dice

1 min read Original article ↗

George Stocker got me looking at AI-generated dice and I’m fascinated/horrified. Here are some examples Grok gave me when I asked for “a pair of dice”:

image (1)
image (2)

And ChatGPT:

ChatGPT Image Jul 21, 2025, 07_22_08 PM
ChatGPT Image Jul 21, 2025, 07_16_57 PM

And DeepAI:

5c290473-801c-4871-bf9b-bdc44be25bc7
666ca28f-1072-4df3-bb71-9708ebf0d438

And OpenArt:

openart-image_vsL7OID2_1753151204724_raw
openart-image_tY7vAL6J_1753151218988_raw

Only a couple of the dice are potentially fair. Most have too many fives. Two dice have 3 fives showing. A more subtle problem is that opposite faces must add up to 7, so it should never be possible to see a 3 and a 4 at the same time. One of the faces can’t decide if it should be a 6 or a 4.

Fundamentally these dice live in the uncanny valley because they resemble real dice, but fail to reflect how dice are meant to function. Sure you could toss these objects, if they were 3D-printed. But you wouldn’t last 30 seconds in a game where the outcome depended on the numbers rolled.

image (3)

2

Finishing with that alleged backgammon board is a nice touch. :slight_smile: