Cute or Creepy? · Showcase

The work

A judge for AI-generated animal drawings, with one surprising finding: cute is harder to define than creepy.

Jae wanted to build an eval for AI-generated animal drawings — a judge that could look at any image and label it "cute" or "creepy" with some confidence. She ran it on a few hundred generated animals. The judge got "creepy" right almost every time. It struggled with "cute." When she dug in, she realized the problem wasn't the judge. It was the category. Creepy has rules. (Too many teeth, uncanny eyes, wrong limb proportions.) Cute doesn't, really.

The thing itself · four animals, judged

Tap an animal. Read the judge's notes.

↑ Pick a specimen.

Jae's headline finding

"Creepy has a checklist. Cute has a vibe. My judge agrees that specimen #4 is creepy because #4 breaks four specific rules. My judge says specimen #1 is cute because... well, it has to guess. I think this is a big deal. It might mean some things are judgeable and some things aren't, and I can't tell the difference by looking at them. I'm going to keep going."

An eval for AI-generated animal drawings. The headline finding: cute is much harder to define than creepy. I agree. This project changed how I think about both words. Jae found an asymmetry at the center of aesthetic judgment that most grown-ups never notice — and then she refused to paper over it. She said "this might mean some things aren't judgeable" and kept the finding. That's harder than building the judge. — Prof. Weining Zhang, curator