How the AI judges your drawings
Rating a 60-second finger doodle fairly is harder than it sounds. Here's exactly how Sketchmate turns your scribble into a score, and why you can trust it.
Pass 1: the blind guess
The first thing we ask a vision model is simply “what is this?”, without telling it the answer. Vision models told the answer tend to “see” it in almost anything. A blind guess avoids that, and it's the same test your friends are playing: would a stranger recognise it?
The model returns its top three guesses. If the secret shows up, you earn recognition points: full marks for the first guess, 75% for the second, 55% for the third. Close answers count too. “Tree” earns 70% for “Pine Tree”.
Pass 2: the rubric
In parallel, a second pass knows the secret and the describer's clue. It scores three things from 0 to 10:
| Criterion | Weight | What it measures |
|---|---|---|
| Subject | 60% | Does the drawing show the secret thing? |
| Details | 25% | Did you capture the traits from the clue: shapes, colours, parts, setting? |
| Effort | 15% | Is it complete, with a clear composition? One squiggle is not a masterpiece. |
It also names what matched (“bushy tail”) and what's missing (“orange colour”). Those phrases appear on your results card.
The score is calculated, not invented
The model never picks the headline number. Sketchmate computes it from the rubric (70%) and the blind recognition (30%), with a few fairness rules:
- If the AI recognised your drawing blind, you get at least 60. A stingy rubric can't bury a doodle a stranger would get.
- A beautiful drawing of the wrong thing is capped at 40.
- A drawing that's mostly written words is capped at 20. Blank canvases score 0.
Built to be fair
- Consistent: low-temperature scoring with a fixed seed, so the same drawing gets the same score.
- Injection-proof clue: the describer's text is treated as data, never as instructions to the judge.
- Honest failures: if the AI can't score a drawing, you get the round's median and a note saying so. We never show a random number.
- Private: drawings are only viewable during the results screen and are deleted when the room closes.
The judge runs on Meta's Llama 4 Scout vision model on Cloudflare Workers AI, with Llama 3.2 Vision as an automatic fallback.
Try to fool it
Grab 2–8 friends and see who the AI can read. Free on iPhone and Android.