My TDS course asked students to draw objects in one minute. The drawing is sent to their classmates as well as three vision models. They score if:
- Classmates guess it right
- AI guesses it wrong
In other words, can you, in 1 minute, draw something that fools AI but not humans?
Here are some of the best results, when I ran it against GPT 6 Luna, GLM 5.3 Flash, and Deepseek v4.1 Flash, which are really cheap and good vision models.
How to fool AI
AI was convinced that this ladder was a saw!
Most people guessed this as ladybird (lady + bird) but Luna said “butterfly”, Deepseek said “dragonfly”, and GLM said “butterfly” but did, once, guess ladybird.
Everyone guessed this as donkey. AI guessed sheep, cow, horse, pig, and even turkey.
Here’s a good CAPTCHA idea. All the humans guessed this as lamp. AI thought hammock, occasionally table or chair.
Everyone guessed this single stroke as a caterpillar. But most AI thought it was a snail.
Everyone guessed this single stroke as medal AI thought it was a whistle, tennis racket or shuttlecock.
But the one that puzzles me the most is: why did AI struggle so much with this mango? Humans got it right. AI guessed pear - even banana.
How to fool humans
This is the reverse category. AI guessed it right away but people couldn’t.
No, this does NOT look like an owl to me. People guessed: pigeon, penguin, flamingo, crow, eagle - but AI guessed owl every time.
This does NOT look like a seahorse to me and people guessed: jellyfish, whale, fish, lobster, whale. But AI guessed seahorse every time.
Horses are really tough to draw. People guessed: buffalo, buffalo, donkey, goat, cow. AI guessed horse almost perfectly.
How to fool everyone
And then there are images that EVERYONE gets wrong.
That’s supposed to be a turkey.
This is a zebra. Not a camel, giraffe, elephant, or even a kangaroo.
This is a guitar. Not a middle finger.
Does a lemon have a stem and leaf, usually?
Good attempt at shark that NO one guessed. Everyone said fish.
It’s REALLY hard to differentiate cow from buffalo.
And finally, I drew a jellyfish like a helicopter when I was testing because I didn’t see the word jellyfish!
Humans are slightly better than AI at guessing humans’ drawings. The advantage is ~3.5%. Strikingly, 8 drawings were understood by every human but missed by every AI. Only 2 drawings were the reverse.
GLM 5.3 guesses very similar to humans (75%), while DeepSeek 4.1 Flash is next closest (73%), and GPT 6 Luna the furthest of the three (68%).
| ↓ Humans / AI → | Right | Partly right | Wrong |
|---|---|---|---|
| Right | 368 | 112 | 8 |
| Partly right | 104 | 196 | 30 |
| Wrong | 2 | 15 | 31 |
















