cat? $dog

A two alternative forced choice with a sample size of one. Nine looks at the same animal. You say what it is each time.

The test

This is the whole animal. The test below measures how much of it people need.

Where the cat ends

The whole animal is four marks of ink.

46% shown

The captcha that died

  1. 2007

    Microsoft Research ships Asirra: to prove you are human, pick the cats out of twelve photographs of pets, supplied by Petfinder. Humans solve it 99.6% of the time in under thirty seconds. Machines are given no better than one chance in 54,000.

  2. 2008

    Philippe Golle trains a support vector machine on colour and texture. It reads a single photograph correctly 82.7% of the time and clears the full twelve at 10.3%, fifty times the estimate it was built against.

  3. 2013

    25,000 of the photographs are reissued as a Kaggle competition. Pierre Sermanet wins it at 98.914%, which is within a point of the humans. Asirra is retired.

Asirra was built on two beliefs: that a cat and a dog are obviously different, and that only a person can see it. The machines spent six years on the second one. This drawing has been quietly working on the first.