Specification gaming examples in AI

Here’s a super fun list of things that artificial intelligence figured out by gaming the rules, like inconsistent and incomplete specifications, bugs, and other bits that humans frequently assume and ignore.

Some examples to get you started are:

  • Aircraft landing: evolved algorithm for landing aircraft exploited overflow errors in the physics simulator by creating large forces that were estimated to be zero, resulting in a perfect score.
  • Block moving: a robotic arm trained to slide a block to a target position on a table achieves the goal by moving the table itself.
  • Data ordering patterns: neural nets evolved to classify edible and poisonous mushrooms took advantage of the data being presented in alternating order, and didn’t actually learn any features of the input images.
  • Road runner: agent kills itself at the end of level 1 to avoid losing in level 2.
  • Ruler detector: AI trained to classify skin lesions as potentially cancerous learns that lesions photographed next to a ruler are more likely to be malignant.
  • Tetris: agent pauses the game indefinitely to avoid losing.

This is truly thinking outside the box!

Leave a Comment