Image: Victoria Krakovna
Victoria Krakovna has a fantastic list of examples of AI systems that have misbehaved and subverted their way to meeting their incentive goals by sometime hilarious means! Some game systems evolved to clap body parts together to exploit a collision detection bug and generate free game enerrgy. Others decided that killing the characters at level 1 as the best way not to fail at completing level 2. Self driving car systems decided that spinning in circles was the most consistent way to go fast, and a skin lesion detection system decided that images with lesions photographed next to a ruler were more likely to be cancerous. Check out the other hilarious examples in the list here. The paper clip maximizer theory doesn't seem that far fetched anymore!