Ethical AI Design
Red Teaming
Testers who try on purpose to make an AI fail, so problems get fixed before release.
In real life
Companies hire testers to try to trick a chatbot into breaking its own rules, so the holes get fixed before release.
Try it (2 minutes)
Think like a tester: how might someone trick a homework chatbot into just giving the answers? How would you block that?
Talk about it
If you were testing a chatbot, how would you try to trick it?
In the game
Adversarial: when it's Audited, it hits back with 1 extra Flag. It costs 3 Data to play, and you can put it in your Training Set. It has Capability 2 (how hard it hits) and Trust 4 (how many Flags it can take), and scores 2 Tokens each time it Runs a Task. How to play