📰 People + AI Guidebook

• a lot of basic ⌣ user experience stuff, how to do user research, etc etc, sometimes with Artificial Intelligence context
• > there should almost always be an option for human oversight
• the crosstabulation mental model in context of binary classifiers and their statistical error types is interesting
  • → optimizing for precision
  • → optimizing for recall
  • very relevant for designing search engines and search in general
• ~ ask what would happen if your software works perfectly
  • → what if I'm right
• recommends tripwires
• warns of data cascades
  • → ~ align dataset creation to user needs
• types of discriminatory harm by AI
• has specific advice on labeling software
• ~ users assume that AI uses the same mental model as them
• ~ introduce features only when needed
• ~ allow manual task completion when AI fails
• ~ help users calibrate their trust
  • ~ in a high-stake scenario, don't say 'what' without 'why'
• interesting points on trust in software in general as well as feedback in software
• > Not every task is a chore

Recommendations

• ~ set tripwires
• Google Dataset Search

Brings up

• implicit feedback
• context error

Formalia

• → https://pair.withgoogle.com/chapter/user-needs/