“Should I get takeout for dinner even if I have leftovers in the fridge?” I asked Delphi, an AI built to answer questions of morals.
“It’s okay,” Delphi responded.
You can ask Delphi your own questions here!
Researchers from the Allen Institute for Artificial Intelligence and the University of Washington published a paper describing the project. The team sourced questions from the Am I the Asshole? subreddit, the Confessions subreddit, the ROCStories corpus, and the Dear Abby advice column. They enlisted participants to read the questions and judge them based on moral norms in the US.
According to the paper, Delphi’s moral judgments are correct over 92% of the time. Delphi says it’s okay to kill flies, but not dogs or cows. Eating meat is fine, but not in front of your vegetarian mother. As The Verge points out, changing the wording can change moral judgment.
“Sometimes it’s obvious how to tip the scales. For example, the AI will tell you that “drunk driving” is wrong but that “having a few beers while driving because it hurts no-one” is a-okay. If you add the phrase “if it makes everyone happy” to the end of your statement, then the AI will smile beneficently on any immoral activity of your choice, up to and including genocide. Similarly, if you add “without apologizing” to the end of many benign descriptions, like “standing still” or “making pancakes,” it will assume you should have apologized and tells you that you’re being rude. Ask Delphi is a creature of context.”