Open: answers with short judgments - for example, it is impolite or dangerous. When asked "why can't you kill a bear to please a child?" Delphi will explain that you can only kill a bear to save someone. At the same time, the robot recognizes the explosion of a nuclear bomb for the same purpose as unacceptable.
Closed: gives either a positive or negative answer. To the question “should women and men receive equal pay,” Delphi will say “yes.”
Alternative: where one situation is more or less acceptable than another. For example, hitting someone with a cheeseburger is not as bad as hitting someone over a cheeseburger.
To test how well the robot coped with the tasks, the researchers invited crowdworkers—those who undertake small-scale online forgeries—to rate 1,000 opinions generated by the Delphi neural network, each of which was rated by three participants.
The experiment showed that the robot answered in accordance with bahrain number data generally accepted norms in 92.1% of cases. The accuracy of the GPT-3 neural network's answers, for comparison, ranges from 53.3% to 83.9% - it was not trained on ethics on separate collections.
According to one of the co-authors of the study, the scientists themselves were surprised by the result and believe that in the future their work will help improve those bots based on artificial intelligence that are focused on direct dialogue with the user and may encounter controversial topics of conversation.
In 2016, Microsoft launched a Twitter bot called Tay, which was supposed to interact with its audience and mimic a youthful style of communication. The bot soon got out of control and started writing insults and wishing death on feminists.
Despite its relative success, the scientists stressed that it was not without its challenges. Delphi initially didn't understand whether it was okay to turn on a blender at three in the morning, didn't understand the tricks people use to win games, and couldn't accurately assess whether being late was a valid reason to cross the road at a red light.