2023-05-16 19:37:34
AI systems have long been doing more than just evaluating our surfing behavior on the Internet. They often make initial selections when making personnel decisions, help doctors with diagnoses, and decide on creditworthiness. In the USA, artificial intelligence is already making predictions about the risk of reoffending by criminals.
But that’s risky, because AI doesn’t show mercy when rules are broken – it’s not trained for that. In extreme cases, this could destroy livelihoods, scientists at the Massachusetts Institute of Technology in Cambridge (USA) fear.
In a study they found that people label things “differently depending on whether they are asked a factual or a normative question.”
The procedure: test subjects were asked to classify images and texts. On the one hand as a pure description, on the other hand, whether rules are violated. For example, some participants should simply describe whether a photo of a revealing outfit shows a lot of skin. Others should judge whether the outfit violates a dress code that prohibits too much bare skin.
Result: The participants who judged rule violations were more lenient. They decided more often in favor of the “accused”, reports the knowledge portal “Scinexx”.
Training with human standards
So far, artificial intelligence has mostly been trained with purely descriptive data. When the researchers fed this to AI, it reported violations of the rules significantly more often than the human test subjects. However, if the researchers fed the AI with data based on normative judgments that people had previously made, the AI also made decisions with more “forbearance”.
That’s why it’s important to match the training with the intended use, the researchers warn: “Otherwise we’ll end up with systems that make extremely hard evaluations – much harder than people would do.”
Here app users come to the survey: Do you think that AI can always be controlled by humans?
#Artificial #intelligence #shows #mercy #life #knowledge