This post is also available in:
עברית (Hebrew)
A new study by Washington University in St. Louis revealed that when humans are told they are training AI when playing a bargaining game, they would intentionally change their behavior to one they deem more ethical, perhaps having the motivation to train AI for fairness.
The study, published in Proceedings of the National Academy of Sciences, included five experiments, with 200–300 subjects each. In each experiment, participants were asked to play the “Ultimatum Game,” where they negotiated a small amount of cash (ranging from $1 to $6), either with other human players or with a computer. Some of the participants were told that their decisions in the game would be used to train an AI bot on how to play the game. Remarkably, those who believed they were training AI were consistently more inclined to have a fair share of the money, even at their own expense. Surprisingly, this tendency to seek fairness continued even after they were told that their choices were no longer being used to train the AI, indicating that the experience of contributing to technology had a lasting effect on their decision-making.
Yet, the motive for this behavior is still unclear. The study didn’t require participants to disclose their reasonings, and it is possible that it was not their kind intentions to make AI more ethical that was the driving factor. Follow-up experiments are currently taking place to understand the motivations of people when training AI. Furthermore, researchers warn that even if it was the case, other people might have different agendas.
However, this information that people who are aware they are training AI might actively alter their behavior is important for developers to consider when working on AI training. The researchers explain that if human biases during AI training aren’t considered, the AI model that was trained will likely also become biased in the same ways. They explained this is something we already see today, as certain facial recognition softwares provide worse accuracy when trying to identify people of color, partly because the data used to train AI was biased, according to TechXplore.
