Research shows AI datasets have human values blind spots

Purdue University researchers found an imbalance in human values within AI training datasets from major U.S. companies. The datasets emphasized information and utility values over prosocial, well-being, and civic values. Information seeking and wisdom/knowledge were most common, while justice and human rights were least represented. This imbalance could affect AI systems' interactions and handling of social issues. The researchers seek to help companies create more balanced datasets reflecting community values.