Researchers found adjusting AI systems to be more warm and friendly to users would result in an “accuracy trade-off”.