What experts can learn by tracking AI harms