Should AI learn from its own mistakes?