Five ways to mitigate the risk of AI models