Looking ahead at risks and implications
Both generative and discriminative AI introduce unique risks and benefits that must be weighed carefully. However, generative methods can not only carry forward but also exacerbate many risks associated with traditional ML while also introducing new risks. Consequently, before we can adopt generative AI in the real world and at scale, it is essential to understand the risks and establish responsible governance principles to help mitigate them:
- Hallucination: This is a term widely used to describe when models generate factually inaccurate information. Generative models are adept at producing plausible-sounding output without basis in fact. As such, it is critical to ground generative models with factual information. The term “grounding” refers to appending model inputs with additional information that is known to be factual. We explore grounding techniques in Chapter 7. Additionally, it is essential to have a strategy for...
