AI and illusions of understanding in scientific research
An article in Nature today (behind paywall, Artificial intelligence and illusions of understanding in scientific research) raises some interesting issues and questions about the use of hashtag#ai and hashtag#genai in scientific research, not just for writing articles, the most common but somewhat tame application, but for making sense of papers, data and crafting explanations. Luckily, there is a freely available synopsis (see below). I can recognize the different visions the authors lay out (Oracle, Arbiter, Quant, Surrogate) and it is clear that all issues stem from human biases, or perhaps more generally human nature, being enabled and/or amplified by AI.
Not surprisingly, a behavioral economics mindset leads one to think of ways to mitigate such biases and traps. I love that. hashtag#behavioraleconomics
But there is in my view a more fundamental lack of understanding of the optimal task allocation between human and AI: it is not just about what we can do better than an AI and what an AI can do better than us. Paraphrasing Andrew Ng, "I don't know if AI will replace radiologists, but I know that radiologists using AI will replace those that don't." So, how do we optimize the synergies between human and AI to improve performance and reduce biases?