Human Judgment and AI Pricing

摘要

Artificial intelligence (AI) is undergoing a renaissance. Thanks to developments in machine learning-particularly, deep learning and reinforcement learning-there has been an explosion in the applications of AI in many settings. In actuality, however, far from providing new forms of machine intelligence in a general fashion, what AI has been able to do has been to reduce the cost of higher quality predictions in a drastic way (Agrawal, Gans, and Goldfarb forthcoming(a)). As deep learning pioneer Geoffrey Hinton put it, "Take any old problem where you have to predict something and you have a lot of data, and deep learning is probably going to make it work better than the existing techniques" (Creative Destruction Lab 2016). Thus, when they are able to utilize AI, decision-makers know more about their environment, including about future states of the world.These developments have brought about discussions as to the role of humans in that decision-making process. The view we take here (see also Agrawal, Gans, and Goldfarb forthcoming(b)) is that humans still play a critical role in determining the reward functions in decisions. That is, if the decision can be formulated as a problem of choosing an action (x), in the face of uncertainty about the state of the world (8) with probability distribution function F(0), in an ideal setting, AI can transform that problem from max_x~ fu(x, θ) dF(θ) into max_xu(x,θ) with actions being made in a state-contingent manner. However, thistransformation relies on someone knowing the utility function, u(x,θ). We claim that, at present, only a human can develop this knowledge.