Abstract

Recent progress in artificial intelligence has intensified claims about AI reaching or surpassing human intelligence. However, much of this discussion is not grounded in scientific measurements of intelligence in humans, animals or AIs. The existing methods for measuring intelligence are a patchwork of heterogeneous approaches that employ different measurement scales and are often limited to a single type of agent. This paper addresses this problem by developing a taxonomy that organizes intelligence measures according to the type of measurement scale (nominal, ordinal, interval and ratio) and methodology (judgment, test batteries, indirect and direct measurement of properties linked to intelligence). Methods that are included in this survey include human intuition, standard tests like IQ and g, test batteries for animals and AIs, and direct measurement of prediction and goal-achievement. This paper argues that direct ratio measures are the most plausible route to a single universal measure that would enable the intelligence of humans, animals and AIs to be compared on a single scale. This could provide a real measure of progress in artificial intelligence, inform debates about AI safety, and help us to develop a scientific understanding of the relationship between intelligence and consciousness.

IPC Classification

Keywords