Apply rich psychological terms in AI with care

Abstract

There is much to be gained from interdisciplinary efforts to tackle complex psychological notions such as 'theory of mind' by combining the rich history of study and debates in cognitive science and recent findings from AI research. However, careful and consistent communication is essential when comparing artificial and biological intelligence, say Henry Shevlin and Marta Halina. Many of the most promising approaches in cognitive science seek to explicate notions like perception, belief, and motivation in information processing terms [1]. A similar move from information processing to psychology is occurring in artificial intelligence (AI) research and machine learning. Currently, few people working in AI would literally attribute beliefs, thoughts, or feelings to machines. However, as new techniques extend the capabilities of artificial systems, it has become increasingly common to use psychological terms to describe processing architectures [2-6]. Such use of psychological terms in AI research may be in many cases justified and unproblematic. Some terms with both narrow computational and psychological meanings, such as "memory" and "reinforcement", have clearly distinct histories in psychology and artificial intelligence and their senses are unlikely to be confused. Others such as "learning" and "behavior" straddle psychology and artificial intelligence, but are broad enough that there is little reason to quibble with their use to describe machine capacities. However, we argue here that there is a third kind of more robustly psychological concepts-including notions like awareness, perception, agency, and theory of mind-that have rich and complex histories in cognitive science. We suggest that these terms-which we call rich psychological concepts-require greater caution when employed to describe the capabilities of machine intelligence.

Performance vs. competence in human–machine comparisons

“…What do such behaviors imply about human-machine similarity? † Importantly, the question of humanlike processing in machines is distinct from whether such systems are truly "intelligent" or "minded" in some deeper sense (27,33,40), which can be asked independently of their similarity to humans. It also differs from the engineering goal of actually building such machines (26), which is separable from evaluation.…”

Visual behavior modelling for robotic theory of mind

“…2 ) comprising a physical robot actor, and an observer agent that watches the actor from above. In order to evaluate the effectiveness of our proposed method, we followed the guidelines proposed by Shevlin and Halina 53 to design our evaluations from both a Machine Learning perspective and a Psychology perspective. During our experiments, we aimed to answer the following two questions: First, can an observer agent visually predict the future plans of the actor robot, without symbolic reasoning.…”

Insightful artificial intelligence

“…Doing so would mean applying a model of one game, such as shogi, to another, such as Go, which would be an unsuccessful strategy, given that the board positions and rules from one game do not transfer to the other. This is not to minimise the power, value or even creative potential (see below) of these systems, but to identify an important difference between them and the capacities believed to be operating in animals when they engage in insightful problem solving (Shevlin & Halina, 2019).…”