Speaker: Marta Kryven, PhD Candidate
Human beings quickly and confidently attribute more or less intelligence to one another. What is meant by intelligence when they do so? And what are the surface features of human behavior that determine their judgments? Because the judges of success or failure in the quest for 'artificial intelligence' will be human, the answers to such questions are an essential part of cognitive science. This thesis studies such questions in the context of a maze world, complex enough to require non-trivial answers, and simple enough to analyze the answers in term of decision-making algorithms.
According to Theory of Mind, humans comprehend the actions of themselves and others in terms of beliefs, desires and goals, following principles of rational utility. If so, attributing intelligence requires them to solve an inverse decision-making problem to infer the quality of another's planning under uncertainty. Alternatively, attributed intelligence can be the result of observing outcomes: billionaires and presidents are, by definition, intelligent.
I applied Bayesian models of reasoning under uncertainty to interpreting data from five behavioral experiments. The results show that on average humans attribute intelligence to efficient agents that optimize over all possible counter-factual scenarios. However, intelligence attribution is idiosyncratic, and a minority of human participants attribute intelligence to outcome.
An insight into differences in attributed intelligence comes from analyzing how humans plan. Although, on average, humans produce normatively optimal decisions, individuals vary in their sensitivity to decision value. Thus, individual decision-making often deviates from optimal. The specific way of attributing intelligence chosen by an observer to efficiency or to outcome is predicted by the observer's planning ability.
Better planners are more likely to attribute intelligence to efficiency. In contrast, less optimal planners are more likely to attribute intelligence to outcome. Moreover, model-based metrics of planning optimality are linked to independent measures of cognitive performance, such as the Cognitive Reflection Test or pupil size.
Eyetracking analysis of spatial planning in real-time shows that participants who score high on independent measures of cognitive ability also plan further into the future. Taken together, these results converge on a theory of attributed intelligence as an evaluation of decision-making, such that recruits the observer's own decision-making mechanisms to interpret and evaluate observed behavior.