Loopholes, a Window into Value Alignment and the Learning of Meaning

Date: Friday, September 24, 2021
Speaker: Sophie Bridgers, Ph.D.
 Simons Postdoctoral Fellow, MIT BCS Early Childhood Cognition Lab (PI: Laura Schulz) and Harvard Computation, Cognition, and Development Lab (PI: Tomer Ullman)

Abstract: Finding and exploiting loopholes, a possible but unintended interpretation of a rule or request, is a familiar facet of fable, law, and everyday life. A child may respond to their parent who says “It’s time to the put the tablet down” by continuing to play with the tablet after physically putting it down on the floor. Engaging with loopholes requires a nuanced understanding of goals, social ambiguity, and value alignment and offers a new lens through which to examine human communication and cooperation. But cognitive, computational, and empirical work on this behavior remains scarce. I will present findings from a parent survey that establishes the ecological validity of loophole behavior in childhood and from two experiments showing that both adults and young children (4 to 10 years) consider exploiting a loophole as less costly than outright noncompliance. I will conclude with a discussion of the development of loophole behavior in neurotypical and ASD populations, and a proposal for a formal framework of goal communication that supports such intentional misunderstandings. This work has implications for improving communication among humans, across developmental and neurodiverse populations, as well as for safer human-AI interactions.​