Sepehr Razavi (Edinburgh, UK): Old ideas, new twist: critical assessment of the signature testing approach

Work in comparative psychology on theory of mind (henceforth ToM) in nonhuman animals and systems is instructive for at least three reasons. First, a historical one, before being used for philosophical, clinical, and scientific interests in humans, the term “theory of mind” was coined in Premack and Woodruff’s (1978) seminal work on the attribution of mental states in chimpanzees. Second, this study in nonhuman animals helps provide a sense of the minimal implementation for ToM, notably by studying species that have “more limited inhibitory control or memory” than humans (Krupenye and Call 2019). Third, and most significantly, for this presentation I will centre the “signature-testing approach”  (Taylor et al. 2022) to mapping measured behaviour to cognitive processes. It is worth shedding light on this for two important reasons. On the one hand, it is a new approach that has yet to be implemented, promising to supplement some important lacunae in comparative intelligence research by challenging predominant assumptions made in previous experimental tasks. On the other hand, its rich foundations were set in place by comparative psychologists and philosophers, drawing from important breakthroughs made in physics in the observation of other phenomena that are not directly observable. My paper will be divided into two parts. 

In the first part (§1), I evaluate the assumptions challenged by “signature-testing” applied to the specific area of mindreading in nonhuman animals and take a closer look at the new approach’s own commitments. From ongoing research, I will conclude that, although Taylor et al.’s approach hints at important shortcomings in comparative psychology and the philosophy of nonhuman minds, it is ultimately not so clear that other researchers are oblivious to these problems or that the stronger version of the many-to-one mapping problem can be overcome with the signature-testing approach. Perhaps the most important assumption of previous research in comparative psychology challenged by Taylor and colleagues concerns the “gold-standard” problem. Indeed, a key motivation for their approach stems from a desire to move past success paradigms that focus on whether or not the compared non-human systems meet a benchmark (Shettleworth 2009; Halina 2023). I will argue that some of the more conclusive studies (e.g., Krupenye et al. 2016) already perform a restriction of hypothesis space advocated more recently by the proponents of the signature-testing approach.

In the second part (§2), starting from the computational assumptions made by the signature testing approach, I will present two recent cognitive models of ToM proposed to explain mindreading in humans (Baker et al. 2017; Zeng et al. 2020). While the first Bayesian model demonstrates convergences toward signature-testing, the second one uses task paradigms that illustrate the problems of success testing as suggest by Taylor et al. Indeed, within the last few years mathematical psychologists and cognitive scientists have made one of their most interesting breakthroughs in the modelling of cognition and perception, that is modelling ToM. Baker et al. (2017) have proposed a Bayesian computational model to account for the inferences we make about the mental states from observed behaviours. The key assumption here is that actions are generated via a rational process aiming to fulfil the agent’s current desire while minimizing costs entailed by her beliefs.