How do we reason about agents who are ignorant? When we interact with someone that has partial or incomplete knowledge, we can flexibly understand and predict their behavior depending on whether their ignorance is easy to remedy (what's inside this box?), out of their control (will it rain today?), or irrelevant to their goals (do we have free will?). Similarly, when we recognize that we don't know something, we can rectify our ignorance through exploration or by watching more knowledgeable agents act.

Phillips et al. make a compelling case that, within theory of mind, knowledge is a more basic representation than belief. But, in doing so, Phillips et al. also treat knowledge and ignorance as two sides of the same representational coin. However, the representational demands of knowledge and ignorance are not necessarily equivalent. For instance, one of the simplest ways to represent ignorance would be as the absence of knowledge. This, however, would require a negation-like representation of a knowledge state. Because the ability to apply negation over mental representations appears to be absent in younger children (Feiman, Mody, Sanborn, & Carey, Reference Feiman, Mody, Sanborn and Carey2017; Mody & Carey, Reference Mody and Carey2016; Nordmeyer & Frank, Reference Nordmeyer and Frank2014; Reuter, Feiman, & Snedeker, Reference Reuter, Feiman and Snedeker2018) and is weak in nonhuman primates (Call & Carpenter, Reference Call and Carpenter2001), even this simple relationship would already predict that representations of ignorance are not an inevitable consequence of representations of knowledge.

Even if children and nonhuman primates could represent ignorance as a consequence of their ability to represent knowledge, this alone would not provide the computations needed to predict and understand the behavior of ignorant agents, making these representations of limited use. Indeed, predicting the behavior of an ignorant agent goes far beyond merely expecting that they will not act in a knowledgeable way: Accurate predictions about ignorant agents involve determining whether they will choose to gather information, and how they will act to maximize their chance of success under uncertainty.

Importantly, these concerns do not simply reflect theoretical questions about the nature of ignorance representations. The few empirical studies that test for an intuitive theory of ignorance suggest that these representations have a tenuous correlation with knowledge representations. Although some sensitivity to ignorance appears early in development (Koenig & Echols, Reference Koenig and Echols2003; O'Neill, Reference O'Neill1996), children's understanding of ignorance continues to develop after children have a mature understanding of knowledge. Young children exhibit egocentric errors, attributing their own knowledge to ignorant agents (Birch & Bloom, Reference Birch and Bloom2003; Hogrefe, Wimmer, & Perner, Reference Hogrefe, Wimmer and Perner1986; Mossler, Marvin, & Greenberg, Reference Mossler, Marvin and Greenberg1976; Sullivan & Winner, Reference Sullivan and Winner1991; Wellman & Liu, Reference Wellman and Liu2004); they fail to predict that agents searching for a hidden object will choose randomly (Friedman & Petrashek, Reference Friedman and Petrashek2009; Ruffman, Reference Ruffman1996); and they do not expect ignorant agents to seek additional information when necessary (Huang, Hu, & Shao, Reference Huang, Hu and Shao2019).

Similarly, there is little evidence that nonhuman primates can predict the actions of ignorant agents (Drayton & Santos, Reference Drayton and Santos2018; Horschler, Santos, & MacLean, Reference Horschler, Santos and MacLean2019; Karg, Schmelz, Call, & Tomasello, Reference Karg, Schmelz, Call and Tomasello2015b; Marticorena, Ruiz, Mukerji, Goddu, & Santos, Reference Marticorena, Ruiz, Mukerji, Goddu and Santos2011; Martin & Santos, Reference Martin and Santos2016). Many experiments examining nonhuman primate theory of mind directly contrast knowledge and ignorance in a single task, which means that subjects can succeed by (1) only representing knowledge, (2) only representing ignorance, or (3) representing both (e.g., Flombaum & Santos, Reference Flombaum and Santos2005; Hare, Call, Agnetta, & Tomasello, Reference Hare, Call, Agnetta and Tomasello2000; Karg, Schmelz, Call, & Tomasello, Reference Karg, Schmelz, Call and Tomasello2015a), making it impossible to discern which representations are guiding subjects' behavior. Even looking-time tasks that probe knowledge and ignorance under different conditions do not provide clear evidence of ignorance representations. For example, after seeing an object hidden in one of two boxes, rhesus macaque monkeys look equally long at the display when an ignorant demonstrator reaches for the correct or incorrect box (Drayton & Santos, Reference Drayton and Santos2018; Marticorena et al., Reference Marticorena, Ruiz, Mukerji, Goddu and Santos2011). Crucially, these results are consistent with two competing explanations: Subjects may be unsurprised because both actions are consistent with their prediction that the ignorant agent will search randomly or they may be unsurprised because they made no prediction at all. The former is consistent with Phillips et al.'s proposal that nonhuman primates are able to make predictions about both knowledgeable and ignorant agents. However, the latter would suggest that rhesus macaques either cannot represent ignorance or cannot form predictions about ignorant agents, despite having expectations about the behavior of knowledgeable agents (Drayton & Santos, Reference Drayton and Santos2018; Marticorena et al., Reference Marticorena, Ruiz, Mukerji, Goddu and Santos2011). Similar concerns also apply to “ignorance” conditions in looking-time studies with infants (Hamlin, Ullman, Tenenbaum, Goodman, & Baker, Reference Hamlin, Ullman, Tenenbaum, Goodman and Baker2013; Luo & Johnson, Reference Luo and Johnson2009).

Taken together, these empirical findings suggest that children and nonhuman primates may not have a rich understanding of ignorance despite being able to successfully reason about knowledgeable agents. This presents an exciting opportunity to reevaluate the common assumption that ignorance representation inevitably accompanies knowledge representation. One possibility is that knowledge is a primary representation out of which ignorance representations are later derived – through negation or otherwise. Such a relationship would explain the developmental lag in ignorance understanding in children and make testable predictions about the status of ignorance representations in nonhuman primates depending on the hypothesized requirements to build this secondary representation. Alternatively, knowledge and ignorance representations may be independent from one another, combining later in life to support reasoning about agents with partial or incomplete knowledge. Critically, in either case, these proposals are consistent with Phillips et al.'s view of the primacy of knowledge representations.

Or perhaps, Phillips et al. are right: Knowledge and ignorance representations may be impossible to disentangle, developmentally indistinguishable (with previous ignorance failures representing only task demands), and best understood in tandem. The task is now to clearly articulate this relationship and design empirical investigations of ignorance representations in their own right, rather than as a control condition for studies of knowledge. A complete account of mental-state representations must explain how ignorance is derived, what (if any) additional representational machinery is necessary, and whether the hypothesized relationship predicts any critical gaps in development of representations of knowledge and ignorance. The answers to these questions are essential not only for understanding this representational capacity, but also for understanding our knowledge representation system and our ability to interpret and predict epistemic actions.