Evaluating information-seeking approaches to metacognition

Abstract

Metacognition has been divided into information monitoring and control processes. Monitoring involves knowing that you know or do not know some information without taking corrective action. Control involves taking corrective action based on the knowledge that you know or do not know some information. In comparative metacognition, considerable attention has been paid toward critically assessing putative evidence for information monitoring in non-human animals. However, less attention has been paid toward critically evaluating evidence for control processes in animals. We briefly review a critique of information-monitoring in animals. Next, we apply these concepts to a number of studies that focus on information seeking in animals. The main type of evidence for control processes in animals come from tube tipping experiments. Before having the opportunity to search for the bait in these experiments, the subject sometimes observes opaque tubes being baited but is sometimes prevented from seeing the baiting. The observations that the subjects look more if baiting was not seen and are more accurate if baiting was seen have been taken as evidence for metacognition in information-seeking experiments. We propose simple alternative hypotheses that are sufficient to explain putative evidence for information seeking in animals without positing metacognition. The alternative explanation focuses on two relatively simple principles: First, an animal has a default "look before you go" response which supersedes random searches in space. Second, spatially guided behavior follows a default rule of "go where something good is." These principles can explain the results of tube tipping experiments without proposing metacognition.

Introduction

Metacognition is the ability to reflect on one's own mental processes, which is a defining feature of human existence (Descartes 1637; Metcalfe and Kober 2005). One of the fundamental questions about cognition in non-human animals (henceforth animals) is therefore whether they have knowledge of their own cognitive states (e.g., Smith 2009; Terrace and Son 2009). Metacognition has been divided into information monitoring and control processes (Nelson and Narens 1990). Monitoring involves knowing that you know or do not know some information. According to this conceptualization, an animal may know that it does not know some information (which it may show through its behavior), but it does not have an opportunity to correct for the perceived lack of information. An inability to correct for perceived lack of information may occur because the capacity to do so does not exist (for example in animals) or because an experimental procedure does not give the subject an opportunity to act on the perceived lack of information (e.g., in monitoring experiments described in the next section). By contrast, control involves information seeking (which it may show through its behavior) based on the knowledge that you know or do not know some information. According to this conceptualization, in addition to an animal knowing that it does not know some information, it also has the ability to take some corrective action for the perceived lack of information. Of course, to infer metacognition in animals from either monitoring or control perspectives, it is necessary for the animal to show its knowledge or lack of knowledge through its behavior. The majority of research in comparative metacognition has focused on information monitoring, but some research has focused on control processing. Recently, putative evidence for information monitoring in non-human animals has received extensive critical assessment (Staddon et al. 2007; Smith et al. 2008; Crystal and Foote 2009a; Crystal and Foote 2009b; Hampton 2009; Jozefowiez et al. 2009a; Jozefowiez et al. 2009b) However, less attention has been paid toward critically evaluating evidence for control processes in animals. We briefly review the critique of information monitoring approaches in animals. Next, we apply these concepts to a number of studies that focus on information seeking in animals.

Information monitoring and control

Nelson and Narens (1990) proposed that a cognitive-executive process regulates the flow of information between an object-level and a meta-level by using control and monitoring processes (see Figure 1). The object-level is a reservoir for an individual’s cognitions, behaviors, memories, and descriptors of a current situation whereas the meta-level monitors and controls the object-level (Son and Kornell 2005). Control and monitoring processes are two “tools” that the central-executive process utilizes to permit communication between the proposed meta- and object-levels. Generally, monitoring processes consist of such phenomena as confidence judgments, feeling-of-knowing judgments, ease-of-learning judgments, and judgments-of-learning (Smith 2005). By contrast, control processes are composed of phenomena which determine the selection and kind of processing, selection of a search strategy, termination of study and search, and the allotment of time for study (Smith 2005).

Figure 1.

Nelson and Naren’s (1990) model of metacognition in humans. A central-executive process monitors internal assessments of knowledge and controls information seeking. (Adapted from “Metamemory: A Theoretical Framework and New Findings,” by T. O. Nelson and L. Narens, 1990, The Psychology of Learning and Motivation, 26, p.129. © 1990 by Academic Press, Inc. Reprinted with permission.)

An example of information monitoring in animals

Investigations of information monitoring in animals has been reviewed extensively elsewhere (Kornell 2009; Smith 2009; Terrace and Son 2009) and will not be repeated here. However, we begin with two examples of information monitoring in animals to ground the development of critiques of metacognition interpretations of data. Such critiques are described in the next section. Although the data reviewed below may be open to alternative explanations that do not involve metacognition, here we emphasize the availability of a metacognition account. In the next section, we outline an alternative account that does not involve metacognition. In the section labeled Parsimony, we outline our views about selecting between multiple, competing accounts of data.

Inman and Shettleworth (1999) established a standard to document metacognition, which we will refer to as the information-monitoring prevailing standard There are two criteria: (1) the frequency of declining a test is expected to increase with task difficulty and (2) accuracy is expected to be higher on trials in which a subject chooses to take the test relative to forced tests, and this accuracy difference is expected to increase as task difficulty increases (this latter pattern is referred to as the Chosen-Forced performance advantage). These criteria are intuitive, and may be understood from a metacognition perspective as follows. If you know that you do not know some piece of information, then you would be expected to bail out of tests for that information, if the opportunity to decline is available. Moreover, if you are prevented from declining a test, then performance will be lower when you are forced to take a test relative to tests that you elect to take; forced tests include trials that would have been declined had that option been available, which thereby impairs forced-test performance.

Representative data from monkeys and rats

We show two examples that meet the information-monitoring prevailing standard, using data from a rhesus monkey (Hampton 2001) and rats (Foote and Crystal 2007). Hampton (2001) used daily sets of four icons in a matching to sample procedure (i.e., reward was contingent on selecting the most recently seen icons from a set of distracters). The procedure is shown in Figure 2. Foote and Crystal (2007) presented a noise from a set of eight durations, which was to be categorized as short or long (i.e., reward was contingent on judging the shortest and longest durations as short and long, respectively). The procedure is shown in Figure 3. Both experiments had the following features. Before taking the memory test, the animals were given an opportunity, on some trials, to decline it. On other trials, the animals were forced to take the test without the option to decline it. Accurate performance on the test yielded a high-value reward, whereas inaccurate performance resulted in no reward. Declining a test yielded a low-value, but guaranteed, reward. The data may be summarized as follows. The rate at which the animals declined to take a test increased as a function of test difficulty (longer retention intervals for the monkey or proximity to the subjective middle of short and long durations for the rats). In addition, accuracy was lowest on difficult tests that could not be declined. Note that the data in Figures 2 and 3 meet the information-monitoring prevailing standard: the animals appear to have used the decline response to avoid difficult problems, and the Chosen-Forced performance advantage increased with task difficulty.

Figure 2.

Schematic representation of information-monitoring design of study and data. Procedure for monkeys (left panel; Hampton, 2001): After presentation of a clip-art image to study and a retention-interval delay, a choice phase provided an opportunity for taking or declining a memory test; declining the test produced a guaranteed but less preferred reward than was earned if the test was selected and answered correctly (test phase); no food was presented when a distracter image was selected in the memory test. Items were selected by contacting a touch-sensitive computer monitor. Data (right side; Hampton, 2001): Performance from a monkey that both used the decline response to avoid difficult problems (i.e., relatively long retention intervals) and had a Chosen-Forced performance advantage that emerged as a function of task difficulty (i.e., accuracy was higher on trials in which the monkey chose to take the test compared with forced tests, particularly for difficult tests). Filled squares represent the proportion of trials declined, and filled and unfilled circles represent proportion correct on forced and chosen trials, respectively. Error bars represent standard errors. (Adapted from Hampton, R. (2001). Rhesus monkeys know when they remember. Proceedings of the National Academy of Sciences of the United States of America, 98, 5359-5362. © 2001 The National Academy of Sciences. Reprinted with permission.)

Figure 3.

Procedure for rats in an information-monitoring design (top left panel; Foote and Crystal 2007): After presentation of a brief noise (2-8 s; study phase), a choice phase provided an opportunity for taking or declining a duration test; declining the test produced a guaranteed but smaller reward than was earned if the test was selected and answered correctly (test phase). The yellow shading indicates an illuminated nose-poke (NP) aperture, used to decline or accept the test. Data (Foote and Crystal 2007): Performance from three rats (bottom panels) and the mean across rats (top-middle and top-right panels). Difficult tests were declined more frequently than easy tests; difficulty was defined by proximity of the stimulus duration to the subjective middle of the shortest and longest durations). The decline in accuracy as a function of stimulus difficulty was more pronounced when tests could not be declined (forced test) compared to tests that could have been declined (choice test). Error bars represent standard errors. (Adapted from Foote, A. L., & Crystal, J. D. (2007). Metacognition in the rat. Current Biology, 17, 551-555. © 2007 by Elsevier Ltd. Reprinted with permission.)

Alternative explanations of information monitoring in animals

Smith and colleagues (2008) used quantitative modeling to demonstrate that low-level mechanisms can produce both apparently functional use of the decline response and the Chosen-Forced performance advantage. Importantly, it is not necessary to propose metacognition in order to implement these alternative, low-level mechanisms. Consequently, the formal modeling suggests that the information-monitoring prevailing standard is inadequate to document metacognition.

Smith et al. (2008) restricted the development of their quantitative model to basic associative and habit formation principles. Their proposal follows¹. Direct reward of the decline response produces a low-frequency tendency to select that response independent of the stimulus in the primary discrimination. Importantly, Smith et al. proposed that the decline response has a constant attractiveness across the stimulus continuum. Therefore, the tendency to produce the response is constant across stimulus conditions. We refer to this class of threshold explanations as a stimulus-independent hypothesis to contrast it with stimulus-response hypotheses (Crystal and Foote 2009a). For the primary discrimination, Smith et al. used standard assumptions about exponential decay of a stimulus (i.e., generalization decrements for an anchor stimulus in a trained discrimination). The exponential decay proposal has independent empirical and theoretical support (Shepard 1961; Shepard 1987; White 2002). Thus, the primary discrimination and the decline option give rise to competing response-strength tendencies, and Smith et al. proposed a winner-take-all response rule (i.e., the behavioral response on a given trial is the one with the highest response strength). A schematic of the model is shown in Figure 4a. Simulations with this quantitative model document that it can produce both aspects of the information-monitoring prevailing standard (Figure 4b): the decline response increases as a function of task difficulty, and the Chosen-Forced performance advantage also increases as a function of task difficulty. Importantly, both empirical features of putative metacognition data are produced by the simulation (Figure 4b) without proposing metacognition.

Figure 4.

Schematic of low-level, response-strength model and simulation of information monitoring. (a) Presentation of a stimulus gives rise to a subjective level or impression of that stimulus. For any given subjective level, each response has a hypothetical response strength. The schematic outlines response strengths for two primary responses in a two-alternative forced-choice procedure and for a third (i.e., decline or uncertainty) response (labeled threshold). Note that response strength is constant for the third response (i.e., it is stimulus independent). By contrast, response strength is highest for the easiest problems (i.e., the extreme subjective levels). Note also that for the most difficult problems (i.e., middle subjective levels) the decline-response strength is higher than the other response strengths. Adapted from Smith et al. (2008). (b) Simulation of schematic shown in (a). Simulation of a response-strength model with a flat threshold produces apparently functional use of the decline response (i.e., intermediate, difficult stimuli are declined more frequently than easier stimuli). The Choice-Forced performance advantage emerges as a function of stimulus difficulty. Adapted from Smith et al. (2008). (From Crystal J.D. & Foote A.L. 2009. Metacognition in animals. Comparative Cognition & Behavior Reviews, 4, 1-16. © Crystal J.D. & Foote A.L. Reprinted with permission.)

The schematic in Figure 4a is tailored to capture the design of the experiment by Foote and Crystal (2007), and the model of Smith et al. (2008) clearly can explain the data from Foote and Crystal. In addition, it is possible to apply the same formal model to explain Hampton's (2001) data without appeal to metacognition. The application of the model follows (Figure 5). Presentation of a stimulus gives rise to a representation of the stimulus. As retention interval increases, the stimulus trace is expected to decay exponentially as depicted in Figure 5. The decline response is modeled by a constant level of attractiveness, as proposed by Smith and colleagues. Therefore, variation of retention interval is equivalent to a trace-decay continuum for a fading stimulus trace. It is possible that the monkey’s performance depicted in Figure 2 could be based entirely on a representation of trace strength. According to this view, use of the decline response is based on the relative strength of a fading memory trace just as the old-new responses from the primary task are based on a fading memory trace by application of a winner-take-all rule. Behavior that is driven directly by a fading memory trace need not be based on knowledge about the fading memory trace.

Figure 5.

Schematic of low-level, response-strength model of metamemory. Presentation of a stimulus gives rise to a fading memory trace after stimulus termination. Trace decay (which is shown on the horizontal axis) grows as a function of retention interval. A low-frequency threshold is used for the decline response. Note that response strength is constant for the decline response. By contrast, memory response strength is highest for the shortest retention intervals. Note that for the most difficult problems (i.e., long retention intervals) the decline response strength is higher than the memory response strength. Also note that the horizontal axis may be viewed as a primary representation (see text for details). (From Crystal J.D. & Foote A.L. 2009. Metacognition in animals. Comparative Cognition & Behavior Reviews, 4, 1-16. © Crystal J.D. & Foote A.L. Reprinted with permission.)

It is helpful to distinguish between primary and secondary representations (Carruthers 2008; Crystal and Foote 2009a). The presentation of a stimulus gives rise to an internal representation of that stimulus (which we refer to as the primary representation). Behavior is often based on a primary representation. For example, when presented with an item on a memory test, it is possible to evaluate familiarity with the item to render a judgment that the item is new or old. By contrast, metacognition involves a secondary representation which operates on a primary representation. For example, a person might know that he does not know the answer to a question, in which case appropriate actions might be taken (such as opting out of an immediate test or obtaining additional information). To document metacognition, we need to be certain that performance is not based on the primary representation before we can assign performance to the operation of a secondary representation. Although it may be difficult to determine which type of representation is used by an animal, we should be extremely cautious about attributing performance to a secondary representation (i.e., metacognition) when the data can be explained by a primary representation. Because the same primary representation (i.e., the same fading memory trace) may be used for both the memory task and the decline response, a secondary representation is not needed to explain Hampton's (2001) data.

The use of two different responses (decline and matching responses) does not, in itself, indicate that the two responses are based on different types of representations. The interpretive problem here is how to determine if the monkey is responding on the basis of a primary representation (i.e., a very weak stimulus representation) or on the basis of a secondary representation (i.e., the monkey knows that it does not know the correct answer). It is not sufficient to claim that the use of a memory task will, by definition, result in secondary representations about memory (thereby definitionally constituting evidence for metacognition). What data specifically implicates the use of a secondary representation? Before Smith and colleagues' quantitative modeling, the answer to this question was that the Chosen-Forced performance advantage could not be explained without appeal to metacognition. However, this pattern of data can be explained by a low-level, non-metacognition model. The burden of proof, in this situation, is on providing evidence that implicates a secondary representation, and until such evidence is provided, the cautious interpretation is to claim that a primary representation is sufficient to explain the data.

The discussion above suggests that an information-monitoring approach to metacognition in animals faces significant barriers. Our view is that studies of putative metacognition from an information-monitoring perspective have yet to demonstrate the use of of a secondary representation, which is a significant probelm for the field of comparative metacognition. Perhaps information-seeking in animals is free from these barriers to assessing metacognition. In the sections that follow, we outline the information-seeking approach and develop an alternative explanation for information seeking in animals.

Information seeking in animals

Although the data reviewed below may be open to alternative explanations that do not involve metacognition, here we emphasize the availability of a metacognition account. In a later section, we outline an alternative account that does not involve metacognition. In the section labeled Parsimony, we outline our views about selecting between multiple, competing accounts of data.

In the tube tipping experiments, a non-human primate is positioned in front of two or more tubes. Food may be placed at the far end of the tubes. The location of food may be determined by visually observing the baiting or by bending down and looking through the tube. Before having the opportunity to search for the bait in tube tipping experiments, the subject sometimes observes opaque tubes being baited (seen trials) but is sometimes prevented from seeing the baiting (unseen trials). The observations that (1) the subjects look more if baiting was not seen and (2) the subjects are more accurate if baiting was seen have been taken as evidence for metacognition using an information-seeking approach (Call and Carpenter 2001; Hampton et al. 2004). We refer to this pattern of data as the information-seeking prevailing standard. If the animal knows that it does not know the baited location, then it would be expected to look inside the tubes before choosing a tube. By contrast, if it knows that it knows the location of a bait, then it would choose the tube immediately (e.g., by tipping it) to obtain food without initially looking inside the tube. Hence, a metacognition account of information seeking predicts more looking before choosing when the exact location of the bait cannot be determined. Of course, accuracy in tube tipping after looking would be higher than when the animals randomly tip tubes without looking. Call and colleagues (Call and Carpenter 2001; Call 2010) reported evidence for metacognition in apes (chimpanzees, orangutans, gorillas, and bonobos) and Hampton, Zivin and Murray (2004) reported evidence for metacognition in rhesus monkeys using information-seeking approaches.

The metacognitive information-seeking perspective predicts that as accuracy on both seen and unseen trial types decreases, the frequency of bending down to look into a tube increases. Thus, using information seeking would allow the subject to increase accuracy if tube baiting was unseen. Importantly, if they know that they know the location of the bait, they should be able to retrieve the bait without looking for it first, and as a result, errors should be rare. Conversely, if observation of the tube baiting is prevented, the subject will not know in which tube the bait was placed. As a result, the subject would need to bend down to look inside the tube first in order to find the food and then obtain it, thereby increasing the number of looks needed to find the reward. Consequently, the subject should look more frequently when baiting was not observed. Although the above account of the data is consistent with an information-seeking metacognitive perspective, it is important to evaluate if alternative non-metacognition perspectives can also account for the data. We develop an alternative explanation in the next sections.

An example of information seeking in animals

Call and Carpenter (2001) tested chimpanzees using an information-seeking approach. Three tubes were positioned in front of the chimpanzee. In the seen condition, one of the tubes was baited within view of the subject. In the unseen condition, a screen blocked the subject’s view of baiting (see Figure 6a). In the unseen condition the subject could seek more information about the location of a hidden bait by bending down and peering into the tubes to find the location of the bait. Call and Carpenter made the following two predictions: 1. The subjects should look in the tubes less often in the seen condition than in the unseen condition, and 2. subjects should be more accurate in the unseen condition when they peer into the tubes before selecting a tube than when they do not look before choosing. Representative data shows that the chimpanzees looked into the tubes more often in the unseen condition than in the seen condition (see Figure 6b). Moreover, they were more accurate after looking in the unseen condition (see Figure 6c).

Figure 6.

Schematic representation of information-seeking design of study and data. (a) Procedure for non-human primates (left panel; Call and Carpenter 2001). In the seen condition, the experimenter (E) baited one of two or more tubes in front of the subject (S). In the unseen condition a screen blocked the subject's view of the baiting. The platform was presented to the subject 5 seconds after baiting. The subject could look before choosing or choose immediately. (b) The bars show the percentage of trials during which the chimpanzees looked in the tubes in a three-tube choice experiment. Chimpanzees looked into the tubes less often in the seen condition than in the unseen condition. (Adapted from Call, J., and Carpenter, M. (2001). Do apes and children know what they have seen? Animal Cognition 3, 207-220.)

Development of an alternative explanation for information seeking in animals

We propose that relatively simple alternative hypotheses that do not posit metacognition may explain putative evidence for information seeking in animals. The alternative explanation uses basic principles of learning (Staddon 2010) and spatial cognition (Brown and Cook 2006). The alternative explanation focuses on two relatively simple principles: First, an animal has a default "look before you go" response which supersedes random searches in space. Second, spatially guided behavior follows a default rule of "go where something good is." These two principles can explain the results of tube tipping experiments without proposing metacognition.

Looking behavior can be explained via the stimulus-independent approach, using relatively simple principles of spatial cognition. Specifically, spatial cognition employs the use of a spatial navigation module and a search module. In the presence of a spatial cue, the spatial navigation module directs an animal to “go where something good is” (i.e., go to the baited tube). By contrast, in the absence of a spatial cue, the search module would direct the animal to “look before you go” (i.e., check the tubes for bait; see Figure 7).

Figure 7.

A stimulus-independent account of information seeking. (a) The response strength of a spatially guided response as a function of spatial location when food is present (the baiting-observed condition in the information-seeking paradigm). The response strength of a spatially guided response is more accurate in the presence of food than the response strength for the default response. On average, a spatially guided "go" response dominates over randomly directed "look" responses. (b) The response strength of a spatially guided response as a function of spatial location when food is not present (the baited-unobserved condition in the information-seeking paradigm). When food is not present, the response strength for the default response is greater than the response strength of the spatially guided response. On average, randomly directed "look" responses dominate over spatially guided "go" responses.

Spatial gradient: "Go where something is good"

The stimulus-independent approach proposes two processes. According to the first process, baiting gives rise to a response strength gradient centered on the location recently seen with food during baiting. Spatially mediated searching is based on a response strength gradient producing searches that are near locations that have recently been observed to have food. Spatially guided behavior follows the rule, “go where something is good” (labeled as “go” in Figure 7a). If the animal has seen the baiting, it would “go” directly to the target on average. Thus, accurate performance in the presence of a seen target would be the result of the response strength being peaked around the spatial location of the target, thereby producing accurate spatially guided search with little error. Figure 7a depicts a typical case with "go" winning out over "look". However, variability in the relative heights and shape of each function in Figure 7a permits exceptions to the typical case whereby a "look" will occur on some rare occasions.

Default response: "Look before you go"

The second proposed process can be thought of as a default rule which leads the animal to “look before you go” (labeled as “look” in Figure 7b). The default response is a relatively low, flat function that does not vary across spatial locations. According to the default rule, when the animal has not seen the baiting, it would be expected to “look” randomly until it finds the designated target, thereby producing a relatively high error rate on average. A random search occurs if the response strength for the default response is greater than the response strength for the “go” response. Figure 7b depicts a typical case with "look" winning out over "go". However, variability in the relative heights and shape of each function in Figure 7b permits exceptions to the typical case whereby a "go" response will occur on some rare occasions.

Predictions

The proposed stimulus-independent model makes the following predictions: If baiting is seen (Figure 7a), then response strength for the “go” response is higher than the default “look” response. Thus, the “go” response wins. Because the “go” response strength is spatially guided (i.e., the response distribution is centered on the target), the “go” response will be directed toward the location of the target (i.e., at or near where the baiting occurred). Moreover, looks (i.e., bending down to examine the tubes) will be rare because the “go” response strength is higher than the “look” response strength on average. If baiting is not seen (Figure 7b), then the spatially guided response distribution (i.e., “go”) is suppressed (to zero in Figure 7b). Thus, the “look” response strength dominates. Note that the default “look” response threshold is constant as a function of spatial location. Consequently, looks are randomly distributed in space. The decision to respond ("go" or "look") may be based on an evaluation of response strength as a function of spatial location (see Figure 7a) without application of information seeking. According to the stimulus-independent approach, accuracy and looking into tubes can be predicted without requiring knowledge of having seen or not seen the tubes being baited. Therefore, looking behavior need not imply that the animal knows that it does not know (i.e., information seeking in the metacognition sense). Thus, although tube tipping experiments (e.g., Call and Carpenter 2001; Hampton et al. 2004) appear to demonstrate evidence of metacognition, the application of the stimulus-independent approach suggests that information-seeking paradigms are also subject to explanation by a simple associative model that is similar to Smith et al.’s (2008) model.

Unlike metacognition in an information-seeking approach, the stimulus-independent approach states that accuracy is a function of the relative response strength for a spatial target and the default threshold. More specifically, if the response strength is viewed as a gradient centered on the location of food, the decision to respond may be thought of as an evaluation of response strength as a function of spatial location (see Figure 7a).

Why is a subject more accurate if baiting was seen?

We propose that the subject samples from the "go" and the "look" distributions shown in Figure 7, and based on a winner-take-all rule, it chooses to go to a spatial location determined from the "go" distribution (if the "go" value is greater than the "look" value). If baiting was seen, a sample from the spatial distribution shown in Figure 7a will likely generate a correct response. If food is not obtained, we propose that the random samples and winner-take-all decision are repeated to generate the next response. Thus, if baiting was not seen, several samples from Figure 7b will likely be required to produce a correct response, thereby producing relatively high error rates.

Why does an animal look more if baiting was not seen?

If baiting was not seen, then it is likely that a sample from the "look" distribution is greater than a sample for the "go" distribution (Figure 7b). If the look does not reveal the location of the bait, we propose that the random samples and winner-take-all decision are repeated to generate the next response. Because the "look" distribution is a constant threshold, random samples from this distribution are equally likely for each spatial location. Therefore, a relatively large number of random samples will be required before a "look" response finds the actual bait.

Potential solutions

In a series of elegant experiments, Call (2010) used a number of experimental designs aimed at eliminating non-metacognition alternative hypotheses. In one experiment, baiting could be seen and/or heard (i.e., the experimenter shook a tube so that the animal could hear that food was present). Another manipulation required greater effort to see the end of a tube because it was placed in an oblique orientation. Each animal had participated in earlier experiments in which they had the opportunity to learn to use the noise made by hidden food to locate a baited cup; fewer than half of the animals learned to use noise. Apes looked less frequently in the visible condition compared to the shaken condition. Apes that had passed a noise pretest (i.e., subjects that learned to use the noise cue derived from shaking) were less likely to look inside the tube in the shaken condition compared to the unseen condition. In addition, subjects were less likely to look when the cost of looking was high. Call argues that the data suggest that the apes did not use a fixed sequence of looking followed by choosing, but instead integrated auditory information. We propose that a spatial representation (activation as in Figure 7a) could be set based on multi-modal cues (as in the shaken condition), which would generate choices ("go where something good is") without looking by animals that have learned the significance of the auditory cue in subsequent spatially guided searching. Of course, an animal that had not learned the significance of an auditory cue for spatially guided searching would not be expected to have a spatial representation activated by an auditory cue (as in Figure 7b), in which case such an animal would frequently look in a tube before making a choice.

In another experiment, Call (2010) manipulated the retention interval between seen baiting and the opportunity to respond. The apes were more likely to look as a function of increasing delay, which also corresponded to increasing forgetting. As we argued above, a fading stimulus trace is expected to reduce response strength for the spatially guided choice (as in Figure 5), which would produce an increase in looking before choosing.

In a final manipulation, Call (2010) manipulated the quality of food type used in the baiting procedure. The rationale was that the apes look in the tube for two reasons. When they do not know the location of food, the animal looks to find it. In addition, perhaps the animal also looks in the tube when they indeed remember the baited location, but in this case it is checking to see that it is correct. Call argued that increasing the value of the reward would increase the frequency of checking by the animal to verify that it is indeed correct about the location. Looks in the visible condition were higher for high-value than low-value foods. It is somewhat disheartening that the same behavior (looking) can be taken as evidence for knowledge and the lack of knowledge of a food location. Indeed these data are not an example of an animal knowing that it does not know something, but rather are offered as an example of the animal knowing that it does know something and doing a behavior that would be expected if it did not know. Despite the above concern over conceptual clarity, the pattern of data is not easily explained by a response strength account. A response strength hypothesis is compatible with the proposal that high-value food generates a higher response strength than low-value food (i.e., a high motivation to go to something good). A higher response strength, overall, would predict less looking. It is not clear why a high value food would generate lower response strength, which would be required to predict higher rates of looking. However, given the conceptual problem outlined above, it would be valuable to have clearer evidence of metacognition.

Parsimony

We have reviewed two views to account for tube-tipping experiments. According to an information-seeking account, an animal has access to internal states of knowledge, discriminates those states, and takes appropriate action when information is needed (i.e., it seeks out the missing information). According to our alternative explanation, animals have two (likely among many other) instinctive rules (go where something is good, look before you go) which can account for putative information seeking without proposing that animals have access to internal states of knowledge.

Concluding that metacognition explains an animal's behavior requires excluding more parsimonious alternative accounts. Inherently, the selection process involves defining evidence of metacognition by exclusion – only when other accounts are not adequate to explain data. Validation of methods to document metacognition is an essential step, and we recommend that researchers do not skip this essential step.

How is parsimony to be evaluated? One approach is to count the number of proposed processes. Yet, we believe that the proposal that animals have access to internal states is inherently more complex than simpler basic rules. We acknowledge that the assessment of parsimony is somewhat subjective. For example, it may be argued that metacognition is efficient or a relatively simple process that may not include the rich and complex aspects that are ascribed to it by people. However, we recommend that it is advisable to accept a metacognition account of information-seeking only if the data cannot be explained by any proposed alternative explanation.

Conclusions

Although considerable attention has been paid recently toward critically assessing putative evidence for metacognition in information-monitoring experiments in animals, we believe that many of the same problems apply to the case of evaluating evidence for metacognition in information-seeking experiments in animals. A large body of research shows that before having the opportunity to search for a bait, animals look more if baiting was not seen and are more accurate if baiting was seen. Although these patterns of data are consistent with a metacognition interpretation of information seeking, they are also consistent with simple alternative hypotheses that do not posit metacognition. The alternative explanation focuses on two relatively simple principles: First, an animal has a default "look before you go" response which supersedes random searches in space. Second, spatially guided behavior follows a default rule of "go where something good is." These two principles can explain the results of tube tipping experiments without proposing metacognition.

New methods are needed to provide an independent line of evidence for metacognition in non-human animals. Importantly, this independent line of evidence would need to be based upon a secondary representation that cannot be explained by simpler proposals. The development of new methods should be guided by careful consideration of alternative hypotheses. We recommend working with quantitative models of alternative hypotheses (e.g., Smith et al. 2008) to validate predictions with simulations.