In their target article, Newell & Shanks (N&S) identify methodological problems in the study of unconscious influences on decision making. Because awareness is indexed by subjects' reports about what they know and how they know it, such studies seek, in effect, to demonstrate dissociations of cognition and metacognition: One component of the design shows that information of some kind is influencing behavior in some way; a second component elicits subjects' reports about the information they possess and the manner in which they are using it. Evidence for unconscious influence is obtained when the (relatively indirect) cognitive measures and the (relatively direct) metacognitive measures paint inconsistent pictures of the underlying mental process.
As the authors note, such studies are only as compelling as the metacognitive measures they use – and measures lacking in reliability, relevance, immediacy, and sensitivity are often employed. Indeed, research on unconscious influence suffers from its own distinctive array of perverse incentives. As in other areas of psychology, the researcher is driven to obtain evidence for a clear effect on the cognitive measure. This incentive fosters practices that make Type I errors more likely (Simmons et al. Reference Simmons, Nelson and Simonsohn2011). In the study of unconscious influences, the researcher typically has a second incentive – to fail to find evidence of a significant effect on the metacognitive measure. This incentive may foster practices that make Type II errors more likely in the assessment of awareness. One effective way of failing to detect awareness is the use of unreliable, irrelevant, insensitive, and/or belated probes. (Ironically, such bias in the choice of probes might itself be unconscious.)
Complicating matters, it is not always easy to formulate, let alone satisfy, the central criterion of relevance – that “assessments should target only information relevant to the behavior” (target article, Table 1). What is relevant to a behavior depends on what causes it. As a result, subtly flawed or imprecise causal theories of behavior can lead even the well-meaning and careful researcher to misidentify relevance in designing a measure of awareness.
Although these methodological problems are important, and although they challenge some influential findings, we believe that the target article understates the full empirical and theoretical case for unconscious influences on complex behavior. To make the empirical case adequately would require a counter-review rather than a commentary. Recent reviews of unconscious cognition that are more comprehensive and, in our view, also more balanced are provided by Kouider and Dehaene (Reference Kouider and Dehaene2007) and by Simons et al. (Reference Simons, Hannula, Warren, Day, Zelazo, Moscovitch and Thompson2007). The empirical study of unconscious cognition has its share of murky bathwater, but we believe the outlines of a baby are distinctly discernible within it.
Critically, the target article also understates the theoretical case for unconscious influences in complex behavior. N&S suggest that such influences make for “good stories,” and that they confirm “strong ex ante beliefs” about mental causation that soften the critical judgment of researchers and journal editors. The explanatory role of unconscious influences is otherwise dismissed, as when the authors state that we do not “need to posit ‘magical’ unconscious processes producing answers from thin air” (sect. 6.2). Are unconscious processes mere explanatory magic?
As we noted above, the hypothesis of unconscious influence is observationally equivalent to the claim that cognition and metacognition are imperfectly coupled and sometimes strongly dissociate (because “conscious awareness” is measured by metacognitive report). In this regard, unconscious processes are no more “magical” than any other functional dissociation in cognition. Such processes are predicted by any cognitive architecture that represents metacognition as a limited subset or partial aspect of the mind.
For example, consider Baars's (Reference Baars2005) global workspace theory (GWT). Contrary to the target article's cursory account of it, this model is motivated by basic computational problems in cognition. Behavioral and neurophysiological investigations suggest the existence of multiple semi-independent “modules” specialized for different facets of information processing. This division of cognitive labor solves some problems, but it also creates a problem. Specifically, information from different modules needs to be integrated to represent arbitrary perceptual combinations, solve unfamiliar problems not handled by any one module, organize motor programs around a coherent action plan, and build an internal model of “the self.” To address this problem, GWT proposes that only a small subset of relevant information is selected for “broadcasting” across the network. Integration is thus obtained, but it is incomplete and comes at the expense of reduced information bandwidth and processing speed. The theory makes sense of the local patterns of neural activity (with relative inactivity in globally connected association areas) that are observed in subliminal priming experiments and in the behavioral automatisms of sleepwalking, epilepsy, and the vegetative state (Baars Reference Baars2005). This architecture implies that non-selected information can bias behavior, but without flexible integration or accessibility to report.
Similar predictions are made by other models that, in principle, distinguish the process of metacognitive report from other processes. For example, well-known models of learning and memory distinguish between procedural and declarative systems (Squire Reference Squire1992). While the systems are thought to interact in the control of complex behavior, declarative knowledge of procedural mechanisms is at best indirect. Other models highlight constraints on the (coarse-grained) format of metacognitive representations. These constraints may limit the kinds of information that are available to report (Winkielman & Schooler Reference Winkielman and Schooler2011). Notably, dissociations are even possible within metacognition, as when we overtly report that certain states are ineffable – we experience more than we can overtly describe or express.
In any event, the claim that some cognitive operations are inaccessible to metacognition is not magical, but conceptually coherent and consistent with current knowledge. It predicts systematic mismatches between cognitive processes and subjects' overt reports about those processes – even when probes of awareness are reliable, relevant, immediate, and sensitive. Indeed, we would be interested to see the authors propose a principled sketch of a cognitive architecture in which cognition and metacognition are inseparable. To us, such a panpsychic architecture sounds like magic.
In their target article, Newell & Shanks (N&S) identify methodological problems in the study of unconscious influences on decision making. Because awareness is indexed by subjects' reports about what they know and how they know it, such studies seek, in effect, to demonstrate dissociations of cognition and metacognition: One component of the design shows that information of some kind is influencing behavior in some way; a second component elicits subjects' reports about the information they possess and the manner in which they are using it. Evidence for unconscious influence is obtained when the (relatively indirect) cognitive measures and the (relatively direct) metacognitive measures paint inconsistent pictures of the underlying mental process.
As the authors note, such studies are only as compelling as the metacognitive measures they use – and measures lacking in reliability, relevance, immediacy, and sensitivity are often employed. Indeed, research on unconscious influence suffers from its own distinctive array of perverse incentives. As in other areas of psychology, the researcher is driven to obtain evidence for a clear effect on the cognitive measure. This incentive fosters practices that make Type I errors more likely (Simmons et al. Reference Simmons, Nelson and Simonsohn2011). In the study of unconscious influences, the researcher typically has a second incentive – to fail to find evidence of a significant effect on the metacognitive measure. This incentive may foster practices that make Type II errors more likely in the assessment of awareness. One effective way of failing to detect awareness is the use of unreliable, irrelevant, insensitive, and/or belated probes. (Ironically, such bias in the choice of probes might itself be unconscious.)
Complicating matters, it is not always easy to formulate, let alone satisfy, the central criterion of relevance – that “assessments should target only information relevant to the behavior” (target article, Table 1). What is relevant to a behavior depends on what causes it. As a result, subtly flawed or imprecise causal theories of behavior can lead even the well-meaning and careful researcher to misidentify relevance in designing a measure of awareness.
Although these methodological problems are important, and although they challenge some influential findings, we believe that the target article understates the full empirical and theoretical case for unconscious influences on complex behavior. To make the empirical case adequately would require a counter-review rather than a commentary. Recent reviews of unconscious cognition that are more comprehensive and, in our view, also more balanced are provided by Kouider and Dehaene (Reference Kouider and Dehaene2007) and by Simons et al. (Reference Simons, Hannula, Warren, Day, Zelazo, Moscovitch and Thompson2007). The empirical study of unconscious cognition has its share of murky bathwater, but we believe the outlines of a baby are distinctly discernible within it.
Critically, the target article also understates the theoretical case for unconscious influences in complex behavior. N&S suggest that such influences make for “good stories,” and that they confirm “strong ex ante beliefs” about mental causation that soften the critical judgment of researchers and journal editors. The explanatory role of unconscious influences is otherwise dismissed, as when the authors state that we do not “need to posit ‘magical’ unconscious processes producing answers from thin air” (sect. 6.2). Are unconscious processes mere explanatory magic?
As we noted above, the hypothesis of unconscious influence is observationally equivalent to the claim that cognition and metacognition are imperfectly coupled and sometimes strongly dissociate (because “conscious awareness” is measured by metacognitive report). In this regard, unconscious processes are no more “magical” than any other functional dissociation in cognition. Such processes are predicted by any cognitive architecture that represents metacognition as a limited subset or partial aspect of the mind.
For example, consider Baars's (Reference Baars2005) global workspace theory (GWT). Contrary to the target article's cursory account of it, this model is motivated by basic computational problems in cognition. Behavioral and neurophysiological investigations suggest the existence of multiple semi-independent “modules” specialized for different facets of information processing. This division of cognitive labor solves some problems, but it also creates a problem. Specifically, information from different modules needs to be integrated to represent arbitrary perceptual combinations, solve unfamiliar problems not handled by any one module, organize motor programs around a coherent action plan, and build an internal model of “the self.” To address this problem, GWT proposes that only a small subset of relevant information is selected for “broadcasting” across the network. Integration is thus obtained, but it is incomplete and comes at the expense of reduced information bandwidth and processing speed. The theory makes sense of the local patterns of neural activity (with relative inactivity in globally connected association areas) that are observed in subliminal priming experiments and in the behavioral automatisms of sleepwalking, epilepsy, and the vegetative state (Baars Reference Baars2005). This architecture implies that non-selected information can bias behavior, but without flexible integration or accessibility to report.
Similar predictions are made by other models that, in principle, distinguish the process of metacognitive report from other processes. For example, well-known models of learning and memory distinguish between procedural and declarative systems (Squire Reference Squire1992). While the systems are thought to interact in the control of complex behavior, declarative knowledge of procedural mechanisms is at best indirect. Other models highlight constraints on the (coarse-grained) format of metacognitive representations. These constraints may limit the kinds of information that are available to report (Winkielman & Schooler Reference Winkielman and Schooler2011). Notably, dissociations are even possible within metacognition, as when we overtly report that certain states are ineffable – we experience more than we can overtly describe or express.
In any event, the claim that some cognitive operations are inaccessible to metacognition is not magical, but conceptually coherent and consistent with current knowledge. It predicts systematic mismatches between cognitive processes and subjects' overt reports about those processes – even when probes of awareness are reliable, relevant, immediate, and sensitive. Indeed, we would be interested to see the authors propose a principled sketch of a cognitive architecture in which cognition and metacognition are inseparable. To us, such a panpsychic architecture sounds like magic.