Publications

A longstanding goal in metacognition research is the quantification of confidence-related performance independently from task-related performance. While there is now a proliferation of generative models of confidence in numerous psychological tasks, the meta-d' model remains a popular approach to estimating metacognitive sensitivity and metacognitive bias in the framework of Signal Detection Theory. However, as research questions in metacognition research increase in complexity, it remains an ongoing challenge to develop extensions of the meta-d' model built to deal with longitudinal, repeated-measures, or other complex experimental designs. Here we introduce the open-source R package hmetad, which allows users to easily fit the meta-d' model for a wide range of experimental designs using a familiar formula syntax. We also review how the meta-d' model can be validated in a Bayesian workflow focused on a range of posterior estimates. Overall, the hmetad package offers a comprehensive interface to fitting, evaluating, and visualizing signal detection theoretic models of metacognitive performance.

Confidence estimates are often “detection-like”—driven by positive evidence in favour of a decision. This empirical observation has been interpreted as showing that human metacognition is limited by biases or heuristics. Here, we show that Bayesian confidence estimates also exhibit heightened sensitivity to decision-congruent evidence in higher-dimensional signal detection theoretic spaces, leading to detection-like confidence criteria. This effect is due to a nonlinearity induced by normalisation of confidence by a large number of unchosen alternatives. Our analysis suggests that detection-like confidence is rational when participants consider a greater number of hypotheses than assumed by the experimenter. Further, we show that a similar dimensionality-driven mechanism can give rise to and modulate the strength of the positive evidence biases in convolutional neural networks, linking our signal detection theoretic analysis to confidence behaviour in artificial systems.

We plan by considering multiple alternatives, yet only one action is ultimately performed. How memory later distinguishes executed actions from those that were merely considered, and what metacognitive confidence reveals about this distinction, remains unclear. Across two experiments, participants planned and executed routes through mazes before judging whether specific locations had belonged to the route they took or to an unexecuted 'path not taken'. Confidence was assessed either before or after route navigation. Confidence carried reliable trial-level information about the quality of route choice, indicating that it functioned as a genuine subjective signal about action-planning quality rather than as a global response tendency. Higher confidence was also associated with greater discriminability between executed and unexecuted routes in memory, indicating more distinct action representations, but showed no reliable relationship with the criterion used to endorse a memory. These effects were equivalent whether confidence was assessed before or after action execution. Together, these findings suggest that confidence indexes the precision of action representations during planning and that false memories for unexecuted alternatives arise from reduced representational distinctiveness rather than from changes in decision policy.

Research on metacognition critically relies on psychophysical measures that are independent of task performance. Although there exist well-validated measures of metacognitive efficiency (i.e., the degree to which confidence ratings distinguish between correct and incorrect decisions), relatively less attention has been paid to developing corresponding measures of metacognitive bias (i.e., the overall tendency to make responses with high confidence). Most researchers currently use mean confidence as a proxy for metacognitive bias, however this measure is confounded by changes in accuracy, response bias, and metacognitive sensitivity. Here, we introduce a new signal detection theoretic measure of metacognitive bias (meta-delta) and demonstrate that it is independent of a range of potential confounds. Then, by analyzing longitudinal data, we show that this measure is reliable enough to use in psychological research. Overall, we advocate for the use of meta-delta as a principled measure of metacognitive bias.

Statistical normality robustly influences causal judgments of binary causal variables (events which either do or do not occur): people judge statistically abnormal events as more causally responsible when causes are jointly necessary for an outcome, but as less responsible when causes are individually sufficient. However, whether this pattern extends to continuous causes, which vary in degree, remains unknown. Across two experiments (N = 1,184 U.S. Prolific participants), we examined this question using a learning paradigm in which participants learned the statistical properties of two continuous causes before making a final causal judgment. We manipulated normality via variance (Experiment 1) or mean (Experiment 2). Despite successful manipulation checks, we found strong evidence that causal judgments were unaffected by normality in both experiments, with no corresponding effects on confidence. These findings suggest that normality effects, while seemingly robust in the field of causation, may be specific to causal structures involving binary variables.

Many cognitive scientists and philosophers take cases of double prevention to be one of the primary motivations for accepting causal pluralism, the view that people have multiple concepts of causation. Thanawala and Erb (2024) argue against Lombrozo (2010)’s account of causal pluralism. They find that the temporal order of events affects people’s causal judgments in double prevention cases, and they argue that this finding is not easily explained by prominent versions of causal pluralism or by counterfactual theories. In contrast to this interpretation, we argue that counterfactual thinking can explain their findings. On this explanation, the temporal order of events affects the extent to which people simulate counterfactual alternatives to these events. We show that under this assumption, a recent counterfactual model of causal judgment can reproduce all qualitative effects of temporal order found in Thanawala and Erb (2024)’s new work. Our findings complement past research that applied the counterfactual framework to temporal-order effects and double prevention cases independently, suggesting that these explanations are highly generalizable.

If Peter prevents Jack from catching a falling bottle that Mike knocked over, most people would think that Mike caused the spill to a greater degree than Peter. Cases of double prevention like these are famously inconsistent with the idea that causal judgments rely on counterfactual dependence; the spill wouldn't have happened if Mike hadn't knocked the bottle over or if Peter hadn't prevented Jack from catching the bottle. But newer counterfactual models are more flexible, and they assume that people imagine different counterfactuals in proportion with their perceived normality. Following recent work showing that these newer models can account for causal judgments in cases of double prevention, here we find that normality affects such judgments. Specifically, when the productive factor is normal and the double preventer is abnormal, we find that participants preferentially rate either the productive factor or the double preventer as more causal depending on the normality of the possible preventer. Contrary to standard interpretations, then, our results suggest that cases of double prevention are actually more problematic for competing theories of causal judgment than they are for counterfactual theories.

Counterfactual theories propose that people’s capacity for causal judgment depends on their ability to consider alternative possibilities: The lightning strike caused the forest fire because had it not struck, the forest fire would not have ensued. To accommodate a variety of psychological effects on causal judgment, a range of recent accounts have proposed that people probabilistically sample counterfactual alternatives from which they compute a graded measure of causal strength. While such models successfully describe the influence of the statistical normality (i.e., the base rate) of the candidate and alternate causes on causal judgments, we show that they make further untested predictions about how normality influences people’s confidence in their causal judgments. In a large (N = 3,020) sample of participants in a causal judgment task, we found that normality indeed influences people’s confidence in their causal judgments and that these influences were predicted by a counterfactual sampling model in which people are more confident in a causal relationship when the effect of the cause is less variable among imagined counterfactual possibilities.

How do people evaluate causal relationships? Do they just consider what actually happened, or do they also consider what could have counterfactually happened? Using eye tracking and Gaussian process modeling, we investigated how people mentally simulated past events to judge what caused the outcomes to occur. Participants played a virtual ball-shooting game and then—while looking at a blank screen—mentally simulated (a) what actually happened, (b) what counterfactually could have happened, or (c) what caused the outcome to happen. Our findings showed that participants moved their eyes in patterns consistent with the actual or counterfactual events that they mentally simulated. When simulating what caused the outcome to occur, participants moved their eyes consistent with simulations of counterfactual possibilities. These results favor counterfactual theories of causal reasoning, demonstrate how eye movements can reflect simulation during this reasoning and provide a novel approach for investigating retrospective causal reasoning and counterfactual thinking.

In cases of double prevention--when one event prevents another from preventing an outcome initiated by a productive factor--people tend to judge the productive factor as causal but the double preventer as non-causal. Recent work demonstrated that this tendency can be explained by appealing to people's agreement with and tendency to consider counterfactuals: asking people to imagine the absence of the double preventer decreased their tendency to view the productive factor as more causal than the double-preventer. These effects were well-explained by the Necessity-Sufficiency (NS) model, which instantiates a particular counterfactual account. Here we asked whether another model, the Counterfactual Effect Size (CES) model, could predict the same effects. We found that the CES model indeed predicted these effects, suggesting that the ability of counterfactual theories to predict causal judgments in cases of double prevention is not unique to the NS model.

When comparing the roles of the lightning strike and the dry climate in causing the forest fire, one might think that the lightning strike is more of a cause than the dry climate, or one might think that the lightning strike completely caused the fire while the dry conditions did not cause it at all. Psychologists and philosophers have long debated whether such causal judgments are graded; that is, whether people treat some causes as stronger than others. To address this debate, we first reanalyzed data from four recent studies. We found that causal judgments were actually multimodal: although most causal judgments made on a continuous scale were categorical, there was also some gradation. We then tested two competing explanations for this gradation: the confidence explanation, which states that people make graded causal judgments because they have varying degrees of belief in causal relations, and the strength explanation, which states that people make graded causal judgments because they believe that causation itself is graded. Experiment 1 tested the confidence explanation and showed that gradation in causal judgments was indeed moderated by confidence: people tended to make graded causal judgments when they were unconfident, but they tended to make more categorical causal judgments when they were confident. Experiment 2 tested the causal strength explanation and showed that although confidence still explained variation in causal judgments, it did not explain away the effects of normality, causal structure, or the number of candidate causes. Overall, we found that causal judgments were multimodal and that people make graded judgments both when they think a cause is weak and when they are uncertain about its causal role.

Mike accidentally knocked against a bottle. Seeing that the bottle was about to fall, Jack was just about to catch it when Peter accidentally knocked against him, making Jack unable to catch it. Jack did not grab the bottle, and it fell to the ground and spilled. In double-prevention cases like these, philosophers and nonphilosophers alike tend to judge that Mike knocking into the bottle caused the beer to spill and that Peter knocking into Jack did not cause the beer to spill. This difference in causal judgment is a difficult puzzle for counterfactual theories of causal judgment; if each event had not happened, the outcome would not have, yet there is a difference in people's causal judgments. In four experiments and three supplemental experiments, we confirm this difference in causal judgments. We also show that differences in people's counterfactual thinking can explain this difference in their causal judgments and that recent counterfactual models of causal judgment can account for these patterns. We discuss these results in relation to work on counterfactual thinking and causal modeling.

People more frequently select norm-violating factors, relative to norm-conforming ones, as the cause of some outcome. Until recently, this abnormal-selection effect has been studied using retrospective vignette-based paradigms. We use a novel set of video stimuli to investigate this effect for prospective causal judgments—that is, judgments about the cause of some future outcome. Four experiments show that people more frequently select norm-violating factors, relative to norm-conforming ones, as the cause of some future outcome. We show that the abnormal-selection effects are not primarily explained by the perception of agency (Experiment 4). We discuss these results in relation to recent efforts to model causal judgment.

Herein we describe CRAMM, a framework for Causal Reasoning via Attention and Mental Models. CRAMM develops and extends assumptions made by a previously developed coun- terfactual simulation model of human causal judgment. We implement CRAMM computationally and demonstrate how it robustly captures human causal judgments about simple two-object interactions at the level of underlying cognitive and perceptual processes, including data on eye-movements that serve as direct evidence for the role of counterfactuals in causal judgment.

Why are agents who act abnormally considered more causal than agents who act normally? Kirfel and Lagnado (2021) claim that agents who act abnormally are better able to foresee the outcomes of their abnormal behavior, giving them an epistemic advantage over agents who act normally. On their view, this advantage explains why people attribute more causality to agents who act abnormally. We tested an extension of this view to cases of negligence. Across three experiments (N = 1713), we found that epistemic advantages are not necessary for inflated attributions of causality to abnormally acting agents. In Experiment 1, participants attributed more causality to an agent who is abnormally negligent relative to an agent who normally does not act, even though they attributed equal degrees of foresight to both agents. In Experiment 2, participants attributed more causality to statistically abnormal negligence relative to statistically normal inaction without any corresponding difference in attributed foresight to either agent. In Experiment 3, we replicated the pattern of results from Experiment 2 and found that people attribute equal degrees of desire and intentionally to normal and abnormal omitters. We argue that Kirfel and Lagnado's epistemic explanation of the abnormal inflation effect in causal judgment does not extend to abnormal negligence precisely because people attribute causality to outcomes that an agent neither foresees nor expects.

Would a penchant for animal cruelty influence how much someone is blamed for lying on a tax return? While extant literature shows that moral character evaluations can influence subsequent moral judgments, it is unclear whether this influence is domain-general or domain-specific. Does receiving information about any morally bad character trait increase blame, or only traits that are similar to the subsequent act? In two experiments (preregistered; N=1320), we investigated how incidental character information influence moral and causal judgments. Participants judged bad agents to be more blameworthy, more causal, and more deserving of punishment than good agents for committing the same act. This effect was robust across how the information was presented (Experiment 1) and different moral domains (Experiment 2). Our findings suggest that moral character exerts a domain-general influence on judgments, challenging the view that only traits that are similar or congruent to a subsequent act influence blame.

How does character information inform judgments of blame? Some argue that character information is indirectly relevant to blame because it enriches judgments about the mental states of a wrongdoer. Others argue that character information is directly relevant to blame, even when character traits are causally irrelevant to the wrongdoing. We propose an empirical synthesis of these views: a two channel model of blame. The model predicts that character information directly affects blame when this information is relevant to the wrongdoing that elicits blame. Furthermore, the effect of character information on blame depends on judgments about the true self that are independent of judgments of intentionality. Across three preregistered studies (N = 662), we found support for all three predictions of the two channel model. We propose that this reflects two distinct functions of blame: a social regulatory function that encourages norm compliance and a pedagogical function that encourages personal improvement.

Moral Foundations Theory (MFT) posits that the human mind contains modules (or “foundations”) that are functionally specialized to moralize unique dimensions of the social world: Authority, Loyalty, Purity, Harm, Fairness, and Liberty. Despite this strong claim about cognitive architecture, it is unclear whether neural activity during moral reasoning exhibits this modular structure. Here, we use spatiotemporal partial least squares correlation (PLSC) analyses of fMRI data collected during judgments of foundation-specific violations to investigate whether MFT’s cognitive modularity claim extends to the neural level. A mean-centered PLSC analysis returned two latent variables that differentiated between social norm and moral foundation violations, functionally segregated Purity, Loyalty, Physical Harm, and Fairness from the other foundations, and suggested that Authority has a different neural basis than other binding foundations. Non-rotated PLSC analyses confirmed that neural activity distinguished social norm from moral foundation violations, and distinguished individualizing and binding moral foundations if Authority is dropped from the binding foundations. Purity violations were persistently associated with amygdala activity, whereas moral foundation violations more broadly tended to engage the default network. Our results constitute partial evidence for neural modularity and motivate further research on the novel groupings identified by the PLSC analyses.

Extant research has shown that previously acquired categorical knowledge affects recognition memory, and that differences in category learning strategies impact classification accuracy. However, it is unknown whether different learning strategies also have downstream effects on subsequent recognition memory. The present study investigates the effect of two unidimensional rule-based category learning strategies – learning (a) with or without explicit instruction, and (b) with or without supervision – on subsequent recognition memory. Our findings suggest that acquiring categorical knowledge increased both hits (Experiments 1 and 2) and false-alarms (Experiment 1) for category-congruent items regardless of the particular strategy employed in initially learning these categories. There were, however, small processing speed advantages during recognition memory for both explicit instruction and supervised practice relative to neither (Experiment 2). We discuss these findings in the context of how prior knowledge influences recognition memory, and in relation to similar findings showing schematic effects on episodic memory.

Previous studies have shown that category learning affects subsequent recognition memory. However, questions remain as to how category learning affects discriminability during recognition. In this three-stage study, we employed sets of simulated flowers with category- and non-category-inclusion features appearing with equal probabilities. In the learning stage, participants were asked to categorize flowers by identifying the category-inclusion feature. Next, in the studying stage, participants memorized a new set of flowers, a third of which belonged to the learned category. Finally, in the testing stage, participants received a recognition test with old and new flowers, some from the learned category, some from a not-learned category, some from both categories, and some from neither category. We applied hierarchical Bayesian signal detection theory models to recognition performance and found that prior category learning affected both discriminability as well as criterion bias. That is, people that learned the category well, exhibited improved discriminability and a shifted bias toward flowers from the learned relative to the not learned category

We provide a new computational model of working memory in the complex span task implemented in the ARCADIA cognitive framework. While there exist implementations of working memory successful enough to account for many of the benchmark findings in the working memory literature, we demonstrate that further progress requires the integration of these models with a rich conception of attention. ARCADIA provides this intersection, allowing for precise control of the focus of attention on a time scale fine enough to begin to disentangle the overlapping effects of interference, temporal decay, and attentional refreshing.

Research on gaze control has long shown that increased visual-cognitive processing demands in scene viewing are associated with longer fixation durations. More recently, though, longer durations have also been linked to mind wandering, a perceptually decoupled state of attention marked by decreased visual-cognitive processing. Toward better understanding the relationship between fixation durations and visual-cognitive processing, we ran simulations using an established random-walk model for saccade timing and programming and assessed which model parameters best predicted modulations in fixation durations associated with mind wandering compared to attentive viewing. Mind wandering-related fixation durations were best described as an increase in the variability of the fixation-generating process, leading to more variable—sometimes very long—durations. In contrast, past research showed that increased processing demands increased the mean duration of the fixation-generating process. The findings thus illustrate that mind wandering and processing demands modulate fixation durations through different mechanisms in scene viewing. This suggests that processing demands cannot be inferred from changes in fixation durations without understanding the underlying mechanism by which these changes were generated.

Objectives

A common finding in the mind-wandering literature is that older adults (OAs) tend to mind-wander less frequently than young adults (YAs). Here, we sought to determine whether this age-related difference in mind-wandering is attributable to age-related differences in motivation.

YAs and OAs completed an attention task during which they responded to thought probes that assessed rates of mind-wandering, and they provided self-reports of task-based motivation before and after completion of the attention task.

Age-related differences in mind-wandering are partially explained by differences in motivation, and motivating YAs via incentive diminishes mind-wandering differences across these groups.

We consider these results in the context of theories on age-related differences in mind wandering, with a specific focus on their relevance to the recently proposed motivational account of such age-related differences.

The recently forwarded family-resemblances framework of mind-wandering argues that mind-wandering is a multidimensional construct consisting of a variety of exemplars. On this view, membership in the mind-wandering family is graded along various dimensions that define more or less prototypical instances of mind-wandering. In recent work, three dimensions that have played a prominent role in defining prototypicality within the mind-wandering family include: (a) task-relatedness (i.e., how related the content of a thought is to an ongoing task), (b) intentionality (i.e., whether thought is deliberately or spontaneously engaged), and (c) thought constraint (i.e., how much attention constrains thought dynamics). One concern, however, is that these dimensions may be redundant with each other. The utility of distinguishing among these different dimensions of mind-wandering rests upon a demonstration that they are dissociable. To shed light on this issue, we indexed the task-relatedness, intentionality, and constraint dimensions of thought during the completion of a laboratory task to evaluate how these dimensions relate to each other. We found that 56% of unconstrained thoughts were “on-task” and that 23% of constrained thoughts were “off-task.” Moreover, we found that rates of off-task thought, but not “freely-moving” (i.e., unconstrained) thought, varied as a function of expected changes in task demands, confirming that task-relatedness and thought constraint are separable dimensions. Participants also reported 21% of intentional off-task thoughts that were freely moving and 9% of unintentional off-task thoughts that were constrained. Finally, off-task thoughts were more likely to be freely-moving than unintentional. Taken together, the results suggest that these three dimensions of mind-wandering are not redundant with one another.

The authors (SSK) make wholesale claims that scientists and lay people suffer not only from misunderstandings, but more dangerous illusions of understanding. However, their case is only supported by intuitive examples rather than solid arguments or data. We argue that SSK misrepresent contemporary scientific thinking, ignore recent advances in methods for causal inference, and exaggerate people’s biases in everyday causal reasoning. Accordingly, we point out that these advances paint a much more optimistic view of scientific understanding.

Reading a book in a coffee shop requires focusing on the task at hand and ignoring task-irrelevant distraction (cognitive stability), while setting aside the book to answer a phone call requires the ability to switch between tasks (cognitive flexibility). Stability and flexibility are often conceptualized as opposing ends of a one-dimensional stability–flexibility continuum, whereby increasing stability (prioritizing task focus) reciprocally reduces flexibility (a readiness to switch tasks), and vice versa. Recent evidence, however, has supported a two-dimensional stability–flexibility relationship, whereby stability and flexibility can be maintained at high levels simultaneously when necessary. Here, we adjudicate between the one- and two-dimensional accounts by fitting competing models to two cued task switching datasets that manipulated the proportion of switch trials (driving contextual adjustments in flexibility) and cross-task congruency effects (driving contextual adjustments in stability). We consider two one-dimensional models: one that assumes a rigid tradeoff where any increase in stability results in a decrease in flexibility, and a more flexible, generalized model that allows but does not enforce such a direct tradeoff. We compare these to two two-dimensional models, one which enforces a strict independence of stability and flexibility, and an unrestricted model that allows interactions between them. Both two-dimensional models, but neither one-dimensional model, were capable of reproducing key behavioral patterns in the original data set. However, the unrestricted two-dimensional model had the best predictive power, indicating that stability and flexibility, while distinct, may trade off in individual- and context-specific ways.

We investigated whether prestimulus alpha-band oscillatory activity and stimulus-elicited recurrent processing interact to facilitate conscious visual perception. Participants tried to perceive a visual stimulus that was perceptually masked through object substitution masking (OSM). We showed that attenuated prestimulus alpha power was associated with greater negative-polarity stimulus-evoked ERP activity that resembled the visual awareness negativity (VAN), previously argued to reflect recurrent processing related to conscious perception. This effect, however, was not associated with better perception. Instead, when prestimulus alpha power was elevated, a preferred prestimulus alpha phase was associated with a greater VAN-like negativity, which was then associated with better cue perception. Cue perception was worse when prestimulus alpha power was elevated but the stimulus occurred at a nonoptimal prestimulus alpha phase and the VAN-like negativity was low. Our findings suggest that prestimulus alpha activity at a specific phase enables temporally selective recurrent processing that facilitates conscious perception in OSM.

Episodic counterfactual thoughts (eCFT) consist of imagining alternative outcomes to past experiences. A common sub-class of eCFT—upward eCFT—involves imagining how past negative experiences could have been better, either because one could have done something differently (internal) or because something about the circumstances could have been different (external). Although previous neuroimaging research has shown that the brain's default mode network (DMN) supports upward eCFT, it is unclear how it is differentially recruited during internal versus external upward eCFT. We collected functional magnetic resonance imaging data while participants remembered negative autobiographical memories, generated either internal or external upward eCFT for the memory, and then rated the plausibility, perceived control and difficulty of eCFT generation. Both internal and external eCFT engaged midline regions of cingulate cortex, a central node of the DMN. Most activity differentiating eCFT, however, occurred outside the DMN. External eCFT engaged cuneus, angular gyrus and precuneus, whereas internal eCFT engaged posterior cingulate and precentral gyrus. Angular gyrus and precuneus were additionally sensitive to perceived plausibility of external eCFT, while postcentral gyrus and insula activity scaled with perceived plausibility of internal eCFT. These results highlight the key brain regions that might be involved in cases of maladaptive mental simulations.

This article is part of the theme issue ‘Thinking about possibilities: mechanisms, ontogeny, functions and phylogeny’.

Hoerl & McCormack (H&M) propose a two-system account of temporal cognition. We suggest that, following other classic proposals where cognitive systems are putatively independent, H&M's two-system hypothesis should, at a minimum, involve (1) a difference in the nature of the representations upon which each system operates, and (2) a difference in the computations they carry out. In this comment we offer two challenges aimed at showing that H&M's proposal does not meet the minimal requirements (1) and (2).

What does it mean to build an artificial agent? In prior work, we have argued at length that the promise of AI introduced into our social milieu calls for a rethinking of the question. We have claimed that agency of the form that all of us naturally recognize requires consciousness to support reasons-responsive choice (ibid). In this short paper, we further claim that consciousness, or something close enough, is often necessary for intentional action as well, and explore the connection between them through thought experiments and computational modeling.

A well-ingrained and recommended engineering practice in safety-critical software systems is to separate safety concerns from other aspects of the system. Along these lines, there have been calls for operating systems (or computing substrates, termed ethical operating systems) that implement ethical controls in an ethical layer separate from, and not amenable to tampering by, developers and modules in higher-level intelligence or cognition layers. There have been no implementations that demonstrate such a marshalling of ethical principles into an ethical layer. To address this, we present three different tracks for implementing such systems, and offer a prototype implementation of the third track. We end by addressing objections to our approach.