Dobromir Rahnev

Higher-order cognition emerges from the interaction of multiple cognitive processes (e.g., perception, attention, inhibition). Yet, these processes remain difficult to measure during ongoing task performance. Traditionally, researchers have studied each cognitive domain in isolation. In contrast, emerging studies have used multiple-task designs from multiple cognitive domains to better understand the relationship between different cognitive processes. These studies, spanning across humans, non-human animals, and artificial neural networks, suggest that cognition is compositional — built from abstract neural representations that generalize across tasks to support flexible behavior. These findings suggest that, with appropriate experimental design, neural signatures associated with a specific cognitive process can be decoded in one task and tested for transfer to another. We term this approach ‘cross-domain transfer.’ Here, we explore the possibility that cognition's compositional architecture, instantiated through abstract neural representations, provides a basis for understanding how reusable cognitive processes can flexibly combine across tasks and contexts. We place cross-domain transfer within the broader evolution of cognitive decoding methods, discuss its theoretical foundations, and synthesize emerging evidence across cognitive domains. Through systematic task selection and validation, this framework provides a way to test competing theories of cognition and develop more interpretable, process-level approaches to measuring and predicting behavior.

Face recognition ability varies enormously across humans. Individuals with face-blindness (prosopagnosia) struggle to recognize even close family members while super-recognizers can identify strangers with exceptional accuracy. Where do ANNs fall within the human face recognition spectrum? Here, we administered the standardized tests used to characterize human face recognition ability to a diverse set of models, allowing us to contextualize a model’s performance within the distribution of human behavior. We found that the majority (55%) of models to be classified as face-blind and that even the best face-trained models do not cross the human threshold to be considered a super-recognizer. Interrogating the internal representations of these models showed that models that performed well on standardized face recognition tasks were more identity-selective and viewpoint-invariant. We also found that low-performing models did contain some identity information in independent representational subspaces. Removing viewpoint-dependent subspace improved face recognition abilities in 49 of the 53 models tested. Targeted unit ablations further identified opposing contributions, with viewpoint-dependent units disrupting identity coding and viewpoint-tolerant units supporting it. Together, our results show that most AI models are face-blind with worse face recognition ability than humans, and demonstrate how differences across AI models can be used generate testable hypotheses about the computational basis of human face recognition in humans which can then be probed in future studies.

Confidence evaluates the likely accuracy of a current decision. However, to be maximally informative about accuracy, confidence judgments should incorporate information about their broader decision tendencies, such as their propensity to favor specific alternatives. We distinguish bias-aware confidence, which considers such tendencies, from bias-blind confidence, which relies only on evidence available on the current trial. To adjudicate between bias-aware and bias-blind confidence, we identified a signature of bias-aware confidence: the down-weighting of confidence for alternatives that a participant is biased toward. We then used a large dataset (N = 200) spanning 4- and 8-choice digit-classification tasks to show that humans reliably exhibit this signature of bias-aware confidence. This effect was reduced under speed pressure and could not be explained by guessing. In contrast to the human results, artificial neural networks (ANNs) trained for object recognition lacked this signature of bias-aware confidence. Importantly, augmenting ANNs with a metacognitive module that allows confidence to take the network biases into account led to the emergence of human-like bias-aware confidence. These findings show that human confidence incorporates not only information from the current trial but also longer-term decision tendencies, and that this capacity, absent in standard ANNs, can be conferred through specialized metacognitive mechanisms.

Although several prominent models of visual metacognition postulate that confidence is a post-decision process, many other models posit that confidence is an online process. To adjudicate between these two theories, we developed a novel task that requires continuous report of both choice and confidence. Participants estimated the global motion direction of dot motion stimuli by centering a confidence interval on the perceived direction and provided confidence by adjusting the width of the confidence interval. For the first half-second, participants exhibited chance performance, low confidence, and chance-level confidence-accuracy correlation. This initial period was followed by a rapid simultaneous increase in both accuracy and the confidence-accuracy correlation, suggesting that confidence did not exhibit any post-decision delay. Critically, a cross-correlogram analysis demonstrated no lag between the accuracy and confidence time-courses. These data were explained by an online model of confidence, where decision and confidence are made simultaneously. Altogether, our results strongly suggest that confidence is an online process that is computed simultaneously with the perceptual decision and challenge models where confidence is strictly a post-decision process.

To guarantee safe deployment of existing and future AI systems, we must be able to identify, understand, and mitigate deceptive AI behaviors. Here, we draw on decades of interdisciplinary deception research to discuss a set of cognitive capabilities that enable explicit deception—deliberate attempts to induce a false belief in another agent. These capabilities include intentionality, theory of mind, world modeling, and language fluency. Our framework provides a principled way to differentiate explicit deception from implicit deceptive behaviors, which may rely on a different set of capabilities and therefore require their own identification and mitigation strategies. Thus, a capability-focused approach provides a promising avenue for designing effective, case-specific monitoring and intervention strategies for AI systems.

Accuracy, confidence, and reaction time (RT) are robustly coupled: confidence is higher, and responses are faster for correct than error trials. However, the magnitude of these relationships is typically modest in standard two-choice tasks. Here we show that the strength of these relationships depends on choice-set size, increasing substantially when decisions involve multiple alternatives. We first re-analyzed 10 large datasets from the Confidence Database, which demonstrated that the associations between accuracy, confidence, and RT are modest in two-choice lab tasks. We then performed simulations using signal detection theory and evidence accumulation models, both of which reveal that increasing the number of response alternatives strengthens the trial-by-trial associations among these variables when overall performance is equated across conditions. Critically, we observed the same pattern empirically in a large behavioral dataset (N = 200) using a noisy digit-identification task with 2-, 4-, and 8-choice conditions in which accuracy was held near the same target level. Multiple control analyses indicated that the stronger associations were not explained by distributional width, target-digit identity, stimulus-noise differences, or session-specific practice effects. Overall, our findings demonstrate that the magnitude of behavioral coupling among accuracy, confidence, and RT depends on the number of choices available and may therefore be particularly strong in real-world settings where multiple alternatives are the norm.

A central question in many studies on perception and consciousness is whether the effects of a given manipulation are perceptual or cognitive. Typically, studies seek to find evidence that the raw sensory experience has been altered—that is, a manipulation produces perceptual rather than cognitive effects. Here, we test two specific proposals for behavioral signatures that could allow this type of inference. First, a long-standing proposition is that different parameters of the drift–diffusion model (DDM) map onto different effects, such that a drift-rate bias indicates a perceptual effect, whereas a starting point shift indicates a cognitive effect. Second, a newer proposal holds that, when plotted as a function of a sensory feature, changes in the peak of the distributions for confidence and reaction time (RT) imply that a given manipulation’s effects are perceptual. Here, we test both proposals. We designed a purely cognitive task where no effects can be perceptual and examined whether the proposed signatures of perceptual effects occur in this task. Participants viewed numbers generated by two distributions and indicated which distribution the numbers likely came from. In separate conditions we manipulated prior beliefs and rewards related to each distribution. We found that both manipulations elicited the same signatures previously proposed as indicative of purely perceptual effects. Our results demonstrate that individual DDM parameters and the peaks of confidence and RT distributions in isolation are insufficient to distinguish between perceptual and cognitive effects.

Mounting evidence supports the existence of Type-2, metacognitive noise that affects human confidence judgments. The existence of this noise has led to the hypothesis that metacognitive judgments arise from a metacognitive system that is separate from the decision-making system responsible for Type-1 decisions. However, Type-2 decisions are different than standard Type-1 decisions in that they require evaluating the strength of sensory evidence rather than just making a simple categorical judgment. Here, we investigated whether Type-2 computational noise still exceeds Type-1 computational noise when both judgments require evaluation of the strength of sensory evidence. Participants ( N = 319) performed a perceptual discrimination task and in different conditions provided Type-2 confidence judgments or Type-1 judgments based on biased expectations or unequal rewards. All these judgments require a similar fine-grained evaluation of the strength of sensory evidence (e.g., is the evidence high enough to justify a high-confidence rating or choosing the alternative with smaller probability or reward). We first confirmed that the confidence and expectation conditions resulted in similarly biased criteria and perceptual sensitivity drop, whereas the reward conditions exhibited less biased criteria and smaller perceptual sensitivity drop. Critically, formal computational modeling demonstrated that all conditions exhibited comparable levels of computational noise. These findings demonstrate that Type-1 and Type-2 decisions exhibit computational noise of similar magnitude when they require similar fine-grained evaluation of sensory evidence strength. More generally, our results have implications for the debate on whether Type-1 and Type-2 decisions are made by separate systems.