To hold someone responsible, we need to assess what causal role their action played in bringing about the outcome. Causality, in turn, can be understood in terms of difference-making -- C caused E if E would have been different had C been different. In previous work, the counterfactual of “C being different” is often construed as “C being absent”. Here, we explore how in social situations, this counterfactual can be alternatively construed by imagining how a different person would have acted in the same situation. We propose a computational model that formalizes this idea of counterfactual replacement. The model compares what actually happened with what would have happened had the person of interest been replaced by someone else, and then predicts responsibility to the extent that the replacement person would likely have changed the outcome. We test our model against people’s responsibility judgments in a variety of scenarios across several experiments.

How do people make causal judgments about other's decisions? Prior work has argued that judging causation requires going beyond what actually happened and simulating what would have happened in a relevant counterfactual situation. Here, we extend the counterfactual simulation model of causal judgments for physical events, to explain judgments about other agents' decisions. In our experiments, an agent chooses what path to take to reach a goal. In Experiment 1, participants either made hypothetical judgments about whether the agent would succeed were it to take a certain path, or counterfactual judgments about whether the agent would have succeeded had it taken a different path. In Experiment 2, participants made causal judgments about whether the agent succeeded or failed because of the path that it took. Our computational model accurately captured participants' judgments in both experiments and we find that causal judgments are better explained by counterfactuals rather than hypotheticals.

How responsible someone is for an outcome depends on what causal role their actions played, and what those actions reveal about their mental states, such as their intentions. In this paper, we develop a computational account of responsibility attribution that integrates these two cognitive processes: causal attribution and mental state inference. Our model makes use of a shared generative planning algorithm assumed to approximate people's intuitive theory of mind about others' behavior. We test our model on a variety of animated social scenarios in two experiments. Experiment 1 features simple cases of helping and hindering. Experiment 2 features more complex interactions that require recursive reasoning, including cases where one agent affects another by merely signaling their intentions without physically acting on the world. Across both experiments, our model accurately captures participants' counterfactual simulations and intention inferences, and establishes that these two factors together explain responsibility judgments.

There is wide agreement that the mind has different mechanisms it can use to make moral judgments. But how does it decide which one to use when? Recent theoretical work has suggested that people select mechanisms of moral judgment in a way that is resource-rational --- that is, by rationally trading off effort against utility. For instance, people may follow general rules in low-stakes situations, but engage more computationally intensive mechanisms such as consequentialist or contractualist reasoning when the stakes are high. Here, we evaluate whether humans and large language models (LLMs) exhibit resource-rational moral reasoning in two moral dilemmas by manipulating the stakes of each scenario. As predicted, we found that the higher the stakes, the more people employed a more effortful mechanism over following a general rule. However, there was mixed evidence for similar resource-rational moral reasoning in the LLMs. Our results provide evidence that people's moral judgments reflect resource-rational cognitive constraints, and they highlight the opportunities for developing AI systems better aligned with human moral values.

Collaboration requires agents to coordinate their behavior onthe fly, sometimes cooperating to solve a single task togetherand other times dividing it up into sub-tasks to work on in par-allel. Underlying the human ability to collaborate is theory-of-mind, the ability to infer the hidden mental states that driveothers to act. Here, we develop Bayesian Delegation, a decen-tralized multi-agent learning mechanism with these abilities.Bayesian Delegation enables agents to rapidly infer the hid-den intentions of others by inverse planning. These inferencesenable agents to flexibly decide in the absence of communi-cation when to cooperate on the same sub-task and when towork on different sub-tasks in parallel. We test this model ina suite of multi-agent Markov decision processes inspired bycooking problems. To succeed, agents must coordinate boththeir high-level plans (e.g., what sub-task they should work on)and their low-level actions (e.g., avoiding collisions). BayesianDelegation bridges these two levels and rapidly aligns agents’beliefs about who should work on what. Finally, we testedBayesian Delegation in a behavioral experiment where partici-pants made sub-task inferences from sparse observations of co-operative behavior. Bayesian Delegation outperformed heuris-tic models and was closely aligned with human judgments.

Some people say they routinely think in an “inner voice,” while others assume this phrase is just a figure of speech. Across two preregistered experiments (N = 190), we investigated whether such subjective differences in internal verbalization predict attested effects of language on perceptual discrimination. Participants completed a canonical discrimination task assessing language-driven categorical perception of facial expressions, plus a self-report measure of internal verbalization and other modes of mental representation. The results replicated the behavioral signatures of categorical perception—faster and more accurate responses for cross-category than within-category discriminations. However, the strength of these effects was unrelated to participants’ propensity for internal verbalization. Our findings suggest that subjective experiences of inner voice use are not a reliable indicator of linguistic involvement in perceptual discrimination. These findings add to an emerging body of research showing that even large differences in mental experiences often lack clear behavioral consequences.

Although subject-complement statements like “girls are as good as boys at math” seem egalitarian, the group in the complement position (boys) is often judged superior. Across two experiments, we examined whether this syntactic framing effect is driven by the ability to discern the pragmatic implications of the syntax. After reading subject-complement statements about the equal math ability of girls and boys or of unstereotyped social groups, participants judged which group was better at math. They also completed a novel measure of pragmatic reasoning ability for subject-complement statements. We found reliable framing effects regardless of stereotype strength, and these effects were uniquely predicted by pragmatic reasoning ability over and above other social-cognitive factors. Moreover, for unstereotyped groups, pragmatic reasoning ability predicted explicit recognition of, and resistance to, the influence of framing. Our findings point to pragmatic reasoning as both a mechanism driving syntactic framing effects and a tool for counteracting them.

Introduction

Medicine prescribing practices are integral to quality of care for leading infectious diseases such as tuberculosis (TB). We describe prescribing practices in South Africa's private health sector, where an estimated third of people with TB symptoms first seek care.

Methods

Sixteen standardised patients (SPs) presented one of three cases during unannounced visits to private general practitioners (GPs) in Durban and Cape Town: TB symptoms, HIV-positive; TB symptoms, a positive molecular test for TB, HIV-negative; and TB symptoms, history of incomplete TB treatment, HIV-positive. Prescribing practices were recorded in standardised exit interviews and analysed based on their potential to contribute to negative outcomes, including increased healthcare expenditures, antibiotic overuse or misuse, and TB diagnostic delay. Factors associated with antibiotic use were assessed using Poisson regression with a robust variance estimator.

Results

Between August 2018 and July 2019, 511 SP visits were completed with 212 GPs. In 88.5% (95% CI 85.2% to 91.1%) of visits, at least one medicine (median 3) was dispensed or prescribed and most (93%) were directly dispensed. Antibiotics, which can contribute to TB diagnostic delay, were the most common medicine (76.5%, 95% CI 71.7% to 80.7% of all visits). A majority (86.1%, 95% CI 82.9% to 88.5%) belonged to the WHO Access group; fluoroquinolones made up 8.8% (95% CI 6.3% to 12.3%). Factors associated with antibiotic use included if the SP was asked to follow-up if symptoms persisted (RR 1.14, 95% CI 1.04 to 1.25) and if the SP presented as HIV-positive (RR 1.11, 95% CI 1.01 to 1.23). An injection was offered in 31.9% (95% CI 27.0% to 37.2%) of visits; 92% were unexplained. Most (61.8%, 95% CI 60.2% to 63.3%) medicines were not listed on the South African Primary Healthcare Essential Medicines List.

Conclusion

Prescribing practices among private GPs for persons presenting with TB-like symptoms in South Africa raise concern about inappropriate antimicrobial use, private healthcare costs and TB diagnostic delay.

Humans are remarkably adept at inferring the causes of events in their environment; doing so often requires incorporating information from multiple sensory modalities. For instance, if a car slows down in front of us, inferences about why they did so are rapidly revised if we also hear sirens in the distance. Here, we investigate the ability to reconstruct others' actions and events from the past by integrating multimodal information. Participants were asked to infer which of two agents performed an action in a household setting given either visual evidence, auditory evidence, or both. We develop a computational model that makes inferences by generating multimodal simulations, and also evaluate our task on a large language model (GPT-4) and a large multimodal model (GPT-4V). We find that humans are relatively accurate overall and perform best when given multimodal evidence. GPT-4 and GPT-4V performance comes close overall, but is very weakly correlated with participants across individual trials. Meanwhile, the simulation model captures the pattern of human responses well. Multimodal event reconstruction represents a challenge for current AI systems, and frameworks that draw on the cognitive processes underlying people's ability to reconstruct events offer a promising avenue forward.