Stefano Palminteri

As large language models (LLMs) are increasingly integrated into decision-making systems (e.g., autonomous vehicles and medical devices), understanding how humans perceive and evaluate AI-generated judgments is crucial. To investigate this, we conducted a series of experiments in which participants evaluated justifications for moral and non-moral choices, generated either by humans or LLMs. Participants attempted to identify the source of each justification (either human or LLM) and indicated their agreement with its content. We found that while detection accuracy was consistently above chance, it remained below 75%. In terms of agreement, there was no overall preference for human-generated responses, even though machine-generated justifications were favored in particularly challenging moral scenarios. Notably, we observed a systematic anti-AI bias: participants were less likely to agree with judgments they believed were AI-generated, regardless of the true source. Linguistic cues, such as response length, typos, first-person pronouns, and cost-benefit language markers (e.g., “lives,” “save”), influenced both detection and agreement. Participants tended to disagree with cost-benefit calculations, possibly due to an expectation that AI would favor such reasoning. These findings highlight the influence of motivated belief and ingroup/outgroup bias in shaping human evaluation of AI-generated content, particularly in morally sensitive contexts.

Contextual valuation is a well-documented phenomenon in reinforcement learning, typically manifesting as range normalization in outcome representation. However, recent findings have revealed systematic deviations from this model, particularly when three options with equally spaced values are presented. In this study, we hypothesize that these distortions in outcome normalization arise from attentional processes. To test this, we conduct three experiments with 105 participants in total while simultaneously tracking their gaze position with eye-tracking. Furthermore, we systematically manipulate attention using both top-down and bottom-up approaches. These manipulations significantly increase the subjective valuation of attended options, supporting a causal role of attention in shaping value representation. To account for these effects, we develop a reinforcement learning model that integrates attentional mechanisms, wherein gaze duration directly modulates the absolute value of options prior to range normalization. This attentional range model outperforms attention-free and choice-repetition alternatives, underscoring the critical influence of attention in value computation.

Social media have fundamentally transformed how we live and communicate. However, the methods to study how our cognitive systems interact with technology platforms are very limited. Computational modelling represents a new avenue to uncover the finegrained cognitive processes driving social media behaviour. Here, we develop a computational model of real-world social media posting data, adapted from the animal reward learning literature. Using a Twitter (currently X) dataset ( n = 2696 users), including a preregistered replication, we show that a hybrid reinforcement learning and habitual cognitive process underlies social media posting behaviour. More frequent posters show more signs of habitual behaviour. Further, younger people and women are more driven by reinforcement learning – updating their strategy more adaptively to maximise social media rewards – while older users and men are more habitual.

Rewards and punishments in reinforcement learning are encoded in both absolute and relative manners. Reference-point dependence, a valuation bias shared by adaptation-level and prospect theories, is often proposed as the computational mechanism underlying relative value encoding. However, the extent to which these behavioural and computational mechanisms are conserved across species remains to be fully understood. We therefore designed parallel reinforcement learning tasks in humans and rats to examine reference-point dependence across species. Behavioural analyses showed robust relative value encoding in both species, and computational modelling confirmed that reference-point dependence reliably accounts for behaviour in humans and rats. Despite these major similarities between species, some differences in behavioural and modelling parameters were observed. Overall, our study demonstrates that relative value encoding is a robust feature of reinforcement learning that is conserved across humans and rats.

It has been well established that, in the context of value-based learning and decision making, the way in which options are paired together (the learning architecture) significantly affects their subjective values as inferred from observed choices. What remains unclear is whether similar distortions emerge when subjective values are assessed in a declarative manner. Across seven experiments, we show that participants remember a systematically biased version of option values that is affected by the context-specific ranking of the option during the learning phase. These results are robust across different value elicitation methods (choice-based and declarative) and decision modalities (experience-based or description-based). Furthermore, we show that presenting the options in different configurations during the learning phase affects their subsequent subjective values. By doing so, we illustrate how principles of context dependence may be employed to design learning environments that enhance value maximisation in decision-making.

Decisions under risk can be characterized by two complementary but distinct behavioral variables: the propensity to choose the expected value-maximizing option and the tendency to select the risky option. Human reinforcement learning (RL) research has traditionally focused more on the former (maximization), providing many insights into how fundamental features of the learning environment influence behavior. For instance, previous studies have shown that counterfactual feedback about unchosen options exerts a strong influence on maximization. However, comparatively little is known about scenarios in which maximization and risk propensity are assessed orthogonally and simultaneously. To address this gap, we designed a series of RL experiments in which we manipulated key factors, including the feedback information regime (partial vs. complete) and the relative expected values of risky and safe options. Contrary to the conventional assumption that “more information is better,” our results clearly show that, in such contexts, the effect of foregone outcomes on risk preference outweighs its influence on maximization. We further compared RL choices with description-based versions of the same decision problems. While replicating established findings on the description–experience gap, we found that complete information had little impact on its expression, suggesting that this gap does not arise primarily from instant feedback or insufficient outcome sampling.

Large Language Models (LLMs) have rapidly become a central topic in AI and cognitive science, due to their unprecedented performance in a vast array of tasks. Indeed, some even see "sparks of artificial general intelligence" in their apparently boundless faculty for conversation and reasoning. Their sophisticated emergent faculties, which were not initially anticipated by their designers, have ignited an urgent debate about whether and under which circumstances we should attribute consciousness to artificial entities in general and LLMs in particular. The current consensus, rooted in computational functionalism, proposes that consciousness should be ascribed based on a principle of computational equivalence. The objective of this opinion piece is to criticize this current approach and argue in favor of an alternative "behavioral inference principle", whereby consciousness is attributed if it is useful to explain (and predict) a given set of behavioral observations. We believe that a behavioral inference principle will provide an epistemologically valid and operationalizable criterion to assess machine consciousness.