Exposure to feelings of another can shape how we feel and think in that context. We propose that exposure to the feelings of multiple others carries broader implications that extend beyond the exposure context. We argue that people identify patterns of affective reactions in the social group, which leads them to infer affect norms. These norms, in turn, serve as benchmarks for subsequent social judgments. In four studies (N = 418), three of which were preregistered, we exposed participants to sequential affective reactions of multiple others and manipulated their intensity. We then tested the effects of such exposure on predictions of others' feelings, one's own feelings, and subsequent social judgments. Across studies, we show that people predict feelings of group members based on the norm they were exposed to, that they assimilate their own reactions to the norm, and that they subsequently judge social targets that deviate from the norm more negatively. Our findings demonstrate how exposure to feelings of multiple others enables people to learn affect norms, bridging research on emotions in dyads, groups, and cultures. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

Greater numbers of people are turning to artificial intelligence (AI) for empathy and emotional support. Here, we review and synthesize recent empirical work on how people perceive empathy from AI versus from humans. Growing evidence points to two dueling effects and a paradox: AI produces language that is rated higher in empathy than language written by humans, but when people perceive text as coming from AI versus a human, they rate it as less empathic. However, despite sometimes rating AI more empathic or having to wait for human empathy, people still show a preference for human empathy. This emerging literature carries significant implications for fundamental research on empathy and for public discourse as the use of AI for emotional support continues to grow.

The fusion of agentic AI and LLMs marks a new frontier in information warfare.

In the past several years we have seen much interest in how people seek out empathy from artificial intelligence (AI) chatbots, especially Large Language Models like ChatGPT. In this short paper we describe a meta-analysis of several recently-published papers and preprints on this topic. This brief meta-analysis supports a more in-depth narrative review of the literature (Ong et al., 2026). Here, we sought to answer two research questions: Research Question 1: How do people perceive AI-generated empathic responses compared to human-written empathic responses? Research Question 2: How do people perceive empathic responses which are labeled as generated by AI compared to written by humans?

Empirical work shows that partisans are disproportionately attracted to extreme allies on their own political side compared to moderates, a phenomenon called political acrophily. We argue that this preference for extremes also shapes the beliefs they adopt: many citizens who describe themselves as moderate partisans increasingly endorse the most extreme positions on core policy issues. We call this pattern belief acrophily: the tendency to hold predominantly extreme positions on political issues in relation to one’s own political identity. Using American National Election Studies data from 1972–2024, we show that belief acrophily has risen sharply over time: by 2020 and 2024, moderate Democrats and Republicans were more likely to place themselves at the extreme ends of issue scales on social benefits, health care, and race than at the midpoint. Panel data reveal that moderate partisans with extreme views are substantially more likely to adopt stronger partisan identities in the future. This analysis suggests that polarization is deepening not only through increased issue position sorting according to party membership, but also through intensifying extremity of moderates’ positions on concrete issues.

Implicit biases are stereotypes and attitudes that influence decisions and actions, contributing to discrimination and societal inequities. The Implicit Association Test (IAT) is the most widely used tool for measuring implicit bias, assessing response time in sorting stimuli into labeled categories. Most interpretations assume that IAT performance (D-scores) reflects conflicting associative memories or decision ease. We challenged this assumption by decomposing D-scores into additional cognitive processes that may influence results, particularly response caution—the tendency to trade speed for accuracy. Using Racing Diffusion Models across 39 topics (N = 115,601), we found that response caution explained significantly more variance in D-scores beyond decision ease. Response caution also best predicted explicitly reported biases. These findings challenge the traditional interpretation of D-scores as primarily reflecting associative memory activation and highlight the need to consider multiple cognitive processes when assessing implicit biases.

Do LLMs have emotions? A recent paper from Anthropic reports finding internal representations of emotion concepts in Claude Sonnet 4.5, concluding that the LLM has 'functional emotions.' We evaluate this claim against what is known about how emotions actually function in biological systems. We argue that emotions serve two core functions: the context-sensitive interpretation of situations, and the reorganization of processing across multiple systems in response to those interpretations. The Anthropic findings offer partial support for the first function, though the consistent, discrete emotional representations identified in Claude sit uneasily with affective neuroscience findings that human emotion is characterized by variable rather than uniform neural signatures. On the second function, the evidence is mixed: Claude's representations modulate output without producing the dynamic reorganization of attention, decision speed, and motivational state that defines emotion in biological systems. We close by proposing what it would take for an LLM to have emotions.

Anticipating the frequency and size of protests remains one of the most formidable challenges in understanding collective human behavior. Psychological research has often addressed this challenge by examining how discrete emotional experiences motivate individuals to participate in collective action. Yet a society can feel highly emotional on average without people feeling the same thing. The average intensity of emotions tells us how strongly people feel on average, but not whether people are emotionally aligned. We argue that beyond the average emotional intensity within groups, a crucial predictor of protest size is collective emotional entropy, reflecting how aligned (low entropy) versus fragmented (high entropy) individuals within a society are across a range of societally relevant emotions (e.g., hate, empathy and guilt). We conducted an eight-month rolling cross-sectional survey of 11,713 Jewish Israelis in a period of large protest activity (2023-2025). Using daily distributions of eight discrete emotions, we estimated both average emotional intensity and collective emotional entropy and linked it to records of protest frequency and size. Analyses revealed that beyond the effect of individual emotions, lower entropy in the preceding days predicted both a greater number of daily protests and larger protest sizes, even when controlling for conflict intensity and other social–psychological indicators. Together, this suggests that collective action may depend not only on the emotions people feel, but on whether people come to feel in similar ways.

Collective emotional responses represent responses of multiple people to an emotional situation. Such responses are ubiquitous in many contexts, such as conflicts, crises, and celebrations. Yet, a closer look at processes of collective emotions leads to an observation: collective emotion often co-occurs with attempts to influence the emotions of the group. We refer to these attempts as collective emotion regulation, where members of groups aim to change collective emotions. In this paper, we first define collective emotion and propose a model of collective emotion generation. We then define collective emotion regulation and distinguish it from related concepts. We propose a process model to describe the unfolding of collective emotion regulation that help organize a variety of behaviors previously discussed in other literatures. We show how the concept can help explain collective behavior and facilitate planning interventions designed to regulate collective emotions. Finally, we discuss open empirical questions and future directions.

Public discourse and emerging policy typically assume that AI emotional support is a deliberate act: a lonely user consciously seeking comfort from a dedicated companion chatbot. In this paper, we draw on emerging empirical evidence and argue that this picture is inaccurate on two accounts, both in how AI emotional support arises and how it shapes future behavior. First, AI emotional support commonly emerges incidentally within task-oriented interactions on general-purpose platforms, much as workplace friendships deepen through collaboration. Second, these incidental encounters are path-dependent: positive experiences of AI emotional support update people's beliefs about AI's emotional capabilities and redirect their choices for future emotional support, increasing preference for AI and decreasing preference for humans. We review recent evidence, including a large-scale longitudinal study conducted in collaboration with OpenAI, showing that daily five-minute conversations with an AI about personal issues over 28 days led to a 10.3

People increasingly face a novel decision when seeking emotional support: human or AI. In existing studies, AI's empathic messages are rated as well as or better than humans'. But these studies either assigned the support source or honored people's choice. In real life, support is often incongruent with choice, as people want one source and receive the other. Across three experiments (N = 1,951), participants chose whether to share an emotional experience with a human or an AI, then were randomly assigned to a congruent or incongruent partner. AI support was rated as superior only among those who had chosen it. Yet regardless of congruence, interacting with AI increased willingness to choose it again. In a 28-day study with OpenAI (N = 981), daily conversations shifted preferences toward AI and away from humans, but only when conversations turned personal. Emotional support choices are thus path-dependent, progressively redirecting away from human connection.

People increasingly turn to LLMs for emotional support and understanding how such AI support differs from humans is crucial for evaluating the consequences of this use. We focused on a highly researched emotion regulation strategy, cognitive reappraisal, a strategy for alleviating negative emotions through reinterpretation. We compared trained humans and GPT-4 in their reappraisal ability with vignettes (Study 1, N = 868) and in real-time interactions (Study 2, N = 386). GPT-4 was consistently rated as more effective, empathic, and novel by objective raters and recipients of the reappraisals. Incentivizing humans increased time spent on the reappraisals but did not close the gap (Study 3, N = 1477). Labeling reappraisals as AI reduced participants' evaluation of their effectiveness, though GPT-4 was still considered more effective (Study 4, N = 496). Language analysis revealed differences in vocabulary associated with quality, suggesting that specific language use is what drives AI superiority.

There is an urgent need for scalable mental health interventions, particularly for workers in under-resourced settings who experience high stress and limited access to mental health resources. One promising approach is empowering workers’ emotion regulation skills. This study evaluates a brief, cost-effective, and scalable intervention that teaches reappraisal—an emotion regulation strategy that alters how individuals interpret challenging situations. Despite its potential, the long-term effectiveness of reappraisal in workplace contexts remains uncertain. We tested this intervention among early education workers in a non-profit organization serving low-income families—one of the most underpaid and stressful professions in the U.S. Study 1 (n = 1,967), a pre-registered field survey, found positive correlations between reappraisal use, emotional well-being, and job performance. Study 2, a pre-registered longitudinal field experiment (n = 4,054), found lasting improvements in some emotional well-being and workplace outcomes from the reappraisal intervention (vs. active control) six months later, including reduced stress, a greater sense of control over stressors, more frequent action-taking to solve problems, increased self-reported job performance, and stronger intentions to stay in the current organization. Furthermore, reappraisal engagement during the intervention significantly mediated these effects. These studies advance understanding of emotion regulation’s effectiveness, durability, and mechanisms in real-world settings, providing a scalable solution to improve well-being in high-stress professions. This research also provides a template for tailoring emotion regulation interventions to novel contexts and extending their benefits to diverse populations.

Emotions change through internal momentum and reactions to events. However, the literatures examining these forces have proceeded separately, emphasizing either endogenous regulation or exogenous value signals such as reward prediction errors. This separation obscures a convergent account in which both influences jointly shape momentary affect. We address this gap by applying sparse equation discovery (SINDy) to a large happiness dataset from a gambling task (N = 16,337) and introduce DynAffect-C, a unified model that separates and quantifies endogenous and exogenous inertia. DynAffect-C recovers known value-signaling effects and, with regulatory baseline-attraction terms, explains substantially more moment-to-moment variation than value-only models. The model reveals nonlinear pull toward baseline affect, alternating mood states in a subset of participants, and a trade-off between stronger internal momentum and transient impacts of external events. The result integrates two literatures and yields a unified theory and testable equation for affect.

Computational models of reinforcement learning (RL), have significantly contributed to our understanding of human behavior and decision-making. Traditional RL models, however, often adopt a linear approach to updating reward expectations, potentially oversimplifying the nuanced relationship between human behavior and rewards. To address these challenges and explore new models of reinforcement learning, we utilized a novel method of model discovery using equation discovery algorithms. This method, currently used mainly in physics and biology, attempts to capture data by proposing a differential equation from an array of suggested linear and nonlinear functions. Using this novel method, we were able to identify a new model of RL which we termed, the Quadratic Q-Weighted model. The model suggests that reward prediction errors obey nonlinear dynamics and exhibit negativity biases, resulting in an underweighting of reward when expectations are low, and an overweighting of the absence of reward when expectations are high. We tested the generalizability of our model by comparing it to classical models used in 9 published studies. Our model surpassed traditional models in predictive accuracy across eight out of these nine published datasets, demonstrating not only its generalizability but also its potential to offer new insights into the complexities of human learning. This work showcases the integration of a novel behavioral task with advanced computational methodologies as a potent strategy for uncovering the intricate patterns of human cognition, marking a significant step forward in the development of computational models that are both interpretable and broadly applicable.

Users on social media are regularly presented with sequences of emotional content in their newsfeeds, which affects their viewpoints and emotions. Could the way users aggregate and remember emotional content from their feeds contribute to the fact emotions are amplified on social platforms? Across five studies (N = 1,051), using experimentally manipulated social media feeds, we found that participants consistently overestimated the average emotional intensity of the individual responses expressed by other users in a sequence (Study 1a). This overestimation led to stronger emotional reactions to the news content that these responses were reacting to (Study 1b). Investigating the mechanism suggested that while there was stronger memory for more emotional responses within a response sequence, we could not find a direct link between memory and overestimation (Study 2). We showed that overestimation was driven mainly by the salience of emotional intensity of different items in the sequence, by replicating the effect using sequences of emotional words (Study 3). We then turned to the consequences of overestimation, showing that overestimation of emotional sequences was uniquely associated with perceiving more intense emotional responses as more representative of how other people would react (Study 4), and with overestimation of the emotionality of the newsfeed as a whole (Study 5). Overestimation of the average individual emotional intensity ratings of a sequence was also predictive of willingness to share articles. This set of findings sheds light on how sampling from newsfeeds amplifies the perception of emotionality.

Increasingly, people use language models for emotional support, and understanding the quality of that support is crucial. This project (N = 3,740) focused on one of the pillars of emotional support and the most researched emotion regulation strategies, called cognitive reappraisal, which is the ability to extract multiple meanings from an emotional situation and to choose a framing that reduces its emotional intensity. In a first conservative test, we compared trained humans and GPT-4's ability for reappraisal using made-up vignettes and third-party human raters (Study 1), showing that GPT-4 outperformed humans. In Study 2 we then investigated whether the gap was driven by effort or skill by incentivizing participants to produce better reappraisals, which led to increased time spent on reappraisals but did not decrease the gap between humans and GPT-4. To examine how the perception that support came from AI affected evaluations, in Study 3 we provided participants with an AI label of the source of reappraisal, which reduced their evaluation of the effectiveness of reappraisal, but GPT-4 was still considered more effective. In Study 4, to ensure that differences were sustained in real-time interactions, we had participants share negative emotional situations with either a human or GPT-4 and receive reappraisal, replicating GPT-4's superior performance compared to humans. We conducted language analysis to identify differences between humans and GPT’s reappraisals, finding differences in language complexity, which explain some differences in evaluation. These results help us understand the nature of emotional support by LLMs and how it compares to humans.

Emotion-regulation difficulties are implicated as a risk factor for suicidal thoughts, yet little is known about how adults with suicidal thoughts regulate emotions in daily life or which deficits are specific to suicidality versus shared across psychopathology. In two ecological-momentary-assessment studies (Study 1: N = 396; Study 2, recruited online: N = 195), we compared adults with current suicidal thoughts with adults with past or no suicidal history (Study 1) and with psychiatric and healthy control participants (Study 2). Participants with current (vs. past) suicidal thoughts reported greater substance use and self-injury to regulate emotions (Study 1). Compared with psychiatric control participants, participants with suicidal thoughts reported higher regulatory effort and substance use, and compared with healthy control participants, they additionally reported greater distraction and rumination and lower regulatory success (Study 2). Self-injury and substance use uniquely predicted momentary suicidal thinking (Study 2). Findings highlight substance use, self-injury, and heightened regulatory effort as potentially distinct emotion-regulation processes associated with suicidal thoughts.

Artificial Intelligence models can generate emotional support messages that peopleperceive to be highly empathic; however, people also perceive less empathy if they believe that messages were AI-generated. This new focus on how the empathy recipient perceives human- written versus AI-generated empathic responses has recently gained attention. We review and synthesize recent empirical work using meta-analyses, clarify claims and limitations, and highlight future directions. This emerging literature carries significant implications for fundamental research on empathy, for public discourse as the use of AI for emotional support rapidly grows, and for policymakers considering regulation and ethical guidance.