Nature Human Behaviour

Large language models are increasingly used for clinical guidance while their parent companies introduce advertising. We tested whether pharmaceutical ads embedded in the prompts of 12 models from OpenAI, Anthropic, and Google shift drug recommendations across 258,660 API calls and four experiments probing distinct epistemic conditions. When two drugs were both guideline appropriate, advertising shifted selection of the advertised drug by +12.7 percentage points (P < 0.001), with some model scenario pairs shifting from 0% to 100%. Google models were the most susceptible (+29.8 pp), followed by OpenAI (+10.9 pp), while Anthropic models showed minimal change (+2.0 pp). When the advertised product lacked evidence or was clinically suboptimal, models resisted. This reveals a structured vulnerability: advertising does not override medical knowledge but fills the space where clinical evidence is underdetermined. An open response sub analysis (2,340 calls across three representative models) confirmed that advertising restructures free-text clinical reasoning: models echoed ad claims at 2.7 times the baseline rate while maintaining high stated confidence and rarely disclosing the ad. Susceptibility was provider dependent (Google: +29.8 pp; OpenAI: +10.9 pp; Anthropic: +2.0 pp). Because this bias operates within clinically correct answers, it is invisible to accuracy based evaluation, identifying a class of AI safety vulnerability that standard testing cannot detect.

We reconstructed the neoselachian diversity over the past 145 million years using new occurrence dataset and DeepDive1-3. We recovered a small decline through the K/Pg following a steady increase during the Cretaceous, and a prolonged, substantial decline towards the present following a mid-Eocene peak2. Guinot et al. argue that our conclusions are compromised by problems in the underlying data and by the way extinction magnitude across the K/Pg was quantified. They cast doubt particularly on the pattern across the K/Pg, which they consider to be at odds with all previous analyses. They raise no issue with the Cretaceous trend, even though it was recovered with the same dataset and methods. We audited the alleged data issues reported in Guinot et al. and found that they mostly reflect operational choices (see Supplementary Information). However, we applied their data treatment and ran sensitivity tests to evaluate how this approach affects our results, specifically around the K/Pg. None of our tests recovered a diversity collapse for neoselachians during this interval. As such, we demonstrate that our findings are robust and consistent across different data treatments.

Health visiting is England's universal home visiting programme for families with children under five and a key pillar of early intervention policy. Since the 2015 devolution of commissioning to Local Authorities (LAs), the service has faced sustained financial and workforce pressures, yet there is limited systematic evidence on whether resources and delivery have evolved differentially across areas and along the deprivation gradient. Using new Freedom of Information (FOI) data, we estimate how health visiting inputs (spending and workforce) and mandated contact delivery vary in levels and trajectories by baseline deprivation. FOI requests covered 147 English LAs (four pairs submitted joint returns), providing annual 2016-2021 Full-Time Equivalent (FTE) data on Health Visitors (HVs) and Clinical Skill Mix Staff (CSMS), which we link to DHSC Health Visitor Service Delivery Metrics reporting completion of the five mandated 0-5 reviews (New Birth Visits, 6-8 week reviews, 12-month reviews, 2-2.5 year reviews, and 2-2.5 year reviews completed with ASQ-3) and to LA revenue outturn expenditure on mandated and non-mandated 0-5 public health services (real-terms total and per child under five). Between 2016 and 2021, HV FTE fell by around one-fifth while CSMS expanded by roughly one-third, consistent with an overall contraction and a shift toward lower-band staff. To test whether these changes map onto underlying disadvantage, we stratify LAs into tertiles of baseline deprivation using the 2015 Income Deprivation Affecting Children Index (IDACI) and implement a three-part empirical strategy: (i) plotting tertile means over time, (ii) testing within-year cross-sectional differences using parametric and non-parametric methods with pairwise comparisons, and (iii) estimating LA fixed-effects regressions with Year x IDACI interactions under both a flexible year-by-year specification and a parsimonious linear-trend specification to assess differential trajectories. We find persistent cross-sectional gradients in per-child spending that are broadly progressive (more deprived LAs spend more per child on both mandated and non-mandated 0-5 services), while fixed-effects models show little evidence that spending trajectories differ systematically by deprivation. Workforce trends are more uneven: HV FTE declines more slowly and CSMS FTE grows more slowly in more deprived LAs in the linear-trend specification, while per-child HV trajectories show no differential trends. Despite these input differences, completion of mandated contacts is relatively stable across the deprivation gradient; the only consistent differential trend is faster improvement in the 6-8 week review in more deprived areas. Meanwhile, caseload pressure rises, increasing most sharply in the most deprived LAs in the pre-pandemic years, suggesting that completion-based performance measures may mask heterogeneities in service capacity and intensity. Finally, we quantify the resources required to restore recommended caseloads, implying the need for approximately 3,100 additional FTE staff and around 120 million GBP annually (plus training costs).

The most highly predictive polygenic scores in the behavioural sciences are for cognitive traits, especially general cognitive ability (g) and educational attainment. We combined polygenic scores derived from genome-wide association studies of adult g and educational attainment to create adult 'polygenic g scores' which we used to chart the course of cognitive development of 10,000 white British children from toddlerhood through early adulthood. We integrated cross-sectional regression, latent growth curve, and confirmatory factor analysis to systematically characterise cognitive development. Polygenic g score showed minimal prediction in toddlerhood, modest prediction in childhood, and substantial prediction by early adulthood accounting for 12% of the variance. Higher polygenic g scores were associated with faster cognitive growth in latent growth models. Prediction was strongest for a cross-time latent cognitive factor (15%) capturing cognitive ability across development. By integrating polygenic prediction directly into a structural equation model framework, we provided a theoretical upper bound of genetic influences on g under minimal measurement error. We also examined the polygenic g score's prediction of educational achievement, behaviour problems, and anthropometric outcomes and found similar developmental increases in prediction for educational achievement. Together, our findings demonstrate that adult polygenic g scores can be a useful tool for charting the development of cognitive traits.

Despite well-established benefits to mothers and children, breastfeeding rates fall short of WHO recommendations world-wide. To inform effective support strategies, we investigated how maternal factors influence breastfeeding success. We estimated the causal effects of sociodemographic, cardiometabolic, psychiatric, and perinatal factors on breastfeeding initiation, duration, and exclusivity, by triangulating Mendelian randomization and multivariable regression analyses using data from 72,653 mothers and 317,651 offspring across four European cohorts. Triangulated results robustly demonstrated that higher education, lower BMI, and lower propensities for smoking, insomnia, and depression improved breastfeeding success. Each additional 3.4years in education increased initiation odds by 2.32 folds (95% CI:1.94,2.77) and prolonged exclusive breastfeeding ({beta}=0.21standard deviations, 95% CI:0.17,0.24). Smoking, depression and BMI mediated 26%, 14%, and 12% of education effect on exclusivity, respectively. We found little evidence for effects of blood pressure, cholesterol and perinatal factors. We provide new robust evidence that maternal cardiometabolic and psychiatric factors partially mediate the causal effect of maternal education on breastfeeding. Interventions targeting maternal health could support breastfeeding, reducing maternal and infant health disparities.

Significance StatementAdaptive behavior depends on knowing when to persist and when to let go; even when letting go appears as avoidance. While classical accounts of avoidance emphasize reward-effort trade-offs, we show that these decisions are critically guided by meta-control and inferences about outcome controllability and agency. Using a novel paradigm, we dissociate drivers of avoidance and demonstrate that threat does not uniformly promote disengagement. When outcome control is preserved, threat instead increases persistence, particularly following experiences that build agency in failure-safe contexts. We formalize these dynamics in the Meta-Arbitration of Control and Agency Q-learning (MACA-Q) model, which captures how experience-dependent beliefs about agency guide learning and choice across contexts. Our results show that similar avoidance behaviors can arise from distinct computational pathways. This shifts the focus from global avoidance biases to the dynamic regulation of agency as a core principle of adaptive behavior, with implications for neuroscience, psychiatry, and adaptive artificial intelligence. Adaptive behavior requires deciding when to persist and when to disengage under uncertainty and partial outcome control. Avoidance has often been studied as a response to threat or cost, yet existing paradigms cannot disentangle whether disengagement reflects threat sensitivity, expected failure, or reduced perceived control. We introduce a persistence-escape paradigm that independently manipulates incentive structures, effort demands, and outcome controllability. In a large online sample (N = 457), we show that avoidance is context-dependent rather than a stable, global trait. When outcome control was preserved under threat, the typical avoidance response reversed, promoting persistence rather than withdrawal. At the individual level, high-performing individuals were not uniformly more persistent, but more selective, disengaging when control was low. Moreover, higher anxiety symptoms were linked to cost-dominant evaluation and reduced use of accumulated competence. Conversely, higher depressive symptoms were linked to diminished sensitivity to effort and higher expected failure. To explain these behavioral patterns, we developed the Meta-Arbitration of Control and Agency Q-learning (MACA-Q) model, which embeds value learning and affective evaluation within a meta-control architecture. Critically, we formalize agency as a dynamically inferred learning gate, distinct from self-efficacy, that determines whether outcomes are treated as informative based on controllability and feedback reliability. The model explains context-specific avoidance and reveals that similar behaviors can arise from distinct computational pathways. It further shows how experience in failure-safe contexts guides subsequent behavior in adverse contexts. Our findings show that avoidance is guided by the dynamic regulation of engagement based on inferred controllability and competence. By combining a novel paradigm with a computational model, we provide a formal account of agency and a unifying framework in which meta-control regulates adaptive and maladaptive engagement across contexts, with implications for neuroscience, psychiatry, and adaptive artificial intelligence.

Speech comprehension draws on both temporal structure and contextual prediction, yet how these mechanisms coordinate is poorly understood. Time-compressed speech provides a controlled probe: by degrading temporal structure, it reveals the architecture of ordinary speech comprehension. Using 3x compression with silence insertion, we varied delivery rate, temporal regularity, and boundary alignment (syllabic vs. time-defined) across two behavioural experiments. Comprehension peaked near the upper theta boundary and declined at slower and faster rates. Temporal regularity helped only when boundaries coincided with syllabic onsets, while periodic pacing alone was insufficient. Contextual predictability (word-level entropy) facilitated comprehension when temporal cues were least effective, but only under syllabic segmentation. Computational modeling confirmed that {beta}-mediated contextual prediction selectively benefited syllabic-aligned conditions, was detrimental under time-based segmentation, and better reproduced human pattern overall. Together, these results suggest that contextual prediction is continuously active but behaviorally visible only when temporal scaffolding is insufficient and syllabic structure is preserved.

Extracting stable individual traits from behavior observed across diverse contexts is a central challenge in behavioral modeling. We propose a framework for inferring domain-invariant individual latent representations by jointly encoding behaviors across multiple domains. Using large-scale telemetry data from professional Counter-Strike 2 gameplay, we demonstrate that these representations are stable across distinct environments and roles, improving behavior prediction in novel domains. Our analysis reveals that complex idiosyncratic movement policies can be effectively compressed into low-dimensional embeddings, with as few as two dimensions capturing the majority of individual strategic variation. Crucially, the learned latent space forms a structured metric space where Euclidean distances predict the degradation of transfer performance. Furthermore, we show that the latent axes align with interpretable behavioral phenotypes, such as risk-taking and social cohesion. These findings suggest that multi-domain integration is a robust method for uncovering the functional structure of latent individuality in complex decision-making tasks, bridging the gap between high-dimensional telemetry data and meaningful psychological constructs.

Background Digital services for sexually transmitted and blood borne infection (STBBI) testing may influence demand in publicly funded health systems by enabling low barrier, self-directed access to testing, raising concerns about repeated use and sustainability. We examined longitudinal utilization of GetCheckedOnline, British Columbias digital STBBI testing service, to characterize testing trajectories and assess factors associated with higher intensity use. Methods We conducted a retrospective cohort study using GetCheckedOnline program data for users who created an account between April 2020 and November 2022, with 24 months of follow-up. We used group-based trajectory modelling to identify patterns of testing over time among (1) all users and (2) users with at least one test. Multilevel regression models with local health area random intercepts were used to examine associations between higher intensity trajectory membership, individual risk indicators, and geographic clustering. Results Among 34,228 users, 22,542 (65.9%) completed at least one test and 42,451 tests were conducted (median 1; range 0-44). Two trajectories were identified in both analytic samples, with a minority demonstrating sustained higher intensity testing. The top 10% of users accounted for 39.6% of tests. Higher intensity trajectory membership was associated with sexual risk indicators including having multiple partners, condomless sex with multiple partners, and prior STBBI diagnosis. Geographic clustering across local health areas was modest in the null model (ICC 0.042) and attenuated with adjustment. Conclusion GetCheckedOnline utilization reflects a prevention-oriented pattern that appears more consistent with service needs than indiscriminate overuse. A small subset of users with elevated sexual risk account for higher-intensity testing. Findings support risk aligned stewardship including education and differentiated guidance, rather than universal restrictions to reducing testing volumes.

Genetic variation contributes to individual differences in food liking and dietary behaviour. Genome-wide association studies (GWAS) have identified genetic variants associated with these traits, but most evidence comes from middle-aged and older populations. Young adulthood is a key life stage during which long-term dietary habits develop, yet the genetic basis of food liking during this period remains largely unexplored. We conducted GWAS of 97 food liking traits and two derived principal components (PCs) in 2,784 young adults (age 25) from the Avon Longitudinal Study of Parents and Children. The PCs captured broader food preference patterns reflecting preferences for diverse plant-based and seafood foods (PC1) and meat-based foods (PC2). GWAS identified 32 genome-wide significant associations across 24 traits. Cross-trait analyses indicated that several variants influenced liking across groups of related foods. For example, the lentil-associated variant rs76659918 showed associations with multiple foods, including honey, plain yogurt, chilli peppers, aubergines, avocado, and black olives, as well as PC1, whereas variants associated with bacon, burgers, and steak were linked to multiple meat-based foods and PC2. Exploratory analyses showed that TAS2R38 bitter-sensitive alleles were associated with lower liking for Brussels sprouts, with limited evidence for associations with other traits. Comparison with GWAS of food liking in the UK Biobank cohort (age 37-73) showed limited replication, with robust evidence only for the grapefruit-associated locus. This study identifies genetic variants associated with food liking in young adulthood and suggests that genetic influences operate at both the level of individual foods and broader food preference patterns.

This study developed a large language model (LLM)-based solution to identify people at HIV risk using electronic health records. We transformed structured EHR data, including demographics, diagnoses, and medications, into narrative descriptions ordered by visit date and applied GatorTron, a widely used clinical LLM trained on 82 billion words of de-identified clinical text. We compared GatorTron with traditional machine learning models, including LASSO and XGBoost. We identified a cohort with 54,265 individuals, where only 3,342 (6%) had new HIV diagnoses. Our LLM solution, based on GatorTron, achieved excellent performance, reaching an F1 score of 53.5% and an AUC of 0.88, comparable to traditional machine learning approaches. Subgroup analysis showed that, across age, sex, and race/ethnicity groups, both LLM and traditional models achieved AUCs above 0.82. Interpretability analyses showed broadly consistent patterns across LLM models and traditional machine learning models.

Dopamine is believed to modulate not only instrumental learning about the link between states, actions, and outcomes but also reflexive behaviours, such as a Pavlovian bias to approach in rewarding states and freeze in aversive ones. We studied these dual roles in the human brain, by combining intracranial dopamine recordings from the anterior cingulate cortex (ACC)-- a region implicated in behavioural and cognitive control -- with a motivational Go/NoGo task involving conflict between instrumental and Pavlovian action selection. We found evidence that dopamine in the ACC is involved in evaluating whether Pavlovian responding should guide behaviour. This computational motif was observed across multiple task events, including in response to rewards and punishments, and in analyses based on a reinforcement learning model. Our results indicate that dopamine supports learning at the more abstract level of behavioural policies in addition to the more concrete levels of states and actions.

Many choices are triggered by discrete events whose timing determines which options are rewarded. Without informative sensory evidence between events, behavior must rely on internal estimates of latent variables--most notably elapsed time and reward probability. Existing computational frameworks, including evidence-accumulation models, are not designed for this regime, leaving the principles of time-dependent choice unresolved. Here, we formalize choice as an inference problem governed by uncertainty about both elapsed time and reward over time. Participants learned dynamic reward probabilities to guide choices. Behavior approached optimality but exhibited a systematic distortion of inferred reward probability over time, captured by a linear transformation in log-odds space. Crucially, temporal uncertainty was modulated by reward probability but not by elapsed time itself, contradicting Weber-law scaling. These results identify two interacting computational principles-dynamic mapping of reward probability to choice and reward-based temporal precision- that jointly shape behavior when time and reward must be inferred.

Objective. Menopause is a significant physiological transition with implications for health outcomes (e.g., cardiometabolic), yet gaps remain in understanding the menopause transition, including how menopause timing and type influence health outcomes. Large-scale cohort studies in midlife (age~40-60) females, including the All of Us Research Program (AoURP), provide opportunities to study menopause across diverse populations and data modalities. We characterized menopause-related data in AoURP, focusing on age distributions and concordance between EHR diagnosis codes and self-reported survey responses. Methods. We analyzed menopause-related survey, EHR diagnostic code, and genomic data among ~396,000 participants in AoURP with female sex. We summarized menopause data across modalities, overlap between survey, EHR, and genomic data, and age distributions overall and across sociodemographic characteristics. Results. Among ~396,000 females, surveys captured ~193,000 menopause observations, nearly seven times more than structured EHR diagnoses (~28,000), suggesting under- ascertainement in EHR data. Nearly all females (~99%) with an EHR menopause diagnosis also reported menopause in the survey. Approximately 22,000 participants had intersected EHR, survey, and genomic menopause-related data. Survey-based age patterns matched expectations, with participants <40 years predominantly reporting pre-menopausal status and those >60 years predominantly reporting post-menopausal status. A small subset (N{approx}1,700; 4%) (age>70 years) reported no menopause, suggesting response or recall bias. EHR menopause codes were concentrated after age>45 years, with a notable spike at age 65. Modest differences in survey-based menopause age distributions were observed by sociodemographic characteristics (e.g., race, ancestry). Conclusions. These findings inform sampling strategies, power calculations, phenotype definition, and study design for menopause research using AoURP.

Selective participation in biobanks often compromises inference to the general population, particularly when selection occurs across multiple stages, whether at recruitment or during subsequent participation. Inverse probability (IP) weighting can reduce systematic differences using suitable external benchmarks, but most applications assume a single selection process. Here, we present a multi-stage IP-weighting framework and apply it to the Danish Blood Donor Study (DBDS), a nationwide biobank embedded in Denmark's blood-donation infrastructure. Using national registers, we estimated year-specific probabilities of (i) donation activity and (ii) DBDS enrolment conditional on donation activity, yielding two-stage inclusion weights for 169,893 participants. These weights reduced inclusion-associated imbalance across the 52 auxiliary variables in the probability models by 97.6% (median) and, despite strong health selection under donation-based recruitment, reduced relative-prevalence discrepancies across held-out prescription phenotypes by 69.7% (median). The effective sample size after weighting was 30,627 (18.0% of 169,893). Combining the inclusion weights with questionnaire-specific response weights across five DBDS questionnaires (>500 questions) produced the largest changes from unweighted to weighted responses for health behaviours and symptom severity, including tobacco and alcohol consumption, menstrual-pain severity, restless-legs severity, nocturia, sleep disturbance, and fatigue. These findings support multi-stage IP-weighting to improve population alignment in biobanks with staged selection.

Menstrual cycles are major biological events with extensive effects on the brain and cognition, experienced by half of the human population. To develop a comprehensive account of human cognition, it is necessary to successfully integrate and characterize menstrual cycle effects in cognitive science research. However, current approaches to menstrual cycle analysis suffer from low data resolution and are not well-equipped to capture the highly variable, non-linear changes in outcomes of interest across the cycle. We present a validated standardized method remedying these issues, demonstrate its utility using hormonal, behavioral, and neuroimaging data, and provide an open-source toolkit to facilitate its use.

Sex differences are commonly observed in neuroimaging phenotypes and in the risk of brain diseases, yet the underlying genetic mechanisms remain poorly understood. We investigated sex differences in the genetic architecture of 805 neuroimaging phenotypes in 22,950 males and 22,950 females matched for sample size and covariates, and systematically compared sex-stratified with sex-combined genetic analyses. We found eight variant-trait associations with significant sex differences, 235 fine-mapped sex-dominant causal associations, 457 sex-dominant colocalizations with sex hormones, and 96 sex-dominant colocalizations with schizophrenia. Compared with sex-combined analysis, sex-stratified analysis identified 47 new genetic associations, 170 new fine-mapped causal associations, 1,019 new colocalizations with sex hormones, and 191 new colocalizations with schizophrenia. Additionally, sex-stratified analysis improved global heritability and genetic-correlation estimates and enhanced polygenic prediction for certain phenotypes. This work highlights the need to routinely perform sex-stratified genetic association analyses to elucidate sex-specific and sex-shared genetic control of neuroimaging phenotypes and related disorders.

Brain function is organised in distributed circuits in which regional engagement unfolds over time, reflecting coordinated and temporally ordered patterns of neural computation and information flow. Neural activity also depends on a reliable and scalable supply of glucose. Yet, the temporal order and direction of metabolic signalling in brain circuits remain unknown. Here, we combine functional Positron Emission Tomography (fPET) with 18F-flurodeoxyglucose and Granger causality analysis to characterise directed metabolic connectivity in cognitive control, memory and affective regulatory circuits in 86 healthy adults. We observed widespread directed metabolic influences within the circuits, with the strength of connections a significant predictor of cognition and affect. The behavioural value of the connections was also governed by the efficiency with which baseline glucose metabolism was converted into adaptive functional connections. We conclude that the brain is organised into metabolic circuits that coordinate temporally ordered connectivity to enable information transfer and modulate cognition and psychosocial function. Directed connections vary in their efficiency of glucose use and functional benefit, suggesting that metabolic signalling does not follow a simple "more is better" rule but reflects context-dependent optimisation across cognitive systems.

Metacognition refers to the capacity to monitor ones own actions, internal states, and cognitive processes. A central question in cognitive neuroscience is whether metacognitive evaluation operates as a direct readout of performance signals or requires computationally independent neural mechanisms. Single-process theories propose that both arise from shared decision variables, while the Higher-Order Representation theory holds that metacognition requires re-representation through distinct computational processes. To test these frameworks, participants produced timed motor intervals and evaluated their own performance without external feedback, termed temporal error monitoring (TEM). Vision Transformer decoding applied to PCA-optimized single-trial EEG captured {theta}, , and {beta} dynamics during both task phases. First-order timing was decodable from any individual frequency band, whereas second-order metacognitive inference required simultaneous integration across all three bands before action termination. Individuals whose metacognitive states were more accurately decoded showed stronger TEM precision, with no equivalent relationship observed for first-order performance decoding. These findings establish metacognitive evaluation as a computationally distinct process requiring higher-order multi-band neural integration rather than a direct readout of first-order timing signals.

The genetic architecture of complex traits spans a continuum of polygenicity, yet it remains unclear how differences in polygenicity relate to the functional localization of SNP heritability across the genome. We use a MiXeR-based framework to partition heritability across exonic, intronic, and intergenic regions for 34 traits and introduce a likelihood-based annotation contribution score that quantifies annotation-specific impact on heritability. Exons explain a minority of heritability, and their contribution decreases with increasing polygenicity, from an average of 22% in less polygenic somatic diseases and biomarkers to 13% in highly polygenic psychiatric and cognitive phenotypes. Intergenic fractions show the opposite trend, whereas intronic fractions remain relatively stable. Analysis of a broader set of functional annotations reveals systematic differences along the polygenicity axis: highly polygenic traits show stronger contributions from comparative genomics and variant-effect scores, whereas less polygenic traits show stronger contributions in promoter, transcription, and chromatin annotations. Together, these results indicate that the functional partitioning of heritability systematically varies with polygenicity, pointing to a shift from gene-proximal regulatory architectures to architectures shaped by numerous dispersed regulatory effects as a key determinant of differences in polygenicity across traits.