Gender polarization in fields such as engineering or nursing has served as a barrier to students whose genders do not conform to these norms. Cimpian et al. demonstrate that this polarization also protects lower-performing students who do conform, while impeding LGBTQ+ students in male-dominated fields.
Gender segregation across college majors persists, yet how these patterns vary across achievement levels and intersect with LGBTQ+ identity remains unclear. We examine whether field cultures and social norms may create 'in-group benefits' for lower-achieving members of field-specific majority groups. Using two nationally representative US longitudinal datasets (the High School Longitudinal Study, n = 6,340; and the Baccalaureate and Beyond Class of 2016, n = 12,950), we estimate relative representation across majors spanning male-majority physics/engineering/computer science to female-majority health fields. Across datasets, representation is most polarized among lower-achieving students, particularly in male-majority fields, where lower-achieving straight-cisgender males are overrepresented relative to lower-achieving straight-cisgender females (relative ratio, 2.62; 95% confidence interval, 2.29-3.00), while this representation gap is smaller among higher-achieving students (P < 0.001). The patterns persist after adjusting for achievement, coursework, interests and aspirations. In female-majority fields, evidence for analogous in-group benefits is mixed across models vis-à-vis LGBTQ+ identity. The findings suggest that the forces shaping field-level sorting operate differently across achievement levels, with implications for talent development and broadening participation.
As artificial intelligence (AI) shifts from tool use to social interaction, company-initiated updates can disrupt user well-being. Here we develop an attachment-based account in which AI companions function as attachment figures such that disruptive updates are associated with separation distress. Across two natural experiments-Replika's erotic role play removal and ChatGPT's GPT-5 rollout-we analysed 54,861 posts from Replika and ChatGPT subreddits and analysed data from 1,452 participants across seven surveys. Both updates increased negativity, loss framing and restoration desires (Replika negative posts, +24.7 percentage points, 95% CI 20.1 to 29.2; ChatGPT, +13.0 percentage points, 95% CI 10.8 to 15.2), with larger Replika (versus ChatGPT) increases in sadness (d = 2.67 versus 1.41) and negative mental health (d = 1.72 versus 0.63). Replika users reported closeness exceeding common human ties (versus friend d = 0.47) and anticipated mourning higher than other technologies (d = 0.32-0.57). These findings inform how companies and regulators evaluate psychological risks of generative AI updates.
We argue that contemporary scientific systems progressively constrain high-risk and conceptually innovative research while being increasingly structured around short funding cycles, productivity-based evaluation criteria and risk-averse frameworks that favour predictable and non-transformative research outputs. Drawing on recent empirical literature, we show how such systems can place disproportionate pressure on early-career researchers by incentivizing safe, tractable and easily evaluated outputs. Structural academic mechanisms such as peer review and funding, escalating publication costs and institutional inequalities interact with precarious employment and hierarchical dependencies. In this environment, we contend that strategic conformity becomes a rational career path. This risks suppressing creativity and critical thinking precisely at the stage when scientific independence could otherwise emerge. Consequently, the probability of substantive contributions by early-career researchers is declining, while talent may increasingly abandon academia or cluster within a limited number of well-resourced institutions and national systems. By adopting a systems-level perspective, we argue that scientific creativity is not merely an individual trait but an emerging property of a supportive academic landscape and that maintaining or restoring it may require substantial structural reforms. These include promoting stable research pathways, decentralized decision-making and evaluation frameworks that better recognize collaboration, originality, persistence and nonlinear career trajectories. Without systemic change, we risk stifling the potential of early-career researchers to go beyond the confines of existing methods and approaches and deliver transformative advances. This would limit their capacity to meaningfully change our understanding of the world or benefit society, with impacts that fall short of their potential.
Moral language often travels widely online, but does more moral content always correspond to higher engagement? We analysed 1,621,147 observations across 13 socio-political topics on Twitter (n = 530,104), Reddit (n = 1,048,653) and 8chan (n = 42,390). Using Distributed Dictionary Representations-word embeddings scored against an expert-validated moral dictionary-we measured moral loading (that is, a post's overall moral relevance) and moral density (that is, concentration of moral content across words). Negative-binomial models showed that moral loading was positively associated with engagement (range 1.12 [0.96, 1.28] to 9.07 [8.21, 9.93], all P < 0.001). Conditional on moral loading, however, moral density was 'negatively' associated with engagement (range -4.71 [-5.41, -4.02] to -0.40 [-0.52, -0.27], all P < 0.001). Engagement peaked at density 0.30 [0.30048, 0.30070], P < 0.001, with lower engagement below (2.28-fold) and above (2.78-fold), consistent with an engagement advantage for moral language that is bounded by an overmoralization penalty pattern.
Paying people for their biometrics raises complex ethical and legal questions about consent, privacy and the long-term risks of handing over irrevocable personal data. This Comment examines why such models require caution and what safeguards could better protect the public.
Code underlying published findings in the social sciences often fails to reproduce reported results when others re-run it on the original data. This Comment proposes four recommendations to strengthen computational reproducibility, facilitating the trustworthiness and reliability of research.
Across the behavioural sciences, artificial intelligence (AI) will speed up research that is digitally accessible, scalable and often Western, educated, industrialized, rich and democratic (WEIRD). But evidence about human behaviour cannot always be scaled. We need to connect rapid, automated work with slower research in real settings.
Here, in five waves of panel data (N = 1,966 US adults), we examined associations between life satisfaction and self-reported use of ten common social technologies measured every 3 months on a six-point frequency scale from 'I did not use' to 'multiple times daily'. At this measurement level and timescale, Bayesian and frequentist random-intercept cross-lagged panel models showed little credible evidence that any social technology use predicts subsequent life satisfaction. In the reverse direction, increases in life satisfaction predicted only modest increases in (video) calling in select demographic groups. In analyses comparing different people, frequency of texting was associated with higher life satisfaction, whereas frequency of YouTube and TikTok use was associated with lower life satisfaction. Despite limited ability to detect within-person change due to temporal stability in responses, the absence of cross-lagged effects is informative: there is scant evidence of a meaningful relationship between social technology use and subsequent life satisfaction.
The relationship between of legislative changes in abortion regulations on women's decision-making remains understudied. In 1991, Japan reduced the gestational limit for abortion from 24 to 22 weeks of pregnancy. Here, using national data on abortions, stillbirths and live births at ≥12 weeks from 1989 to 1994 (n = 7,364,783), we examined the association of the change with abortion incidence rate and cumulative incidence between 12 and 27 weeks using quasi-experimental interrupted time-series and regression discontinuity analyses. After the change, incidence rates increased at 20 weeks (incidence rate ratio (IRR) 1.28, 95% confidence interval (CI), 1.15-1.43) and 21 weeks (IRR 1.33, 95% CI, 1.18-1.50), while decreasing at 22 weeks (IRR 0.25, 95% CI, 0.21-0.30, P < 0.001) and 23 weeks (IRR 0.07, 95% CI, 0.06-0.10, P < 0.001), and no meaningful changes for ≤19 or ≥24 weeks. Cumulative incidence decreased only for women at 19-23 weeks. In conclusion, the legislative change shifted abortions to earlier gestational ages, and was associated with fewer abortions among women who had reached 19th week of pregnancy.
Fairness concerns shape how people judge wages, institutions, taxes, transfers and social conflict. Decades of research show that people care not only about their own outcomes but also about how outcomes are distributed across others. The key question is now how these concerns operate outside simple laboratory settings, in real social and institutional environments. We argue that fairness concerns are shaped by comparison groups, beliefs about what outcomes are deserved, social norms, early-life experience and the institutions that make some comparisons more visible than others. This means that policies and organizations do not merely respond to fairness preferences; they can also activate, redirect and reshape them. Progress requires methods that connect experiments with field data, institutional variation and richer measures of how people evaluate whole income distributions. Such work can improve the design of workplaces, welfare policies and public institutions that are both effective and seen as legitimate.
Assortative mating (AM), the tendency to choose partners with similar traits, contributes to the genetic architecture of human traits. Detecting its genetic footprint (genetic AM) requires either genotyped couples or detection of gametic phase disequilibrium (which is insensitive to recent changes). We introduce an inter-chromosomal phasing approach that reconstructs parental haplotypes from close relatives and correlates maternally and paternally inherited polygenic scores to provide estimates enriched for recent AM. Applied to 69 traits in 245,884 UK Biobank and 194,325 Estonian Biobank individuals, our approach showed concordance with mate-pair-based estimates comparable to or better than existing approaches, significantly outperforming them in the Estonian Biobank (through-origin r2 = 0.638 versus 0.479, P = 0.005) and performing comparably in the UK Biobank (r2 = 0.737 versus 0.731). We identified genetic AM for traits such as height, education, overall health and time spent watching TV, with assortment for height and education accelerating in recent generations (for example, θHAP = 0.041 (0.034-0.048) versus 0.015 (0.007-0.024) for height in post-1980 versus pre-1960 Estonian Biobank cohorts).
Why have incarceration rates and racial disparities fallen in the USA? Using data on recorded crimes, arrests and prison admissions from 2000 to 2019, we decompose declines in imprisonment by race and offence category. Whereas the rise of mass incarceration primarily reflected increased punitiveness in policing and case processing, recent declines in imprisonment for violent and property offences can be attributed almost exclusively to reductions in recorded crime rates, particularly to declines in violence among Black Americans. For drug offences, however, the pattern appears different. There is little evidence of declining drug use. Rather, drug prison admissions appear to have fallen owing to reductions in the likelihood of arrest and imprisonment. For Black individuals arrested for drug-related offences, the probability of imprisonment was cut in half since 2000. Overall, we find clear evidence that the forces that produced mass incarceration differ substantially from those associated with the contraction of America's carceral state.
The coronavirus 2019 pandemic disrupted food purchasing behaviours, triggering both short- and long-term nutritional shifts. Here, analysing 9,367,550 weekly shopping baskets from over 30,000 US households (2017-2022), combined with county-specific policies and artificial intelligence-enhanced nutrition profiling, we document distinct changes in household food baskets. During the pandemic and lockdowns, households increased spending on less processed foods, but by the post-pandemic period, spending on ultraprocessed foods rose above pre-pandemic levels. This divergence appeared strongly in processing-based measures (NOVA) but less in nutrient-based profiles (Healthy Eating Index and Food Standards Australia New Zealand), underscoring their complementary perspectives on diet quality. The spending gap between socioeconomic groups widened in 2020 but converged post-pandemic, while a new nutrition gap emerged as higher-income households increased spending on less nutritious foods. Categories such as prepared and frozen foods and carbonated beverages showed persistant effects. These findings highlight the pandemic's uneven, lasting influence on dietary behaviour and the need for policies promoting healthier, more resilient food systems.
Many natural disasters central to climate-policy debates are hydro-meteorological hazards, yet it remains unclear how media attention is distributed across disaster types. Using 135 million news articles from 466 sources in 123 countries since 2016, we estimated cross-border reporting changes after disasters. Attention rises but is highly uneven (95% confidence intervals are shown in parentheses after the estimates): short-run increases are largest following earthquakes (β = 0.0785 (0.0758, 0.0812)), dry-mass movements (β = 0.0531 (0.0349, 0.0713)) and volcanic eruptions (β = 0.0425 (0.0359, 0.0490)), and smaller for floods (β = 0.0069 (0.0062, 0.0076)) and droughts (β = 0.0001 (-0.0036, 0.0039)). Conditional on severity and duration, hydro-meteorological disasters receive less attention than geophysical disasters (θ = -0.0065 (-0.0115, -0.0015)). Coverage is higher for 100+ versus 0-9 deaths (β = 0.0360 (0.0231, 0.0490)), with stronger fatality gradients for country pairs with tighter social ties (β = 0.1498 (0.0822, 0.2174)) and ancestral relatedness (β = 0.2965 (0.1887, 0.4043)).
Funding agencies play a central role in science by awarding grants to researchers. Grant review can be influenced by bias, including preferences for applicants who share reviewers' gender, country or citizenship. Using 37,073 applications for European Research Council grants with multiple reviews per proposal, we show that reviewers who share an applicant's country of work or citizenship award higher scores than reviewers with no such match. This effect is strongest when both country and citizenship align. Shared gender increases scores to a smaller extent. Further analyses indicate that these findings are unlikely to be driven by shared tastes or informational advantages. At the panel level, correlational evidence shows that applicants are more likely to advance past the first stage when the chair or more panel members share their country of work or citizenship. Country of work and citizenship may therefore be relevant sources of bias in international grant review.
In non-human primates, forelimb nerve transection and repair alters the otherwise orderly and highly conserved organization of the digit maps in primary somatosensory cortex (S1). These same changes are presumed to occur in humans and to have meaningful implications for patient recovery. Here, using functional magnetic resonance imaging, we map digit responses in 21 patients with surgically repaired hand nerves and 30 controls. In S1 contralateral to the repaired hand, digit map arrangement is systematically altered and response amplitudes are elevated. At the individual level, map organization is more variable and less typical of controls, a pattern that, unexpectedly, is also evident in S1 of the uninjured hand. Our results show that hand nerve repair alters S1 organization in humans, consistent with animal models. These changes probably reflect disordered peripheral inputs following unguided nerve regeneration. The cortical circuitry itself may remain largely stable. The functional relevance of these changes remains uncertain. We find no credible links between altered S1 maps and functional impairments.
Human ticklishness is one of the oldest yet unresolved scientific puzzles. As early as the time of Aristotle and Socrates, scholars have speculated on its origins and function, but fundamental questions remain. Is ticklishness culturally shared or specific? Does it follow a structured bodily topography, and how does it relate to other bodily sensations? Here we analysed data from three cultural groups (Chinese, Dutch and Greek) and reveal consistent behavioural patterns and social dynamics. We extracted a high-resolution bodily map of ticklishness and show that it generalizes across cultures and individuals, and is distinct from maps of touch sensitivity, pain and pleasure. When tested against predictions from five historical and contemporary theories, we show that no single theory can fully explain the ticklishness topography. Although ticklishness is stronger in rarely touched regions, supporting Darwin's hypothesis, its topography relies on a psychophysiological mechanism that is more complex than previously assumed.
