The research responds to a persistent gap in the literature on computational thinking (CT) integration in science education. Although a common rationale for weaving CT into science curricula has been the promise of boosting learning in both domains simultaneously, few empirical studies have actually demonstrated equivalent gains in science learning. Across two systematic reviews of CT integration literature published between 2020 and 2022, only 19 empirical studies assessed both science and CT learning, and just 10 were impact studies that compared students’ knowledge and skills before and after an intervention. Several of those studies found statistically significant gains in both domains, but one found no significant gains, and there has been little consensus on how or why CT integration might deepen science understanding rather than simply compound domain-related challenges.

The theoretical foundation of the Decoding Approach is mechanistic reasoning, a construct that science educators have long sought to cultivate. Following frameworks developed by Russ and colleagues and later consolidated by Schwarz and colleagues, mechanistic reasoning spans three levels: Level 1 describes only the surface characteristics of what happened; Level 2 addresses how a process unfolded over time or space through co-occurrence and sequencing; and Level 3 addresses why the phenomenon occurred, attending to entities, their actions, and cause-and-effect chains. Because modern science standards such as the Next Generation Science Standards emphasize understanding how and why phenomena occur — not merely retaining facts — the researchers reasoned that decoding, which examines causality in both the real-world phenomenon and the modeled world, could push students toward Level 3 reasoning and enable them to transfer their understanding of causal mechanisms to new systems.

Decoding is rooted in established ideas from cognitive science. The real world and the modeled world serve as “contrasting cases” that focus learners’ attention on what to notice, and decoding adds the explicit identification of similarities between the two contexts, making it a form of analogical reasoning. The approach also draws on the professional practice of model validation, though the researchers distinguish it from expert-level validation: the middle school participants in decoding were just beginning to learn the phenomena being modeled, and the goal was not to produce expert-caliber predictive models but to help learners inch toward fuller understanding by bridging code, simulation, and scientific information. The approach sits between a “programming-first” model of CT integration, in which students create models of their own, and a “simulation-only” approach, in which students manipulate a model through an interface without ever seeing the code that generates the simulation.

The intervention took place over three cohort years from 2021 to 2023, initially planned as an in-person three-week summer program followed six months later by a two-day continuation workshop. Implementation varied considerably due to the COVID-19 pandemic: the first cohort totaled 40 contact hours conducted entirely virtually, the second was hybrid with 60 contact hours, and the third totaled 70 contact hours conducted fully in person in museum classrooms and exhibit halls. Participants were New York City youth between the ages of 11 and 13, recruited with strategies designed to reflect the demographics of the city’s school-age population, including partnerships with public schools and organizations serving under-resourced students from groups underrepresented in STEM and computing. Applicants were bucketed by gender, ethnicity or race, and school or program affiliation, then randomly drawn into treatment and comparison groups; students selected for the comparison group were offered spots in the following year’s full program.

The curriculum focused on the authentic scientific problem of rising Lyme disease rates in New York City, combining real-world investigation, data collection and analysis, and computational modeling with StarLogo Nova, a block-based modeling environment developed at MIT. Students began with instruction on ecosystem concepts — food webs, carrying capacity, ecological relationships — and explored systems through physical models and embodied simulations in which they acted out the roles of individuals in a system. The first focal model depicted population dynamics and energy flow among acorns, mice, and foxes. In early “Use” activities, students manipulated population sizes with slider variables; by day four, code was introduced and students began decoding, mapping mechanisms in the real world to those in the code. Later “Modify” activities required students to change code so foxes could eat both acorns and mice, and to compare mouse code across two different models and justify the differences. In the final week, students created new code, adding agents such as wild turkeys, dogs, and opossums to the tick-mouse-deer Lyme disease model, using decoding in both directions as they decided what behaviors to include and how to represent them.

The continuation workshop was designed partly as a research vehicle for studying transfer. After a three-hour refresher with increasingly difficult coding challenges, student teams took part in a five-hour “Hackathon” centered on a new ecological system — zebra mussels in New York waterways — modifying an existing model to represent the invasive species. Because the zebra mussel scenario was new to students, this activity allowed the research team to observe whether students transferred mechanisms such as predation, energy transfer, and disease transmission from earlier models to an unfamiliar context. Teams then presented their modified models, including adaptations of mechanisms from past models.

Quantitative results rested on the Survey of Knowledge and Skills in Computational Thinking (KSCT), a 20-item instrument combining an ecosystems scale, decoding “triplets,” and difficult computer science items, administered before the summer program, immediately after it, and again at the start of the continuation workshop. Baseline comparisons revealed that the treatment and comparison groups were not equivalent — the comparison group skewed toward higher-scoring individuals, which the researchers attributed to attrition driven by the pandemic. Nevertheless, treatment students (n = 46) improved significantly from a pre-test mean of 9.435 to a post-test mean of 12.174, with p < .01 and a large effect size of d = 0.96, while comparison students showed no significant change. Because the groups were non-equivalent, the researchers also conducted an analysis of covariance, which confirmed that group assignment had a significant effect on score change while pre-test scores did not.

Scores by subscale added nuance. On the ecosystems scale, treatment students gained significantly (d = 0.586), though they entered the program with relatively strong understanding — 80 percent could correctly predict the impact of an ecosystem change at pre-test, compared with a typical 54 percent found in prior research. The largest gains appeared on items about ecosystem equilibrium and identifying the logistic growth curve, which rose from 33 percent to 52 percent correct. Understanding of energy loss across food-chain levels changed little. Most striking were gains on the decoding scale, where students matched coded mechanisms to scientific processes: means rose from 1.913 to 3.391 with a large effect size of d = 1.157. A durability analysis of the 27 treatment students who took all three surveys found no significant difference between post and continuation scores, suggesting that learning persisted over the six-month interval, though the researchers note that self-selection into the continuation workshop may have influenced this result.

The qualitative strand of the study centered on artifact-based interviews, in which students thought aloud while investigating agent-based models, plus student work collected during the program. Two reviewers coded transcripts for mechanistic reasoning levels, comparisons between real-world and simulated contexts, and decoding behaviors, working through rounds of coding and discussion until reaching agreement. The researchers distinguished "basic" decoding — identifying a mapping between a code mechanism and a science process without further elaboration — from "advanced" decoding, in which students applied the mapping to make predictions or assess a model's validity. Interview transcripts show students spontaneously identifying that collision code represented infection, or that a "wiggle" procedure controlled movement, even when those words did not appear in the code itself.

Because only four focal students completed the full set of three interviews and three surveys across all cohort years — two cases in year one and two in year three — the qualitative analysis is deliberately limited in scope and was not treated as quantitative evidence. The researchers also acknowledge that the treatment and comparison groups were non-equivalent at baseline, that attrition shaped both quantitative and qualitative samples, and that the durability findings may reflect sampling effects. The implementation itself varied across cohorts in format and contact hours, spanning pandemic-era virtual instruction to fully in-person museum programming, which complicates direct comparisons between years.

Even with these limitations, the study's implications are noteworthy. The evidence from the four case studies shows students using coded mechanisms as an "active" and "executable" representation with which to reason about scientific phenomena, and in some cases transferring that reasoning to entirely new systems such as the lionfish overpopulation scenario posed in the final interview, where students were asked to imagine the code they would write. The authors argue that the results reveal untapped opportunities for deepening students' understanding of ecosystems through CT integration with an explicit decoding emphasis — and, more broadly, that reading, interpreting, and critiquing models built by others constitutes a valuable learning experience in its own right, one that does not require every student to become a model-builder for the synergy between computational thinking and science learning to take hold.

Breast cancer is the most commonly diagnosed cancer among women worldwide, and its outcome depends heavily on the stage at which it is detected. Tumors identified through organized mammography programs are, on average, smaller and less advanced than those found after symptoms appear, which is why health authorities across high-income and middle-income countries have invested heavily in population-wide screening. South Korea operates one of the most comprehensive such programs, offering biennial mammography to women over designated ages through a national infrastructure administered in connection with the National Health Insurance Service. The program’s reach makes the country an unusually informative setting for studying whether organized screening delivers equitable participation.

Screening is most effective when it is repeated at regular intervals, yet most studies of screening disparities have examined only single rounds of testing. The new study takes a longer view, asking not simply whether women ever received a mammogram but whether they completed every one of the five biennial screening cycles recommended across a ten-year window. This distinction matters because organized screening programs are designed to detect tumors early through consistent, population-wide participation, and gaps in adherence can allow cancers to progress undetected between rounds. A woman who attends one appointment but misses the next three receives far less of the program’s protective benefit than the single-round statistics would suggest. Prior research has already documented that women with disabilities experience disparities not only in screening but also in cancer detection and treatment, contributing to poorer outcomes overall. The Korean team set out to quantify how disability status, disability type, and disability severity shape the odds of completing the full course of recommended screening.

To answer that question, the investigators assembled a retrospective cohort by linking two national data sources: the Korean National Cancer Screening Program Data, which records participation in the country’s organized screening services, and the Korea National Disability Registration System, which contains medically verified records of disability. The cohort comprised 4,956,744 women who participated in screening in 2011 or 2012 and were then followed for ten years. Because disability registration in Korea requires medical verification, the researchers were able to classify participants not only by whether they had a disability but also by its officially defined type and its severity. The fifteen disability types recognized under the national system were grouped into six broader categories: physical disabilities, brain injury, communication disabilities, major internal organ disabilities, intellectual disability and autism, and mental disorders.

This administrative linkage gave the study a foundation that survey-based research often lacks. Self-reported screening behavior is known to be imperfect, with respondents sometimes recalling tests that records do not confirm, and self-reported disability can be inconsistent across conditions and cultures. By drawing on government records of both screening attendance and disability status, the investigators minimized recall bias and misclassification, producing a dataset in which the exposure and the outcome were each documented independently of the participants’ own reports.

The primary outcome was deliberately stringent: completion of all five recommended biennial screening cycles over the follow-up period. This measure captures the kind of sustained, long-term engagement that screening programs depend upon, rather than a single episode of care. By anchoring the analysis in a universal, government-organized screening program, the study also offered an opportunity to examine whether disparities persist in a system where cost-related obstacles have been substantially reduced for participants, including provisions designed to ease access for people with disabilities.

The findings were unambiguous. Women with disabilities had lower odds of completing all five screenings compared with women who had no registered disability. The disparity was not uniform, however. The researchers observed a gradient pattern by severity: women with severe disabilities showed the lowest adherence, while those with mild disabilities had slightly higher adherence than their severely disabled counterparts, though the overall pattern still reflected disadvantage relative to women without disabilities. This gradient suggests that the barriers to long-term screening are not simply a binary matter of having or not having a disability, but intensify with the degree to which a disability affects daily functioning and health care access. In practical terms, the more support a woman needs to manage daily life, the less likely she was to remain within the screening program across an entire decade.

Disability type also mattered. Adherence differed across the six categories, with the lowest participation among women with intellectual disability and autism and the highest participation among women with physical disabilities. That physical disabilities were associated with comparatively better adherence, while intellectual and developmental disabilities carried the greatest disadvantage, points to barriers that extend beyond the logistical challenges of physically attending a screening appointment. Communication difficulties, differences in health literacy, caregiver dependence, and the ways screening services are designed and delivered may all play a role in shaping whether women with cognitive and developmental disabilities return for repeated mammograms over a decade. Mammography itself can be an uncomfortable and confusing procedure, and for women who have difficulty understanding explanation, consenting to the examination, or tolerating the positioning it requires, each successive appointment may become harder to arrange and complete without individually adapted support.

A central implication of the study is that financial support alone does not close the gap. Korea’s National Cancer Screening Program provides organized, population-based mammography with reduced or eliminated cost barriers, meaning that the disparities observed cannot be attributed primarily to an inability to pay. Instead, the authors argue, the persistence of these gaps points to structural and procedural obstacles embedded in how screening is offered, scheduled, communicated, and experienced by women with disabilities. The study’s conclusion calls explicitly for policy efforts such as a disability-responsive design of screening services and routine equity monitoring within organized screening programs, so that participation gaps can be tracked and addressed systematically rather than left to emerge unnoticed in aggregate statistics.

The research carries considerable weight because of its scale and design. A cohort of nearly five million women, drawn from nationwide administrative data and followed for a full decade, offers a level of statistical power and generalizability within the Korean context that smaller, clinic-based or survey-based studies cannot match. The size of the cohort also allowed the investigators to examine subgroups that smaller studies often cannot analyze reliably, including women with rarer disability types, without losing precision. The use of medically verified disability registration data also reduces the risk of misclassification that can arise when disability status is self-reported. The study was approved by the Institutional Review Board of the College of Medicine at The Catholic University of Korea, and the requirement for informed written consent was waived because the analysis used de-identified data originally collected for administrative purposes. The work was funded by the National R&D Program for Cancer Control through the National Cancer Center, supported by the Ministry of Health and Welfare, and by a National Research Foundation of Korea grant funded through the Ministry of Science and ICT.

Like all observational research, the study has limitations that shape how its findings should be interpreted. Because it relies on registered disabilities, women who have impairments but are not formally registered would be counted among those without disabilities, potentially diluting measured differences. If anything, this misclassification would tend to make the observed disparities appear smaller than they truly are, which lends additional weight to the gaps the study did detect. The cohort consists of women who participated in screening at the outset, in 2011 or 2012, which means the analysis focuses on long-term adherence among initial participants rather than on initial uptake itself; women with disabilities who never entered the program in the first place would fall outside the cohort entirely, so the full extent of screening inequity may be still greater. As with any retrospective cohort design, the study can identify associations between disability characteristics and screening adherence but cannot definitively establish the mechanisms through which those associations arise. The findings are also rooted in the specific structure of Korea’s organized screening program and disability registration system, and the degree to which they generalize to other health systems with different financing, delivery, and disability classification arrangements would require further study.

Nevertheless, the study adds an important international data point to a growing body of evidence that people with disabilities face systematic disadvantages across the cancer care continuum, from prevention and screening through diagnosis and treatment. International health frameworks have increasingly emphasized that persons with disabilities have the same right to the highest attainable standard of health as everyone else, and that health systems are obligated to remove the barriers that prevent them from exercising that right. By demonstrating that these disadvantages persist in a system with universal organized screening and reduced financial barriers, the Korean findings challenge the assumption that coverage and cost subsidies are sufficient to achieve equity. The severity gradient and the especially low adherence among women with intellectual disability and autism suggest that interventions will need to be tailored: accessible facilities and transportation support may help some women, while others may require adapted communication, longer appointments, caregiver involvement, or individualized reminder and navigation systems.

The authors’ call for disability-responsive design implies rethinking screening programs from the ground up, considering how appointment scheduling, facility layout, mammography equipment, staff training, and follow-up communication accommodate the full range of disability types and severities. Routine equity monitoring, meanwhile, would embed the kind of analysis performed in this study into the ongoing operation of screening programs, allowing administrators to detect widening gaps and evaluate whether interventions are working. As organized screening programs in many countries confront aging populations and growing attention to health equity, the Korean cohort study offers both a warning and a template: long-term adherence data, disaggregated by disability status, type, and severity, can reveal disparities that single-round participation statistics conceal, and addressing them will require deliberate design choices rather than the removal of financial barriers alone.

The investigation rested on semi-structured interviews with 131 academics — 73 men and 58 women — drawn from five universities and colleges across Iran and Canada, two national contexts chosen deliberately for their contrasting governance structures, language ecologies, and levels of technological integration. Purposive sampling targeted faculty who had personally used AI in research writing within the previous twelve months, and candidates without that hands-on experience were excluded. Interviews lasted roughly twenty to thirty minutes and were conducted over Zoom or in person, primarily in English, with a Farsi subset translated and back-checked for conceptual equivalence. Recordings were transcribed with the help of Otter.ai and verified manually, then analyzed through Braun and Clarke’s six-phase reflexive thematic analysis, a qualitative technique that moves from systematic coding through theme development, review, definition, and reporting. Two analysts independently double-coded twenty percent of the transcripts, stratified by discipline and career stage, reaching 85 percent agreement before discrepancies were resolved through discussion, and sampling continued until thematic saturation was achieved.

The dominant first impression was one of efficiency. Faculty described deploying AI for literature reviews, data analysis, manuscript organization, and early drafting, with one social scientist marveling that the technology structures reviews “in minutes, something that used to take weeks.” Well over half reported that AI was especially useful for grammar, language accuracy, and the overall clarity of academic prose. But participants described something subtler than speed. Shahini’s analysis suggests AI is redistributing cognitive labor across the research pipeline: the locating, organizing, and synthesizing that once consumed a scholar’s intellectual resources are increasingly automated, which pushes distinctly human contribution toward judgment, contextual reasoning, and disciplinary understanding. Participants implicitly separated information processing, which machines perform well, from knowledge construction, which they regarded as irreducibly human. In that reading, AI does not diminish scholarly expertise but relocates it — provided researchers remain actively engaged rather than drifting into the role of passive evaluators of machine output, a transition many feared would erode the analytical habits underpinning rigorous science.

The enthusiasm collided hard with the technology’s failure modes. An applied linguistics participant captured the paradox in two sentences: AI “helps me refine my writing,” the participant acknowledged, “but when it generates references, they are often made up.” Roughly half of the faculty reported encountering fabricated or inaccurate citations, a hallmark of large language models, which generate statistically plausible text without querying any verified database of real sources. Participants also flagged factual inaccuracies and superficial analyses, noting that fluent, convincing prose often conceals a lack of analytical depth. Roughly a third raised algorithmic bias: because models are trained on historically skewed datasets, they reproduce existing hierarchies of knowledge and privilege dominant discourses while sidelining minority viewpoints, emerging theories, and context-specific interpretations. “The real danger of AI in research is that it often reinforces biases in its responses,” one sociology professor warned. The concern is epistemic as much as technical — unchecked, bias shapes which questions get asked, which evidence surfaces, and whose scholarship becomes visible.

Ethical questions dominated the interviews, and none proved thornier than accountability. “Who is responsible when AI generates false information in a research paper?” a philosophy faculty member asked. “The ethical implications are unclear.” Participants converged on a clear answer in principle: the human researcher who uses and submits the content carries the responsibility, because AI cannot exercise moral judgment, justify methodological decisions, or answer for scholarly claims. That position mirrors the authorship policies of major publishers, which bar AI from authorship on the grounds that authorship demands accountability and intentionality that machines lack. Yet most participants found current institutional policies dangerously vague, and an engineering professor’s warning — “Universities need strict policies on AI use, or else we risk a crisis of academic dishonesty” — reflected widespread anxiety about AI-generated text passing as original work. More than half favored mandatory disclosure whenever AI contributes to idea generation, writing, or data analysis, reflecting a broader reconceptualization of academic integrity from simply protecting originality toward governing the entire pipeline of AI-assisted knowledge production.

Participants also worried about a quieter casualty: scholarly voice. More than two-thirds identified the homogenization of academic writing as a significant risk, with a literature scholar complaining that AI-written papers “sound generic” and “lack the nuance that comes from real scholarly engagement,” and a historian fearing that AI “will gradually erase individual academic voices.” More than half said AI-generated writing lacks a personal scholarly perspective, while about 39 percent worried that continued advances could eventually diminish the role of human authors altogether. The technical explanation lies in how language models work: they predict the most statistically probable response given the patterns in their training data, which structurally favors consensus over novelty and convention over intellectual divergence. Participants framed the stakes in strikingly epistemic terms — drafting, revising, and organizing arguments are not merely acts of communication but part of how researchers develop their own thinking, so outsourcing them risks converting intellectual discovery into the editing of machine-generated content, thinning the conceptual diversity on which scientific progress depends.

Perhaps the study’s most striking statistic is its near-unanimity on governance: 97 percent of participants said the absence of clear institutional policies has left researchers navigating AI with individual judgment rather than established standards. “We need a structured way to integrate AI into research. Right now, it’s a free-for-all tool,” one computer scientist complained. Roughly two-thirds called for formal AI literacy training covering both the capabilities and the limitations of these systems, ethical considerations, responsible research practices, and critical evaluation of AI-generated content. “Professors and students need AI training to use it responsibly,” a business studies participant insisted. That consensus aligns with influential work in AI ethics, including Floridi and Cowls’s framework emphasizing transparency, accountability, fairness, and meaningful human oversight, and with the research priorities articulated after ChatGPT’s emergence in journals such as Nature. Participants, in other words, were not demanding prohibition — they wanted clearer institutional expectations defining where machine assistance is legitimate and where human responsibility cannot be delegated.

The answer Shahini proposes is HAKI, a framework built from three interlocking domains and four operational processes. The domains are Human-Centric Knowledge — the tacit and explicit knowledge generated by human cognition, experience, and creativity; AI-Assisted Knowledge — structured, retrieved, or synthesized data produced by algorithms, large datasets, and pattern recognition; and Ethical-Regulatory Frameworks — the guidelines that govern how the two interact. The processes turn those abstractions into workflow. Human Analytical Control requires researchers to define the task before touching the tool: specifying keywords, date ranges, inclusion and exclusion criteria, and quality thresholds so that AI use is purposeful and bounded. AI Ethical Mediation mandates open disclosure of any significant AI assistance, along with logging of prompts, model versions, and parameters to create an auditable trail, with AI credited as a tool rather than an author. Critical Interpretation obliges scholars to treat AI output as raw material to be triangulated against primary sources and woven into original analysis. Recursive Verification closes the loop: every fact, citation, and data point is checked against primary evidence, and failed checks send the researcher back to earlier stages.

Conceptually, HAKI is anchored in Sociotechnical Systems Theory, which holds that technological and human elements should be jointly optimized for both effectiveness and ethical responsibility, and in Epistemic Agency Theory, which insists that knowledge ultimately depends on human judgment, evaluation, and accountability. Shahini argues the model fills a hole left by better-known theories: the SECI model of knowledge creation cannot accommodate AI, which lacks the lived experience and intentionality needed for genuine tacit knowledge; Actor-Network Theory grants technologies an agency that cannot exercise consciousness or moral judgment; and the Extended Mind thesis fails to distinguish passive cognitive tools from autonomous generative systems. HAKI instead positions AI as a “dependent cognitive collaborator” inside a human-centered knowledge ecosystem — neither an autonomous knowledge creator nor a mere calculator. It also aims to move beyond existing human-centered initiatives such as human-in-the-loop oversight, Google’s People + AI Research design guidebook, and the ethical principles issued by UNESCO, the OECD, and the European Commission’s high-level expert group, which, Shahini contends, articulate ideals without providing the executable, discipline-sensitive procedures that academic writing actually requires.

The author is candid that HAKI remains a conceptual and exploratory framework, not an empirically validated operating model, and that its contribution is theoretical — a foundation for future investigation rather than a finished product. The framework assumes human oversight can effectively regulate AI-generated knowledge, yet as models grow more sophisticated, their outputs become harder to distinguish from authoritative scholarship, and subtle biases may evade even diligent verification. Enforcing ethical-regulatory frameworks across institutions, publishers, and disciplines that lack standardized disclosure policies poses another hurdle, compounded by uneven AI literacy among researchers. And the model does not yet grapple with what comes next: AI systems that generate original hypotheses, interpret complex data autonomously, or simulate critical reasoning. Shahini’s roadmap calls for AI-assisted fact-checking tools, comparative studies of institutional disclosure policies, standardized authorship guidelines, and the embedding of AI education into research methods training. Until then, the framework’s message is stark and simple: the more machines can do, the more human accountability matters.

Subject of Research: Faculty perceptions of AI-assisted academic research and the development of the HAKI conceptual framework for ethical human–AI collaboration in knowledge production

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02041-4

Keywords: Artificial intelligence, Academic research, Authorship, Knowledge production, AI ethics, HAKI model, Academic integrity, AI literacy

What distinguishes the research is the quality of its evidence and the moment at which it was collected. The team — John Jerrim, Kirtana Krishnakumar, Nikki Shure and Gill Wyness, described in the paper as joint first authors — harnessed a rare pairing of datasets: England’s responses to the 2015 Programme for International Student Assessment (PISA), the OECD’s three-yearly test of 15-year-olds in reading, mathematics and science, and the National Pupil Database, the administrative archive holding every English pupil’s examination record. A successful link was achieved for 4,914 of the 5,194 pupils sampled from 206 English schools, a 95 percent match rate. Timing matters enormously here. English teenagers sit PISA roughly six months before taking their GCSEs, the high-stakes qualifications that publicly rank schools, determine post-16 pathways, decide which subjects students can pursue in the age-18 examinations that gate university entry, and leave measurable traces in lifetime earnings according to government analyses. That half-year stretch — when adolescent distractions multiply and revision pressure mounts — is precisely when motivation and concrete ambition might plausibly translate into extra effort and better grades.

Central to the study is a conceptual distinction that much earlier work has blurred. Achievement motivation was measured with PISA’s validated five-item scale, on which pupils rated statements such as “I want top grades in most or all of my courses,” “I want to be the best, whatever I do” and “I see myself as an ambitious person” on a four-point agreement scale. Item-response-theory modelling converted those answers into a continuous score, standardised within England so that every estimate can be read directly as an effect size. University aspiration, by contrast, was deliberately concrete. England used its allowance for national questionnaire items to ask pupils how likely they were ever to apply for a degree and, if they answered “very likely” or “fairly likely,” to name up to three target institutions in an open-text field. Roughly 2,800 pupils in the analytical sample named at least one destination, with Oxford (11 percent), Cambridge (8 percent), Manchester (5 percent), Birmingham (4 percent) and Leeds (4 percent) topping the list. An “elite” aspiration was defined as naming a Russell Group university — the association of 24 research-intensive British institutions that includes Oxford and Cambridge — and the paper reports that 58 percent of England’s 15-year-olds intended to submit a Russell Group application.

The demographic patterns are arresting. Boys were nine percentage points more likely than girls to plan a Russell Group application in raw comparisons, and the gap persisted at seven percentage points once prior achievement, demographic background and school fixed effects entered the models. Girls made up 58 percent of Russell Group aspirants but only 46 percent of pupils naming Oxford or Cambridge, and the shortfall held across the ability distribution, among low, middle and high achievers alike. First- and second-generation immigrant students — teenagers born abroad, or with two foreign-born parents — were 11 to 12 percentage points more likely than non-immigrant peers to hold elite plans, a difference that survived rich controls and even widened slightly once teenagers were compared within the same school. Socioeconomic status told a more tangled story. Pupils from the top fifth of the ESCS index were 15.5 percentage points more likely than the bottom fifth to name a Russell Group first choice, but roughly half of that gap dissolved once age-11 Key Stage 2 test scores and PISA results were controlled, and the remaining six to eight percentage points lost statistical significance once schools were compared like-for-like.

Motivation itself, however, displayed a socioeconomic gradient that refused to wash out. The raw difference between the top and bottom ESCS quintiles stood at 0.4 standard deviations, and even among pupils with identical prior attainment attending the same school, advantaged teenagers remained 0.31 standard deviations higher on the motivation scale — a gap the authors interpret through the social environments that make high performance feel valuable in the first place. Gender differences in motivation itself were modest, with girls edging boys by 0.09 standard deviations, while immigrant pupils out-motivated their non-immigrant classmates. Strikingly, simply naming a Russell Group institution as a first choice was associated with a 0.4 standard deviation motivation advantage over pupils planning no university at all, conditional on prior ability and background — evidence that drive and concrete ambition tend to travel together.

The statistical machinery behind the headline number was deliberately conservative. The team estimated value-added regression models in which each pupil’s standardised total GCSE points score — capped across the best eight subjects — was regressed on motivation quartiles and university application plans, alongside an unusually dense control set: Key Stage 2 scores from the age-11 national tests, all 30 PISA plausible values across reading, mathematics and science to suppress measurement error, free school meal eligibility, immigrant background, gender, stated career plans at age 30, parental educational support and private tutoring, and separate socio-emotional scales for test anxiety, academic self-efficacy, sense of belonging at school and attitudes to teamwork. School fixed effects confined comparisons to pupils within the same institution, absorbing school culture, peer climate and neighbourhood factors, while the survey’s clustering, stratification and replicate weights were handled through the specialised REPEST software. Previous research showing that PISA scores correlate at 0.78 with later GCSE grades in mathematics underscores why these controls bite so hard. The authors concede one technical wrinkle — PISA’s plausible values were imputed without the study’s fuller administrative controls — but robustness checks using a single value per domain produced substantively identical results.

The results unfold in layers. Unadjusted, Russell Group aspirants out-scored pupils planning no university by 1.13 standard deviations in GCSE points, and the most motivated quartile beat the least motivated by 0.58. Most of that, unsurprisingly, reflects prior ability. But after the core value-added adjustment — before school fixed effects were even added — Russell Group aspirants had made roughly 0.32 standard deviations more academic progress over the final six months than non-planners, and about 0.18 more than peers aiming for non-elite universities, while top-quartile motivation independently added 0.16 to 0.17. Layering in school fixed effects and the full battery of socio-emotional controls barely moved the estimates, leaving the combined advantage for a highly motivated Russell Group aspirant over a low-motivation pupil with no university plans at roughly 0.37 standard deviations. Because the model already conditions on attainment measured at age 11 and again at 15, a jump of that magnitude across a six-month window is, by the standards of education research, very large — and alternative models using GCSE mathematics grades alone reproduced the same pattern.

The authors read these patterns through three established psychological lenses. Situated expectancy-value theory holds that motivation springs from two ingredients: a student’s expectancy of success, rooted in self-concept of ability and self-efficacy, and the value they place on achievement. Achievement goal theory separates learning goals, aimed at developing competence, from performance goals, aimed at demonstrating it, and notes that the PISA motivation scale — with its language of top grades and being “the best” — leans performance-oriented. Social cognitive theory adds that self-efficacy is built or battered by environment, through feedback, peers and teachers. Against this backdrop, the gender gap in elite aspirations becomes intriguing. Girls matched boys on general motivation yet held back on naming elite destinations. The researchers speculate that high-status university plans may feel fundamentally competitive — more about out-performing others than learning — an arena in which girls, who typically favour mastery goals, may be less comfortable staking a claim, especially with GCSEs looming and the decision feeling newly high-stakes.

Structural forces loom equally large. The stubborn socioeconomic gap in motivation — persisting within schools and at fixed levels of ability — fits a social-capital account in which families and school communities transmit norms about what achievement is worth and how attainable ambitious goals really are. Evidence cited in the paper reinforces the point: research on French teenagers found that disadvantaged pupils’ weaker grasp of educational options and their habit of underestimating their own ability explained roughly three-quarters of the socioeconomic aspirations gap there. The policy implications in England follow directly. The team argues that mentoring programmes, structured goal-setting support, clearer information about university routes and stronger partnerships between schools and families could nurture the motivation and concrete ambition the data show paying measurable dividends — particularly for disadvantaged students facing structural barriers that make sustained aspiration harder. The study was funded by the British Academy.

Honest caveats accompany the headline. Like nearly all work in this field, any causal claim rests on the assumption that, once observable characteristics are controlled, motivated and ambitious teenagers resemble their less driven peers in every relevant respect — a standard only large randomised trials could decisively test. The authors also flag a subtler problem: because prior attainment controls may themselves have been shaped by earlier aspirations, part of the true motivational effect may have been absorbed into the controls, making the 0.37 estimate conservative. And because pupils were surveyed roughly 18 months before actual applications would be filed, stated plans are not the same as submitted forms. Still, the stakes are difficult to overstate. In England, GCSE results echo through A-level choices, university admissions and lifetime earnings, and anyone who misses a pass in English or mathematics must resit. Against that backdrop, the study’s message is blunt: in the months before the exams that help shape a life, staying driven — and keeping a specific, ambitious destination firmly in view — appears to be worth roughly a third of a standard deviation. For teenagers, teachers and policymakers alike, that is a number worth taking seriously.

The survey, led by R. A. Higgoda, A. A. Alahakoon and J. Indrakumar, captured responses from 400 undergraduates spanning both the pre-clinical and clinical phases of Sri Lanka’s medical curriculum, with women making up 65.8 percent of the sample. It received ethics approval from the Faculty of Medical Sciences at Sri Jayewardenepura and was carried out in accordance with the Declaration of Helsinki, with informed consent obtained from every participant. Published online on 29 August 2026 as an open-access research article, the study set out to measure five intertwined dimensions of the chatbot phenomenon: what students actually know about these tools, how often they use them, how they judge their effectiveness and trustworthiness, which barriers stand in their way, and what they expect from AI as their careers unfold. The authors note that evidence on this question from Sri Lanka has so far been limited — a gap that matters, because most of what the medical-education community knows about chatbot adoption comes from wealthier countries whose students, infrastructure and assessment cultures differ in important ways.

Methodologically, the team relied on an instrument built for the local context rather than an imported questionnaire. The survey was piloted among Sri Lankan students before deployment, and its internal consistency — the degree to which items intended to measure the same underlying attitude produce aligned answers — was quantified with Cronbach’s alpha, a statistic running from zero to one that educational researchers conventionally treat as acceptable above 0.7 and excellent above 0.8. This instrument scored 0.856, signalling strong reliability. The researchers then moved beyond raw percentages into exploratory univariable binary and ordinal logistic regression, reporting unadjusted odds ratios (ORs) with 95 percent confidence intervals. Ordinal logistic regression is the natural tool when the outcome is a ranked scale — students rating chatbot effectiveness from poor to very effective, for example — because it models the odds of landing in a higher rather than a lower category. Crucially, the team tested the proportional-odds assumption, the premise that a predictor’s effect is identical across every cut point of the ranked outcome, and withheld common odds ratios wherever that assumption broke down — a conservative choice that trades statistical tidiness for honesty about what the data can cleanly support.

The perception data painted an overwhelmingly positive picture, edged with caution. Two-thirds of respondents — 67.5 percent — rated chatbots as effective or very effective learning aids, while 79.3 percent judged them trustworthy and 85.8 percent said the tools had improved their learning experience. Asked what the technology actually delivered, students converged on three benefits: 81.7 percent cited time savings, 81.5 percent pointed to constant availability — an always-on tutor that never sleeps, cancels a session or charges by the hour — and 71.7 percent valued personalised learning, the ability to interrogate a topic at one’s own pace, in one’s own words, as many times as needed. These advantages map directly onto the structural pressures of medical school, where dense syllabi, high-stakes multiple-choice and essay examinations, and limited access to one-on-one teaching make scalable, individualised support disproportionately valuable. For any student juggling those demands, the appeal of a tireless, on-demand digital study partner is not a gadget-lover’s whim; it is a rational response to genuine academic pressure, and the near-universal uptake figures suggest students have priced that trade-off accordingly.

Yet the warning flags flew high alongside the applause. Fully 76.2 percent of students identified inaccurate information as a major drawback — the most widely reported concern in the study — followed by reduced critical thinking at 66.9 percent and overdependence at 49.6 percent. The accuracy worry has solid technical grounding. Modern chatbots are built on large language models, neural networks trained on vast text corpora to predict plausible sequences of words. Because such models optimise for linguistic coherence rather than factual verification, they can generate confident-sounding but false statements — the phenomenon known as hallucination — including invented drug doses, distorted mechanisms of disease or fabricated references that look impeccable on the page. In a discipline where a wrong fact can eventually become a wrong prescription, that failure mode is not an inconvenience; it is a hazard. The critical-thinking concern cuts just as deep, because educators have long argued that outsourcing synthesis to a machine short-circuits the productive struggle through which students genuinely build clinical reasoning. And with almost half the cohort voicing fear of overdependence, the students themselves seem to sense that a helpful crutch can quietly harden into a cognitive cast.

The sharpest statistical divides separated pre-clinical from clinical students. In models that satisfied the proportional-odds assumption, clinical-year students — those already rotating through wards and clinics — reported significantly greater perceived effectiveness (OR 1.88, 95 percent CI 1.29–2.73, p = 0.001), a better overall learning experience (OR 1.97, 95 percent CI 1.35–2.88, p < 0.001), greater perceived accessibility (OR 2.40, 95 percent CI 1.66–3.47, p < 0.001), stronger conviction that AI proficiency will matter for their future careers (OR 1.77, 95 percent CI 1.23–2.55, p = 0.002) and, strikingly, stronger feelings that chatbot use is ethically inappropriate (OR 1.86, 95 percent CI 1.24–2.80, p = 0.003). Because these are unadjusted odds ratios, each figure reads as a simple multiplier: a clinical student carried roughly two and a half times the odds of rating accessibility higher than a pre-clinical peer. The overall pattern implies that proximity to real patients sharpens both sides of the ledger at once — deepening appreciation of what chatbots can do for a harried learner while simultaneously heightening awareness of where such tools do not belong.

That simultaneous elevation of enthusiasm and ethical alarm is the study’s subtlest finding. Clinical students were more likely than their juniors to insist that AI skills will be professionally indispensable — and more likely at the same time to call chatbot use ethically inappropriate. The two judgments are not contradictory; they are two faces of bedside realism. Senior students have seen enough of actual clinical decision-making to recognise where a language model genuinely accelerates learning, and where it intrudes on territory that belongs to supervised human judgement. Their wariness may also reflect a growing instinct for the boundaries that protect patient confidentiality, professional accountability and the integrity of medical advice. In effect, the students standing closest to medicine’s front line are sketching, unprompted, the very boundary map that professional bodies around the world are still struggling to draw.

Gender produced a second, unexpected fault line. Men were a minority of the sample — 65.8 percent of respondents were women — yet male students reported significantly greater knowledge of AI chatbots (OR 2.26, 95 percent CI 1.53–3.36, p < 0.001), more frequent use (OR 1.54, 95 percent CI 1.00–2.37, p = 0.049), greater perceived accessibility (OR 1.62, 95 percent CI 1.10–2.38, p = 0.014) and stronger feelings that such use is ethically inappropriate (OR 1.91, 95 percent CI 1.25–2.93, p = 0.003). The pairing is intriguing: male students were simultaneously more immersed in the technology and more critical of it, a combination that echoes broader findings in digital-health research where heavier exposure tends to sharpen, rather than dull, awareness of a tool’s limits. Whether the pattern reflects differing exposure to computing culture, divergent confidence in self-assessed knowledge or simply different response styles in surveys, the data cannot say — but it flags a gendered gap in technology adoption that medical educators would be unwise to ignore.

Perhaps the most quietly consequential number in the study is 58: the percentage of students who first learned about AI chatbots from their peers. Formal channels apparently trailed far behind peer word-of-mouth, meaning that the most transformative educational technology of the decade spread through Sri Lanka’s medical schools largely beneath the radar of the faculty charged with overseeing it. That informal diffusion route carries a double edge. On one side, it demonstrates genuine organic demand and the swift social contagion that viral tools enjoy among digitally native students. On the other, it means the norms governing use — when a chatbot is a legitimate study aid rather than a shortcut that hollows out learning, how rigorously an answer should be verified, what counts as academic honesty — were set by students for students, with no curricular guardrails, no ethical framing and no quality control. With assessment formats from multiple-choice questions to structured essay questions in the mix, that unregulated usage touches precisely the ground where medical faculties have the most to lose.

The authors’ prescription is unambiguous. AI chatbots are already woven into Sri Lankan medical training and are generally perceived favourably, they conclude, but that integration demands structured AI literacy education, explicit ethical guidance and faculty-led curricular integration — neither prohibition nor indifference. Translated into daily practice, that agenda means teaching students how large language models actually work and where they predictably fail; training them to verify machine-generated content against primary sources before it touches their notes, let alone their future patients; defining institutional policy on permissible use across coursework and examination preparation; and folding AI competencies into the formal curriculum instead of leaving them to peer folklore. The stakes stretch far beyond one island. Sri Lanka’s findings give the global medical-education community a rare window into how AI adoption unfolds in a health system outside the wealthy world, where student demand can run far ahead of institutional capacity to govern it. And the study’s closing message to educators everywhere lands with blunt force: your students are already using these tools. The only genuinely open question is whether you will teach them how.

Brain fog became one of the defining and most feared features of long COVID, the syndrome in which fatigue, breathlessness and cognitive complaints persist for months after the acute illness has cleared. The cognitive dimension is broad but specific: it typically involves attention, memory and executive function, the higher-order machinery the brain uses to plan, focus, juggle competing demands and restrain impulses. Earlier variants of SARS-CoV-2, including the original strain, Alpha and Delta, were repeatedly linked to persistent cognitive deficits in survivors, raising fears that infection might scar the mind as deeply as the lungs. Omicron changed the calculus. The variant generally causes less severe disease than its predecessors, yet it infected an enormous share of the world’s population, including vast numbers of older adults. That shift left scientists with an urgent, unresolved question: would a milder variant still leave a cognitive mark, and if it did, would the deficits persist, improve with time, or — the darkest possibility — keep worsening, particularly in older people whose brains carry less reserve to absorb an insult?

To find out, a research team led by Professor Yan Wan and Professor Bo Hu of the Department of Neurology at Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, followed patients hospitalized with COVID-19 during the Omicron-predominant period from December 2022 to March 2023 across nine hospitals in China. In total, 3,419 patients were enrolled in the two-year multicenter cohort study. Because gathering thousands of recovering patients for laboratory testing is impractical, the researchers assessed cognition by telephone at six months and two years after infection using two complementary, validated instruments. The Telephone Interview of Cognitive Status-40, or TICS-40, is a structured cognitive screen that probes orientation, attention, calculation, memory registration, delayed recall, learning ability and language-based reasoning, producing a numerical score that can be tracked over time. The Informant Questionnaire on Cognitive Decline in the Elderly, or IQCODE, adds a different lens by asking a relative or close contact whether the person’s memory and everyday thinking have actually deteriorated in daily life. Among the enrolled patients, 2,087 survivors completed both the six-month and two-year assessments, and their paired scores allowed the team to assign every participant to one of four cognitive trajectories: stable, reversible decline, non-progressive decline or progressive decline.

The results sketch a portrait of unexpected resilience. Two years after discharge, 67.1 percent of participants had maintained stable cognitive function throughout the entire observation period. Another 22.0 percent had experienced reversible cognitive decline: their scores sagged in the first six months and then climbed back by the two-year visit. Irreversible decline, meaning deficits that never recovered during the study window, affected 10.8 percent of survivors. That group divided into two starkly different fates. Some 7.2 percent showed non-progressive decline, a fixed deficit that neither improved nor worsened, while 3.6 percent — the smallest and most worrying slice of the cohort — followed a progressive trajectory of continuing deterioration. Put simply, nearly nine in ten hospitalized Omicron survivors were either cognitively stable or firmly on a path to recovery two years after their illness, a proportion that defies the grim expectations set by earlier variants of the virus.

The raw scores behind those trajectories reveal what recovery actually looks like in the brain. Among patients whose decline reversed, average TICS-40 scores rose by three points, from 21.00 at six months to 24.00 at two years — a meaningful climb on an instrument where each point reflects real cognitive performance. Crucially, the improvement was not diffuse but strikingly domain-specific. Patients regained ground in attention and calculation, delayed recall, memory registration, learning ability and executive function. Registration, the capacity to hold new information in mind for its first few seconds, and delayed recall, the ability to retrieve that information minutes later, are the bedrock of everyday memory: remembering a shopping list, an appointment or a familiar name. Attention and calculation, often tested by counting backward by sevens, underpin everything from managing household finances to following a conversation in a noisy room. The progressive group moved in precisely the opposite direction, shedding four points, from 17.00 to 13.00, with deterioration concentrated in attention and calculation, delayed recall and registration — the very domains that rebound in others, and exactly the skills that determine whether an older adult can continue living independently.

Just as important as who recovered, however, is who did not — and the risk factors form a coherent biological picture. Being 60 years or older carried more than a fourfold increase in the odds of progressive decline, consistent with the aging brain’s thinner cognitive reserves and decades of accumulated vascular wear. Reinfection stood out as perhaps the most consequential and modifiable red flag: each additional SARS-CoV-2 infection more than doubled the odds of a progressive course, hinting that repeated viral exposures may compound injury to the brain’s blood vessels and immune environment. Severe COVID-19, longer hospitalization, hypertension and the presence of white matter lesions — bright scars visible on brain imaging that signal damage to the delicate wiring of the nervous system and the small vessels that feed it — were each independently associated with progressive decline. Taken together, the profile suggests that patients fared worst when an aging or previously injured cerebral microvasculature met the systemic inflammation, oxygen deprivation and immune storm that accompany severe or repeated infection.

Two forces pushed the odds in the opposite direction, toward protection. Every additional year of education lowered the risk, in line with the well-established concept of cognitive reserve — the idea that education, occupational complexity and lifelong mental engagement build denser, more flexible neural networks that can absorb injury before symptoms ever surface. And every additional dose of vaccine also shifted a patient’s trajectory away from progressive decline, an observation with obvious public health weight. Vaccination is known to blunt viral replication, reduce disease severity and lower the risk of reinfection, any of which could spare the brain both direct viral insult and the collateral damage of systemic inflammation. The researchers suggest these same mechanisms may help explain why Omicron-era cohorts, in whom immunity from vaccination and prior infection is widespread, appear to fare cognitively better than patients infected by the harsher variants that dominated the pandemic’s first years.

“Our findings indicate that cognitive decline after Omicron infection is not necessarily permanent,” said Professor Wan. “The substantial recovery observed over two years is encouraging, but the persistence of deficits in a vulnerable minority shows why patients at higher risk should not be overlooked.” Her co-leader, Professor Hu, sketched the clinical translation: “Our results support a more targeted approach to long-term cognitive care after COVID-19. Identifying high-risk patients early could help guide monitoring and rehabilitation, while broader prevention measures such as vaccination may reduce the likelihood of progressive decline.” In practice, that means clinicians could weigh a short list of signals — age over 60, repeated infections, severe initial illness, hypertension, prolonged hospitalization and white matter lesions on imaging — to decide who needs closer cognitive surveillance, rather than placing every survivor under indefinite observation. The domain-specific pattern of recovery hands rehabilitation specialists a concrete target list as well: training that hones attention, calculation, memory encoding and retrieval, learning and executive planning, rather than generic mental stimulation.

The study is not without limitations, and its authors are careful to name them. Cognition was measured by telephone rather than through face-to-face neuropsychological testing, which captures a narrower slice of mental function and cannot fully assess language, visuospatial ability or mood. Telephone instruments also depend on a patient’s hearing and cooperation, and the informant questionnaire rests on the accuracy of a third party’s observations. The two-year horizon, long by the standards of long-COVID research, still cannot exclude deterioration or recovery beyond that window, and because the cohort consisted of hospitalized patients, the findings may not extend to the far larger population whose Omicron infections never reached a hospital ward. The researchers call for longer follow-up and comprehensive, in-person neuropsychological assessments to clarify how cognitive trajectories continue to evolve as the pandemic’s aftermath lengthens.

Even with those caveats, the study’s central message lands with force, and it is likely to reshape conversations in clinics and around kitchen tables. Post-Omicron cognitive decline — the frightening brain fog that patients describe — is largely reversible, and for the majority of survivors time appears to be an ally. Yet the study is equally clear that recovery is not universal. For older adults, the repeatedly infected, the severely ill, the hypertensive and those whose scans reveal damaged white matter, vigilance remains essential, because their decline can progress quietly while everyone else recovers. The strategy the authors propose is one of balance: reassure most patients that improvement is common, monitor the few who carry clinical or imaging risk factors, rehabilitate the specific cognitive functions that falter, and vaccinate to prevent the next infection from tipping the scales. As the world accumulates infections and reinfections, understanding which brains recover and which continue to slide may prove one of the pandemic’s most consequential open questions — and this two-year journey through the fog offers the first long, hopeful answer.

News Publication Date: 27-Aug-2026

Web References: https://doi.org/10.1002/nep3.70050

References: Wan, Y., & Hu, B. (2026). Post-Omicron cognitive decline is largely reversible: Two-year multicenter cohort study of domain-specific outcomes in hospitalized patients. Neuroprotection. Advance online publication. https://doi.org/10.1002/nep3.70050

For decades, the search for why girls and women remain underrepresented in science, technology, engineering and mathematics has focused overwhelmingly on the learners themselves—their confidence, their interests, their sense of whether they belong in physics or programming. A new nationwide study from Germany turns that lens around and points it squarely at the institutions. Researchers at the University of Regensburg and the University of Erlangen-Nuremberg surveyed 209 providers of out-of-school STEM education and found that the most powerful and consistent predictor of gender-inclusive practice was not how capable an organization believed it was, but how deeply promoting females in STEM was woven into its institutional self-image—an organizational identity that predicted every single marker of program success the team measured. Writing in the International Journal of STEM Education, the researchers argue that closing the gender gap may require changing what educational organizations believe about themselves, not only what girls believe about themselves.

At the individual level, the psychological machinery behind STEM engagement is well mapped. A strong science identity—seeing oneself as the kind of person associated with science—predicts students’ choices of optional science learning experiences and their intentions to pursue STEM careers. Forming that identity is harder for girls, who frequently face barriers in reconciling their self-concept with the stereotypically masculine norms of STEM fields, and the early development of STEM-specific self-efficacy is a critical driver of their academic choices. What has been almost entirely missing, the authors contend, is the equivalent analysis one level up. Out-of-school providers—workshops, maker spaces, science clubs, student research labs, summer academies and career-orientation events—offer hands-on learning that schools cannot always deliver and are uniquely positioned to spark and sustain girls’ interest during adolescence, precisely the period when STEM interest otherwise erodes without additional support. Yet whether these organizations’ own collective beliefs shape what they actually deliver had never been systematically tested.

To fill that gap, the team borrowed two constructs from organizational psychology. The first, organizational identity, traces to Albert and Whetten’s classic 1985 formulation: the central, distinctive and enduring characteristics that define who an institution is—”who we are and what we stand for.” The second, organizational efficacy, descends from Albert Bandura’s theory of collective efficacy, defined as a group’s shared belief in its conjoint capabilities to organize and execute the courses of action required to produce given levels of attainment, and asks, in effect, “what we are capable of doing.” Crucially, the researchers transferred a two-dimensional model of individual science identity to the institutional plane, distinguishing between how an organization perceives itself as an actor in the field of promoting females in STEM, and how it believes influential outsiders—participants, funding agencies, STEM teachers, policymakers and the wider STEM community—recognize it. Identity sets the strategic direction; efficacy provides the confidence to act on it.

The empirical test came in the form of a nationwide online survey conducted in November 2022 on behalf of MINTvernetzt, Germany’s federally funded competence and networking centre for extracurricular STEM education. The questionnaire, implemented in Lamapoll, went to 1,504 member organizations; 527 responded, a return rate of 35 percent, of whom 414 were actual providers of STEM programs. The final analyses drew on the 209 providers with complete responses. Organizational identity was captured with a six-item scale adapted from established science-identity research, yielding a Cronbach’s alpha of .93, while organizational efficacy was measured with three items adapted from a validated German general self-efficacy instrument, with an alpha of .85, both on four-point Likert-type formats. Five theoretically derived success indicators served as outcome variables: consideration of fundamental gender-inclusive principles such as stereotype-free concepts and appropriate female role models (alpha .72); interest in networking and professional development on gender-fair STEM promotion (alpha .89); a focus on long-term support rather than isolated events (alpha .72); integration of environmental influences such as parents, teachers, peers and media (alpha .62); and self-assessed success in getting girls excited about STEM (alpha .81).

Statistically, the team computed Pearson correlations and then a series of multiple linear regression models run in two steps, mean-centering the predictors before forming an interaction term, with all analyses performed in IBM SPSS Statistics 29. Organizational identity correlated between .25 and .52 with the five success criteria and emerged as a significant predictor of all five. Its strongest effects appeared for two practices: inclusive program planning—embedding gender-fair language, role models and critical reflection on bias into advertising, design and delivery—and interest in networking and professional development, the latter being the largest single effect observed for identity in the entire study. Providers who saw themselves, and felt seen by others, as institutions committed to advancing females in STEM were markedly more likely to seek collaboration, training and continuous improvement in that domain.

Organizational efficacy told a more complicated story. It significantly predicted four of the five criteria, with its strongest effect on self-assessed success—confident organizations rated themselves as better at exciting girls about STEM. But one result inverted expectations: efficacy negatively predicted interest in networking and professional development. The more capable providers believed themselves to be, the less they wanted external exchange. The authors interpret this as a compensation effect: organizations that doubt their collective capabilities seek outside support to close perceived gaps, while highly efficacious organizations lean on internal resources and may underestimate both the benefit of exchange for themselves and their own potential role as mentors to struggling peers. Notably, identity’s pull toward networking remained high regardless of efficacy level, hinting that a strong institutional identity fuels a genuine commitment to learning rather than mere deficit-filling.

Because programs differ enormously in structure and theme, the second regression step added covariates: whether more than half of a provider’s offerings targeted girls only (12 percent of cases), four target-group categories spanning early childhood through educational professionals, four format types from hands-on labs to mentoring, and seven disciplinary fields. Explained variance rose from between 8 and 33 percent in the first step to between 27 and 41 percent with controls, and the effects of identity and efficacy remained robust, though slightly weaker. One telling shift occurred: after accounting for program characteristics, efficacy—rather than identity—became the significant predictor of integrating environmental influences, suggesting that institutional confidence is what enables organizations to weave parents, teachers and peers into their programs once structural heterogeneity is considered. The interaction between identity and efficacy, only marginally significant before controls, vanished entirely. Among the covariates, informatics programs showed the strongest networking interest and long-term focus, technology-oriented programs were marginally less likely to emphasize inclusive practice, and providers concentrated on vocational or higher education reported lower networking interest and lower self-assessed success.

The outcome measures themselves were grounded in a substantial evidence base on what actually works for girls in STEM. Prior studies show that recruitment advertisements using only masculine linguistic forms and male-typed imagery reduce women’s self-ascribed fit, interest and application intentions, while female role-model interventions—in which women employed in STEM visit schools—reduce gender stereotypes about mathematical ability and increase enjoyment, perceived importance and career aspirations. Duration matters as well: follow-up research shows motivational gains from short science interventions can decay within six months, and a systematic review of 32 interventions for secondary-school girls found roughly two-thirds lasted between twenty minutes and two weeks. A review of 165 studies on female underrepresentation found that 61 percent identified environmental factors—parental involvement, peer pressure, teachers’ gender bias—as contributors to low participation. In the authors’ framing, identity orients the organization toward these evidence-based practices; efficacy determines whether members believe they can actually execute them.

The practical implications operate on three levels. Organizationally, providers can embed gender equity in their mission statements, internal discussions and outreach materials, and use peer-learning formats, training workshops and professional networks to build collective confidence in enacting gender-inclusive practice. Programmatically, the evidence supports deliberate female role models, mentoring structures, hands-on and real-world content, long-term support structures aligned with short-term activities, and the active involvement of parents, teachers and peers in girls’ educational pathways. At the network level, the results carry a pointed warning for funders and consortia: networking offers will naturally attract lower-efficacy organizations—the ones who need them most—while high-efficacy providers must be actively recruited with different strategies, positioned not as beneficiaries but as contributors whose experience can lift an entire community of practice.

The authors caution that each organization was represented by a single respondent, that success was self-assessed rather than independently verified, and that the cross-sectional design cannot capture how identity and efficacy develop over time; several adapted scales also showed only moderate internal consistency. Future work, they suggest, should survey multiple representatives per institution, complement subjective ratings with objective measures such as website content analyses and participant feedback, and test interventions designed to strengthen organizational identity and efficacy directly. But the central message stands: the battle for gender equity in STEM is not fought only in the minds of girls deciding who they want to become. It is fought, just as decisively, in the collective self-understanding of the organizations that invite them in—and when an institution genuinely believes that advancing women in science is part of who it is, gender-inclusive practice follows.

Hepatitis C virus, or HCV, is a single-stranded RNA virus of the Flaviviridae family that spreads through blood-to-blood contact. In adults, most acute infections go unnoticed, and the majority progress to chronic disease that can smoulder silently for decades, progressively scarring the liver toward cirrhosis and, in some patients, hepatocellular carcinoma. For children, the dominant route of infection is not needles or contaminated equipment but childbirth: the virus can cross the placenta or infect the infant during delivery, a phenomenon known as vertical or perinatal transmission. Population studies generally place the vertical transmission risk at around five percent, which is precisely why international guidelines urge systematic follow-up of babies born to viremic mothers. What makes the Vienna findings especially consequential is that HCV is now among the most treatable viral diseases known to medicine. Direct-acting antivirals — including the sofosbuvir/velpatasvir and glecaprevir/pibrentasvir regimens referenced in the study — cure more than 95 percent of patients in eight to twelve weeks, and are approved for children as young as three.

The study, published open-access in the International Journal for Equity in Health, exploited a clinical clue that most infection-surveillance efforts ignore: neonatal abstinence syndrome, or NAS. NAS is a withdrawal disorder seen in newborns exposed in utero to opioids and other substances; affected infants show tremors, irritability, feeding problems and, in severe cases, seizures, typically within days of birth. Because the syndrome only develops when the mother used substances during pregnancy, a NAS admission functions as a marker for a population at heightened risk of blood-borne infections. The Vienna cohort, assembled by a team led by Caroline Schwarz and corresponding author Michael Gschwantler of Klinik Ottakring together with pediatric colleagues, comprised 243 children — including five sets of twins — of whom 50.2 percent were male, with a median gestational age of 39 weeks and a caesarean delivery rate of 45.5 percent. Their 238 mothers had a median age of 30 years, and 88.1 percent were on opioid agonist therapy, the maintenance treatment with methadone or buprenorphine that stabilizes substance use during pregnancy.

To map the virologic landscape, the researchers reconstructed each mother’s testing history using the two-step logic that defines HCV diagnostics. First-line serology detects anti-HCV antibodies, signalling exposure to the virus at some point in the past. Because antibodies can persist even after the immune system clears the infection, a positive screen must be followed by a polymerase chain reaction test for HCV-RNA, which detects active viral replication. Of the 238 mothers, 227 — 95.4 percent — received some form of HCV testing, a strikingly high figure that reflects conscientious clinical practice at the centre. Among the 218 women who underwent serology, 116 — 53.2 percent — were anti-HCV positive, meaning more than half had been exposed to the virus. Of these, 110 proceeded to confirmatory HCV-RNA testing, and nine additional women were tested directly by PCR. In total, 53 mothers were identified as HCV-viremic. By comparison, 10 of 211 mothers tested, or 4.7 percent, were positive for HIV antigens, and 4 of 237, or 1.7 percent, carried hepatitis B surface antigen.

The infection’s footprint on the next generation was sobering. The 53 viremic mothers had borne 55 children. Only 42 of them — 76.4 percent — ever underwent HCV-RNA testing, the only reliable method in infancy because maternally derived antibodies can circulate in a baby’s blood for up to 18 months, making antibody screens uninterpretable in the first year of life. Four of the 42 tested children were viremic, yielding an estimated vertical transmission rate of 9.5 percent — nearly double the background risk typically cited in the literature. The gaps in follow-up harboured at least one more hidden case: among seven children born to antibody-positive mothers who never received RNA testing, one had been repeatedly HCV-viremic since birth. That pattern, the authors note, implies the mother’s active infection had never been diagnosed during pregnancy and the virus had been transmitted perinatally — a child carrying a diagnosable, curable infection whom the system’s diagnostic net had quietly failed to catch.

The arithmetic of what fell through the cracks is stark. Seventeen of the 238 mothers — 7.1 percent — did not receive adequate HCV testing, a figure that combines eleven women who were never screened at all with six antibody-positive mothers who never underwent confirmatory RNA testing, leaving it impossible to distinguish a cleared past infection from an active, transmissible one. On the pediatric side, 13 of the 55 virus-exposed children — 23.6 percent — never had their HCV status assessed. Each gap represents a missed opportunity for a cure that, as of 2026, takes as little as two months. The contrast with other blood-borne viruses is particularly pointed: HIV and hepatitis B screening are built into Austria’s routine prenatal care program, documented in the “parents-child health passport” that accompanies every pregnancy, while HCV — which in this population infected more than one in five women, compared with 4.7 percent for HIV and 1.7 percent for hepatitis B — is not.

The paper’s central recommendation is a technical fix with outsized consequences: HCV reflex testing in routine prenatal care. In a reflex-testing scheme, a positive antibody screen automatically triggers RNA confirmation on the same blood sample, without a new appointment, a second blood draw, or a clinician’s discretionary order. The approach is engineered to eliminate precisely the cascade failures the Vienna data lay bare — the antibody-positive mothers lost between screening steps, and the patients who never return for the follow-up visit that a two-step system implicitly demands. For a population contending with stigma, unstable housing and fragmented access to care, that simplification is not a convenience but the difference between diagnosis and blindness. Reflex testing also fits the laboratory economics: the blood is already drawn, the sample is already in the lab, and the cost of an added molecular test is small measured against the price of decades of undiagnosed chronic hepatitis, liver monitoring, and late-stage disease.

The timing matters against a global backdrop. The World Health Organization has set 2030 as the deadline for eliminating viral hepatitis as a public health threat, and perinatal transmission sits squarely on the critical path: every undiagnosed viremic mother is a potential new chronic infection in the next generation, seeding chains of transmission that will not surface clinically for decades. The United States Centers for Disease Control and Prevention has recommended universal HCV screening for every pregnancy since 2020, and European and international clinical guidelines call for screening pregnant women and evaluating infants born to viremic mothers. Austria’s prenatal program, the authors argue, has not kept pace, leaving a population that is doubly marginalized — by substance use and by the circumstances surrounding it — outside the screening architecture that is supposed to protect mothers and their children.

The study’s design carries caveats that the authors themselves acknowledge. It is retrospective, drawing on records from a single centre, and some data were incomplete — including the seven untested children of antibody-positive mothers whose status remains unknown, one of whom turned out to be infected. The true vertical transmission rate may therefore be somewhat higher or lower than the 9.5 percent estimate, and maternal viremia may be undercounted. Yet the sheer size of the cohort — 243 children and 238 mothers assembled over eleven consecutive years — makes it one of the most detailed portraits available of hepatitis C within a European perinatal population affected by substance use, and the pattern it reveals is unlikely to be unique to Vienna. The authors frame the findings as evidence that these families constitute an underserved, at-risk population whose infections are neither systematically sought nor systematically treated.

For the four children confirmed to be viremic, and the probable fifth whose mother’s infection was never diagnosed, the clock is now the decisive variable. Chronic hepatitis C acquired at birth is typically silent through childhood, quietly laying down fibrous scar tissue in the liver while the patient grows up unaware — which is exactly why early identification, followed by antiviral cure from age three onward, can compress decades of risk into a non-event. The Vienna study’s message to policymakers is unusually concrete: clinicians are already screening nearly all of these mothers, the diagnostic tools already exist, the cures already work, and what remains is a line in the national prenatal care program that has not yet been written. Until hepatitis C joins HIV and hepatitis B in that program, the researchers conclude, a curable infection will keep moving silently from one generation to the next.

The intellectual urgency behind the proposal rests on a growing body of evidence documenting childhood’s quiet retreat from the outdoors. Research cited by the team indicates that children today spend up to 90 percent of their time indoors and engage extensively with digital devices, a displacement driven by urbanisation and technological immersion and compounded by the confinement of the COVID-19 pandemic, which measurably weakened children’s connection to nature. The consequences extend beyond nostalgia: diminished contact with natural environments limits opportunities for exploration, environmental empathy and essential life skills, with long-term implications for well-being and pro-environmental attitudes. Richard Louv’s influential notion of “nature deficit disorder” crystallised these concerns, while Louise Chawla’s research has shown that early experiences in nature play a decisive role in shaping ethical values, emotional bonds with the living world and enduring pro-environmental orientations. Nature engagement, the literature suggests, fosters not only environmental stewardship but spiritual development expressed through wonder, respect for life and awareness of interdependence—the emotional and spiritual architecture of who children become.

Central to the framework is a deliberately secular, rigorously operationalised definition of spirituality. The authors are explicit that in early childhood spirituality is not confined to religious instruction; it encompasses children’s sense of wonder, empathy, meaning-making and connection with the natural world. Drawing on Brenda Hyde’s account of young children’s innate capacity to relate to something greater than themselves, and on Jennifer Mata-McMahon’s characterisation of spirituality as a universal developmental domain transcending cultural and religious boundaries, the team separates the construct from its nearest conceptual neighbours. Moral education, they note, targets ethical values, moral reasoning and behaviour; socio-emotional learning emphasises emotional regulation, self-awareness and interpersonal competence. Spirituality, by contrast, concerns a child’s developing sense of awe, connectedness and meaning—a distinct yet complementary strand of holistic development. The distinction carries practical weight in Malaysia’s multicultural, multi-religious society, where any classroom initiative touching the spiritual must avoid doctrine while remaining educationally substantive. The NLET framework is engineered to navigate that boundary, omitting religious doctrines, theology and rituals entirely and anchoring every activity in the religiously neutral space of the natural environment, allowing adaptation across national, vernacular and private preschool settings.

The proposal emerges from a policy landscape that is ambitious on paper yet uneven in practice. Malaysia’s National Preschool Curriculum Standard, known as KSPK, already recognises spiritual values through its emphasis on moral and religious development, and the revised curriculum, KP2027, strengthens that direction by foregrounding holistic well-being, values formation, sustainability awareness and meaningful learning connected to the environment. Yet classroom implementation frequently remains limited and ritualistic—centred on routine activities such as singing songs, with minimal opportunity for reflection or nature-based inquiry—because teachers often lack the pedagogical knowledge and resources to translate spiritual goals into daily practice. The result, the authors contend, is that spirituality functions as a hidden or passive curriculum rather than an intentionally nurtured aspect of child development. Their answer is a structured teaching module that equips preschool teachers with practical strategies for integrating nature-loving education into everyday routines, supported by the love-based teacher–child relationships that researcher Yen-Hui Shih identifies as the baseline from which children confidently explore their inner worlds. Such an approach, they argue, could simultaneously advance children’s spiritual growth, moral and ethical development and environmental responsibility.

Conceptually, the module stands on four theoretical pillars woven into a single pedagogical architecture. Constructivist learning theory—from Jean Piaget’s portrayal of children as active builders of knowledge to Lev Vygotsky’s emphasis on social interaction—frames nature as a spiritually meaningful learning context in which children construct inner meaning through direct encounters, reflective dialogue and symbolic expression, extending constructivism beyond cognition into value formation. The Biophilia Hypothesis, Edward O. Wilson’s proposition that humans possess an innate affiliation with other living things, supplies the rationale for nature as the primary learning medium, with the authors positioning biophilic engagement as a foundational mechanism for care, empathy, responsibility and reverence for life rather than a mere affective response. James Fowler’s Faith Development Theory legitimises the imaginative, intuitive meaning-making of preschoolers—the intuitive–projective stage—as an authentic, observable form of spiritual expression rather than immature cognition, treating children’s storytelling and symbolic interpretations of nature as genuine spiritual processes. Finally, Lisa Miller’s Spiritual Child Theory conceptualises spirituality as an innate but nurture-able human capacity that can be strengthened through relationships, reflective practices and engagement with nature. Together, the authors argue, these frameworks converge in structured nature-loving pedagogy.

What elevates the paper beyond purely theoretical territory is its insistence that spirituality be made observable and measurable. The framework operationalises children’s spirituality through three interrelated dimensions: awe or wonder, defined as children’s spontaneous experiences of fascination and curiosity toward nature; connectedness, their emotional and relational awareness of themselves, others and the natural world; and meaning-making, their capacity to construct, interpret and express personal understanding through meaningful encounters with nature. Each dimension is captured through a multi-method triangulation strategy designed to minimise subjectivity. Structured classroom observations will employ standardised behavioural checklists and anecdotal record templates, coding awe through sustained attention, spontaneous verbal expressions of curiosity and focused exploration, and connectedness through documented relational behaviours such as gentle handling of living things, verbal expressions of care and collaborative helping actions. Child drawings produced after nature-based activities will be analysed against thematic criteria for symbolic representations of ecological relationships and visual emphasis on environmental features of personal significance. Brief, semi-structured, child-friendly interviews, conducted immediately after the drawing sessions in familiar classroom surroundings, will elicit children’s narratives through open-ended prompts standardised to reduce the risk of leading questions. Converging evidence across the three streams, the authors argue, reduces observer bias and strengthens the trustworthiness of interpretations of young children’s spiritual expressions.

The research design follows Design and Development Research, or DDR, a methodology suited to systematically designing, developing and evaluating educational products grounded in both theory and real-world practice. The study is planned in three sequential phases. Phase I, a needs analysis, will survey a purposive sample of 100 preschool teachers and conduct semi-structured interviews with seven parents to map their understanding, challenges and instructional needs around nature-based spiritual learning, while curriculum documents including KSPK and KP2027 are reviewed to determine how thoroughly spirituality, character education and nature engagement are addressed in policy. Phase II moves to design and development, in which learning objectives, module components and nature-based activities are constructed and subjected to formal expert validation. Here the team deploys the Fuzzy Delphi Method, a consensus-building technique that converts qualitative expert judgment into quantified agreement, with a panel of 11 specialists in early childhood education, nature-loving education and children’s spirituality. The method will establish mathematical consensus on essential module elements, pedagogical strategies and the behavioural items within the measurement tools, ensuring indicators that are theoretically sound, culturally sensitive and developmentally appropriate for the Malaysian preschool context.

Phase III will carry the prototype into authentic preschool classrooms. Before implementation, the accompanying assessment instruments—observation protocols, behavioural checklists, drawing tasks and interview guides—will undergo expert validation for content validity and developmental appropriateness. Selected teachers will then implement the module over a specified period, with data gathered through teacher evaluation surveys gauging usability, practicality and appropriateness, alongside structured classroom observations, children’s drawings and brief interviews examining the three spiritual indicators during nature-based activities. Importantly, the authors frame this phase as formative evaluation: the goal is to collect feedback for further refinement rather than to establish summative effectiveness, a deliberately cautious stance reflecting the framework’s provisional status. Each phase, in the logic of the design, generates the evidence needed to justify the next, creating a structured pathway from abstract theory to validated instructional product.

The implications stretch well beyond Malaysian preschools. Nationally, the NLET module aligns with KP2027’s pillars of holistic development, values formation and sustainability, offering teachers a structured yet flexible framework for translating abstract curricular aspirations into observable classroom practice and repositioning educators as facilitators of reflective, relational learning rather than transmitters of prescribed content. Globally, the authors connect the framework to United Nations Sustainable Development Goal 4.7, which calls for education promoting sustainable development, environmental citizenship and global values, arguing that long-term environmental stewardship is rooted in a child’s earliest relationship with the natural world. The framework also embraces a relational, ecological worldview—evoked metaphorically as a “one world family”—in which children perceive themselves not as dominators of nature but as interdependent members of a shared moral and ecological community. Whether deliberate pedagogy can reliably cultivate awe, compassion, gratitude and connectedness at scale remains, the researchers acknowledge, an open empirical question, and they present their model as provisional pending systematic investigation. But by articulating a coherent conceptual and methodological foundation—and insisting that even something as intangible as a child’s sense of wonder can be defined, observed and nurtured without doctrine—the study offers educators a credible starting point for reintroducing the natural world, and the meaning it affords, into the earliest years of education.

Run every four years, TIMSS is among the most consequential measuring instruments in education, and until recently every participant in every country sat the same kind of test: printed booklets, pencils, and hand-processed answer sheets. That changed in 2019, when about half of the participating countries migrated to a digital version delivered on computers, laptops, or tablets, with most of the remaining systems completing the migration in the 2023 cycle. The motivations were compelling. Digital delivery opens the door to interactive items and Problem-Solving and Inquiry (PSI) tasks that paper cannot replicate, promises more colorful and dynamic formats that hold children’s attention, reduces the volume of responses requiring human scoring, and eliminates the laborious data entry of paper materials. Crucially, the digital assessment was engineered to be comparable with its paper predecessors, so that the long-term trend lines on which education policymakers depend would remain interpretable across the transition.

Engineering intent, however, is not the same as psychometric certainty. Any difference in performance caused not by what students know but by how they are asked to show it—a phenomenon measurement specialists call a mode effect—can silently corrupt a trend line. To guard against this, every country that switched to digital in 2019 also administered a parallel paper-based “bridge” study with a nationally representative sample of students. The international evidence was sobering. A pilot item-equivalence study conducted before the 2019 cycle, together with the subsequent technical report, concluded that a statistical adjustment built into the international scaling procedures could preserve the linkage between formats—yet even after that adjustment, ten countries, among them Austria, Canada, Georgia, Hong Kong, Italy, Lithuania, the Netherlands, Portugal, Qatar, and Sweden, still showed residual country-level mode effects in at least one subject or grade. By 2023 the bridge study was no longer internationally required, because paper-to-digital links had already been established. Only four jurisdictions voluntarily ran one: Ireland, Azerbaijan, Bahrain, and Flanders.

The new report, compiled by Aidan Clerkin of the Educational Research Centre in Dublin together with colleagues in Belgium, Bahrain, and Azerbaijan, collates all four national comparison studies into a single picture. The design is elegantly simple. In each country, students were randomly assigned at the school or class level to sit either the digital main study or the paper bridge study, so the two groups should, on average, possess identical underlying proficiency; any systematic difference in scores can therefore be attributed to the mode of administration itself. Flanders ran its comparison at fourth grade only; Ireland and Bahrain covered both grades; and Azerbaijan’s bridge study covered eighth grade alone. The digital samples were nationally representative and large, ranging from 4,336 to 6,174 students per grade, while the paper samples ranged from 1,128 to 1,779 students per grade, selected through equivalent procedures and completing an abridged assessment built solely from trend items carried forward from 2019. Achievement was estimated with five plausible values per subject and analyzed using the IEA’s IDB Analyzer, allowing comparisons of overall means, gender differences, content and cognitive domains, and access to devices at home.

The headline verdict is unambiguous: at the level of overall achievement, the change of medium left scores untouched, in mathematics and science alike, in all four countries. A faint directional pattern was visible—differences tended to favor the digital test among fourth graders, while eighth-grade differences tended to favor paper in Ireland and Bahrain—but almost none of these gaps proved statistically significant once measurement error was taken into account. Azerbaijan broke the pattern in the opposite direction, with its eighth graders tending to score higher on screens. In practical terms, the authors conclude, the trend measurements linking 2019 to 2023 in these systems can be treated as valid: any rise or fall in the 2023 results reflects real changes in what children know, not an artefact of the keyboard.

The exception came from Azerbaijan, and it is the study’s most eye-catching finding. There, boys who took the science assessment digitally scored 408 points on the international reporting scale, while boys who took it on paper scored 382—a 26-point gap equivalent to one quarter of a standard deviation, far from trivial by the standards of international assessment. No corresponding difference appeared among girls, and the researchers found no gender-related mode effects anywhere else: in Ireland, Flanders, and Bahrain, the scores of boys and girls were statistically indistinguishable across formats. That reassurance matters now more than usual, because the TIMSS 2023 cycle reported widening gender gaps in several countries, and researchers trying to establish whether girls are genuinely losing ground need to know that the test itself is not distorting the picture. In Azerbaijan, however, the mode of delivery appears to interact with gender in a way that demands explanation.

The researchers also dissected the results by content domain within each subject, and by cognitive domain—the progression from Knowing, through Applying, to Reasoning. Here the findings were sparse and unsystematic in three countries. In Bahrain, eighth graders tested on paper outperformed their digital peers in the mathematics domain Data and Probability; in Ireland, paper-tested eighth graders led in Number and in Physics; and Bahrain recorded a solitary difference on the eighth-grade mathematics domain of Applying. Flanders showed no significant differences in either subject, and neither did any fourth-grade comparison anywhere. Azerbaijan, once again, stood apart: students tested digitally scored significantly higher in three of the four science domains—Physics, Chemistry, and Earth Science—as well as in Geometry and Measurement. The cognitive asymmetry was more telling still: on the higher-order domains of Applying and Reasoning, Azerbaijani students performed significantly better on the digital version of both the mathematics and the science assessments, while scores on the recall-oriented Knowing domain differed little—precisely the signature one would expect if screens somehow helped students think through complex, multi-step items.

The strongest pattern in the entire study, however, had nothing to do with test mode and everything to do with what children have at home. Across Flanders, Bahrain, and Azerbaijan, students with access to a computer or tablet at home outperformed those without—most sharply when they took the test on a screen. In Flanders, digitally tested fourth graders with home computer access scored 526 in mathematics and 494 in science, against 482 and 433 for peers without such access; in Bahrain the corresponding gaps were 468 versus 450 in mathematics and 481 versus 463 in science. Azerbaijani eighth graders showed the same asymmetry on the digital assessment, while in Ireland differences were small and non-significant at both grades and in both modes. The exception that illuminates the rule came from Bahrain’s eighth grade, where computer access predicted scores even among paper testers—hinting that there, a home computer functions partly as a proxy for socioeconomic status and broader learning resources rather than as digital familiarity alone.

The authors are candid about the limits of their evidence. With dozens of comparisons, some significant results are expected by chance alone—so-called Type I errors, or false positives—and the domain-level and device analyses were explicitly exploratory, since the bridge studies were designed primarily to test overall achievement. There is also a deeper structural asymmetry: the plausible values used to score the paper samples rest exclusively on trend items carried over from 2019, whereas the digital samples’ scores also absorb responses to PSI tasks and other innovative item types available only on screen. The two estimates, in other words, are not built from identical information. Nor is equivalence guaranteed elsewhere in the assessment world: sizeable mode effects have been documented in PISA 2015, in PISA reading short-text responses, and in PIRLS 2021, and Irish test developers once recorded a “substantial mode effect” among second graders on a national test—so pronounced that only a paper version was released at that grade level.

That is why the report singles out Azerbaijan for urgent follow-up: verifying that the digital and paper samples were truly equivalent in sociodemographic terms, drilling down to item-level comparisons—particularly on digital tools such as the on-screen ruler—and even cognitive interviews to understand how Azerbaijani children engage with low-stakes tests in each format. For the other three countries, the message is quiet confidence. Policymakers can now read changes between the 2019 and 2023 cycles as substantive signals, whether they stem from pandemic aftershocks, curriculum reforms, or shifting home and school contexts, rather than as artefacts of the switch from pencil to pixel. Countries that transitioned in 2023 without running their own bridge studies enjoy no such clarity, and may struggle to disentangle genuine changes in proficiency from procedural ones. In a world where a few points on an international test can trigger budget battles and political storms, that distinction is the difference between responding to a real crisis and chasing a ghost in the machine.