The premise stands on the extraordinary progress of LLMs, sophisticated AI systems trained on vast amounts of textual data from the web, books, and academic literature. These models, initially designed for natural language processing tasks like translation or summarization, have evolved into remarkably nuanced predictors of complex human traits. Wolfram’s study methodically benchmarks the predictive power of LLMs against genomic data and expert human evaluations, uncovering insights that could redefine assessment metrics in psychology and education.

Genomics, which has long been heralded as a critical avenue to understanding individual differences in cognition, relies on identifying specific gene variants linked to intelligence and learning ability. While powerful, genomic predictors often require extensive datasets, are prone to ethical controversies, and frequently struggle to capture the environmental and sociocultural components influencing cognitive development. Wolfram’s research posits that LLMs, grounded in linguistic and contextual world knowledge, offer a complementary—and in some cases superior—approach.

The methodology deployed in the study involves applying state-of-the-art LLMs to naturally occurring textual outputs associated with individuals, such as essays, social media posts, and academic writing. By analyzing syntactic complexity, semantic richness, and thematic coherence, the models generate cognitive profiles without explicit phenotype data. These AI-derived predictions are then directly compared to polygenic scores derived from genome-wide association studies (GWAS) and to expert assessments conducted by seasoned psychologists and educators.

Notably, the results demonstrate that LLMs achieve predictive accuracy on par with genomic methods, a finding that challenges the long-held assumption that genetic markers remain the gold standard for identifying cognitive aptitude. The AI’s ability to contextualize language within broader narratives and cultural frameworks allows it to capture subtle cognitive and educational signals that genetic data may overlook. Moreover, when combined with expert assessments, LLM-generated predictions enhance overall accuracy, indicating a complementary relationship rather than a competitive one.

The implications of this research extend beyond academic curiosity into practical applications. In education, for instance, AI-powered assessments could provide real-time, scalable, and non-invasive evaluations of student learning styles, comprehension, and potential cognitive challenges, facilitating personalized learning experiences at an unprecedented scale. This prospect could democratize access to educational resources, particularly in under-resourced settings where expert evaluators are scarce.

Furthermore, the study addresses concerns related to privacy and data security by emphasizing that LLM predictions can be made from publicly available or consented textual data without the need for genetic sampling, which is costlier and more intrusive. This advantage positions large language models as ethically favorable tools, provided that transparency and consent are rigorously maintained in data collection practices.

Critically, Wolfram also explores the limitations inherent in relying solely on AI models. While LLMs demonstrate remarkable capacity, they are sensitive to biases encoded in training data, including cultural, socioeconomic, and linguistic biases. These factors could skew predictive outcomes if not carefully mitigated through refined model training and validation techniques. The study calls for an interdisciplinary approach where AI specialists collaborate closely with cognitive scientists and ethicists to ensure equitable and responsible deployment.

In the realm of cognitive science, the ability to quantify mental constructs such as working memory, fluid intelligence, and verbal reasoning through language-based AI tools opens new avenues for research. Traditionally challenging to measure with precision, these dimensions are accessible by LLMs analyzing discourse patterns and conceptual complexity. This reframing could accelerate hypothesis testing and theory development, transforming the way intelligence is operationalized and measured.

Moreover, the predictive use of large language models may influence neuropsychological assessments, psychiatric evaluations, and even workplace talent identification. Early indications suggest that nuanced verbal outputs captured by LLMs correlate with cognitive function and educational attainment, offering auxiliary data points that can supplement clinical and administrative decision-making processes. The integration of these models could streamline assessments and offer continuous monitoring capabilities unobtainable by conventional methods.

Wolfram’s study further engages with the ethical dimensions of employing AI in predictive psychology. The paper underscores the necessity of safeguarding individuals from potential misuse of predictive data, highlighting risks such as stigmatization, discrimination, and privacy breaches. It advocates for stringent regulatory frameworks and continuous monitoring to balance innovation with respect for human rights.

Looking ahead, the research hints at the prospect of synergistic models that integrate genomic, linguistic, and expert inputs, leveraging the strengths of each modality. Such hybrid approaches promise more comprehensive and nuanced forecasts of cognitive ability and educational outcomes, establishing a new frontier in predictive accuracy.

Importantly, this emerging AI-driven paradigm democratizes knowledge by enabling non-invasive, cost-effective, and scalable approaches to measure cognition and learning. It offers a potent tool to bridge disparities in educational achievement and cognitive science research infrastructure worldwide, potentially transforming policy development and individualized support services.

In summary, the study by Wolfram marks a watershed moment in cognitive and educational assessment, revealing that large language models offer a predictive capacity that challenges long-established methodologies. By harnessing the intrinsic link between language and cognition, these AI systems stand poised to revolutionize our understanding of the human mind, with profound implications for education, psychology, and beyond.

As large language models continue to evolve, their integration into scientific inquiry and practical applications must be guided by ethical considerations, interdisciplinary collaboration, and rigorous validation. The promise of AI as a complementary or even superior predictor of cognition beckons a future where technology and human expertise converge to unlock unprecedented insights into the fabric of intelligence and learning.

Subject of Research: Cognitive and educational outcome prediction using large language models compared to genomics and expert assessment.

Article Title: Large language models predict cognition and education close to or better than genomics or expert assessment.

Article References:
Wolfram, T. Large language models predict cognition and education close to or better than genomics or expert assessment. Commun Psychol 3, 95 (2025). https://doi.org/10.1038/s44271-025-00274-x

Image Credits: AI Generated

The study, published in the esteemed journal Genetics in Medicine, offers the first data-driven approach to harmonizing the selection of genes for NBSeq programs worldwide. The researchers harnessed advanced machine learning techniques to distill complex patterns from an extensive dataset comprising thousands of genes selected by diverse screening programs. This methodological innovation presents a transformative tool capable of guiding policymakers and clinicians through the multifaceted decision-making process inherent in genomic newborn screening, ensuring that gene inclusion reflects not only scientific rigor but also practical considerations relevant to public health.

Central to the research is the observation that despite 27 NBSeq programs collectively analyzing 4,390 unique genes, only a small subset — precisely 74 genes, or about 1.7% — appear consistently in over 80% of these initiatives. This stark disparity reveals the heterogeneity in how different programs define clinical utility, evidence strength, and public health value when curating their gene panels. Such inconsistency poses a significant barrier to creating unified standards that could facilitate broader adoption, equitable access, and interpretable results for families worldwide.

The study identifies key predictors that strongly influence whether a gene is included in NBSeq panels. Among these, the presence of a gene-associated condition on the U.S. Recommended Uniform Screening Panel (RUSP) emerged as a top determinant. This reflects the weight of preexisting public health frameworks that prioritize conditions with established newborn screening protocols. Moreover, the availability of robust natural history data—a comprehensive understanding of the disease trajectory in the absence of intervention—is a crucial factor. Equally important is the demonstration of effective treatments, which validates the clinical actionability of detecting the gene variant in newborns.

To translate these insights into a practical tool for global NBSeq governance, the research team developed a sophisticated machine learning model incorporating 13 distinct predictors encompassing clinical, epidemiological, and therapeutic evidence metrics. This model achieved high accuracy in recreating gene selection patterns across existing programs, suggesting its capacity to reliably predict gene candidacy for inclusion. Importantly, the model’s adaptability permits continuous refinement as new genetic discoveries, treatment modalities, and regional health priorities emerge, fostering dynamic and evidence-responsive screening frameworks.

The implications of this research resonate profoundly within the precision medicine and public health communities. By providing a transparent and data-driven gene prioritization strategy, this tool could serve as a foundation for harmonizing NBSeq efforts internationally. Such harmonization is pivotal not only for scientific consistency but also for addressing ethical, legal, and social issues surrounding the return of genomic information in newborns, such as equity of access, informed consent, and the management of uncertain findings.

Moreover, the involvement of the International Consortium of Newborn Sequencing (ICoNS)—founded by leading figures in the field including Dr. Robert C. Green of Mass General Brigham and Dr. David Bick of Genomics England—anchors this publication in a global collaborative framework. ICoNS embodies the international effort to consolidate expertise, data, and policy perspectives to navigate the complex landscape of genomic newborn screening. The consortium’s commitment reflects the growing recognition that tackling genetic disorders at birth requires coordinated action transcending borders.

The use of machine learning in this context exemplifies a broader trend in biomedical research, leveraging computational intelligence to manage large-scale genomics data and extract actionable insights. Traditional gene selection processes for newborn screening have often relied on expert panels and consensus, which, while invaluable, may be limited by subjective biases and knowledge gaps. The data-driven approach demonstrated here underscores how quantitative methods can augment human expertise, enabling more transparent, scalable, and reproducible decision-making.

Importantly, this new model also allows for regional customization, acknowledging that genetic disorder prevalence, healthcare infrastructure, and treatment availability vary globally. This flexibility ensures that NBSeq programs are not only scientifically grounded but also contextually appropriate, thereby maximizing their clinical relevance and cost-effectiveness. Policymakers and healthcare providers can thus tailor screening panels to optimally serve their populations while maintaining core standards informed by robust evidence.

The study’s findings also highlight the challenges ahead. The vast majority of genes included in NBSeq programs lack consensus inclusion, reflecting ongoing uncertainty about their clinical significance, returns on investment, and ethical considerations related to possible overdiagnosis or incidental findings. Efforts to expand newborn genomic screening must therefore proceed cautiously, balancing innovation with responsible stewardship to protect the best interests of infants and their families.

Furthermore, as treatments for genetic disorders proliferate—driven by advances in gene therapy, enzyme replacement, and personalized medicine—the pressure to incorporate newly actionable genes into NBSeq panels will grow. The proposed machine learning framework equips stakeholders with a scalable mechanism to evaluate emerging candidates swiftly and systematically, avoiding fragmented rollouts and ensuring equitable access to cutting-edge interventions.

In summary, the Mass General Brigham-led study marks a milestone in advancing genomic newborn screening from disparate pilot projects toward a harmonized, evidence-based global initiative. By leveraging computational modeling and international collaboration, this work lays the foundation for a future where newborn screening programs are consistent, scientifically validated, and responsive to evolving medical knowledge. Such progress promises to enhance the early detection and treatment of genetic disorders, ultimately improving health outcomes from the very start of life.

Subject of Research: Genetic disorders consideration and gene selection for genomic newborn screening programs using machine learning.

Article Title: Data-driven consideration of genetic disorders for global genomic newborn screening programs

News Publication Date: 9-May-2025

Web References:

• Genetics in Medicine Article
• BabySeq Project
• Mass General Brigham Press Releases
• Mass General Brigham DNA Sequencing Study

References:
Minten T, et al. “Data-driven consideration of genetic disorders for global genomic newborn screening programs” Genetics in Medicine. DOI: 10.1016/j.gim.2025.101443

Keywords: Human genetics, genomic newborn screening, machine learning, genetic disorders, public health genomics, gene panel prioritization, precision medicine, newborn care, international collaboration