News & Research
The latest AI research and news with real-world stakes — each item sourced, dated, summarized in plain English, and tagged by impact area. Every item is checked against its source before it appears.
Research
Preparing Adult Learners for the Future of Work in the Age of Artificial Intelligence
Jie Ke
International Journal of AI in Pedagogy Innovation and Learning Futures · 2026-09-06
This paper examines how a Bachelor of Science in Professional Interdisciplinary Studies degree-completion program at a Mississippi HBCU can be redesigned to prepare adult learners for an AI-transformed regional labor market. Drawing on environmental scanning and document analysis, the authors identify a significant mismatch: the Deep South is absorbing roughly $75 billion in announced AI capital investment, yet regional bachelor's attainment sits at 27.0% and the Black workforce faces concentrated occupational exposure to AI displacement. The paper proposes a five-phase, 36-month action plan that infuses AI competencies into the existing credential, frames adult learners' workplace experience as an asset for evaluating AI outputs, and addresses the finding that AI-skill job postings carry a 28% wage premium and 51% of such postings fall outside IT roles. The work matters because it targets a population and institution type largely bypassed by current AI investment in higher education.
- Workforce
- Certifications
Research
H&E to recurrence score: A step forward, but not yet a substitute for genomic testing
Shiqiong Zhou, Q. Ke
Translational Oncology · 2026-09-06
Shamai and colleagues trained a multimodal deep-learning model to predict Oncotype DX recurrence scores from routine H&E slides and clinicopathological variables in hormone receptor-positive, HER2-negative early breast cancer. Validated across the TAILORx trial and six external cohorts totaling over 5,000 patients, the model achieved an AUC of 0.898 for identifying high recurrence scores and identified 31% of clinically high-risk postmenopausal women as potentially low-risk, suggesting a path to reducing overtreatment. The authors caution, however, that limitations including intratumoral heterogeneity, unvalidated performance in node-positive disease, calibration uncertainty near thresholds, and digital pathology infrastructure gaps mean the tool should complement rather than replace genomic testing. The work advances the case for AI as a decision aid in precision oncology but underscores that genomic assays remain the gold standard for borderline or discordant cases.
- Quality assurance
- Certifications
Research
When Words Become Obligations: Terminology, Contracts and Compliance in Artificial Intelligence
Rafael Alberto Patron
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-06
This research note examines how terminology differences in the EU AI Act—such as 'user' versus 'deployer,' 'provider' versus 'vendor,' and 'compliance' versus 'conformity'—carry distinct legal, contractual, and evidentiary weight that can create compliance risks for organizations. The author traces each definition to primary sources and identifies two key findings: a concept can be relabeled during the legislative process (e.g., 'user' became 'deployer'), and an obligation can be substantively rewritten while retaining its original name (as happened with Article 4 of the AI Act via Regulation (EU) 2026/1744). The paper proposes a TERM–CLAIM–EVIDENCE model as an operational instrument to help organizations ensure their vocabulary holds up under contract, procurement, or audit scrutiny. This matters for enterprises and certification bodies that must navigate EU AI governance documentation with precision.
- Enterprise
- Certifications
- AI policy
Research
Occupational Vulnerability to AI‐Driven Change: The Role of Skill Composition, Task Structure, and Psychosocial Buffers
Thinuri Welithotage, Behdin Nowrouzi‐Kia
American Journal of Industrial Medicine · 2026-09-06
Analyzing 664 U.S. occupations across 128 skill dimensions, this study finds that vulnerability to AI-driven change is multi-dimensional, shaped not just by automation exposure but also by psychosocial job features like autonomy, task variety, and social interaction. Using K-means clustering, the authors identify five distinct occupational clusters that differ systematically in both AI Exposure (AIOE scores) and a Psychosocial Buffer Index (PBI), showing that high-exposure jobs with strong structural buffers may be less vulnerable than raw automation risk scores suggest. The findings indicate that AI exposure aligns with cognitive skill intensity while buffering capacity varies independently, meaning some low-exposure roles can still face significant occupational stress. The authors argue this framework supports targeted workforce interventions such as retraining, workflow redesign, and collaborative AI integration.
- Workforce
- AI policy
Research
When Words Become Obligations: Terminology, Contracts and Compliance in Artificial Intelligence
Rafael Alberto Patron
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-06
This research note examines how legal terminology in the EU AI Act carries distinct contractual and evidentiary weight that differs from ordinary professional usage. The author traces definitions such as 'user' versus 'deployer,' 'provider' versus 'vendor,' and 'certificate' versus 'certification' to primary sources, demonstrating that a single concept can be relabelled during the legislative process and that obligations can be materially rewritten while retaining their original names. A key example is Article 4 of the AI Act, which was substantively altered by Regulation (EU) 2026/1744 while preserving its position in the text. The paper proposes a TERM–CLAIM–EVIDENCE model to help organizations ensure their vocabulary holds up under contract review, procurement, or audit scrutiny.
- AI policy
- Certifications
- Enterprise
Research
AI Integration and Job Satisfaction among Chinese Secondary School Teachers: A Moderated Mediation Analysis in the Chinese Educational Context
Yajie Zou, Jamil Bin Ahmad, Bity Salwana Alias
Journal of Educational and Social Research · 2026-09-05
This study of 500 secondary school teachers in mainland China finds that AI integration into schools positively predicts teachers' intrinsic job satisfaction, with work environment perception explaining roughly 64% of that relationship. AI literacy moderates the pathway, meaning teachers with higher AI literacy experience stronger improvements in how they perceive their work environment and, in turn, greater satisfaction. The findings suggest that investing in teacher AI literacy training and supportive work environments could amplify the benefits of school AI adoption, though the cross-sectional, non-probability sample limits causal conclusions.
- Workforce
Research
Human-guided continual learning for multifaceted improvement of self-driving vehicles
Haohan Yang, Yi Zhou, Xiaosong Hu et al.
Nature Communications · 2026-09-05
This paper presents a human-guided continual learning method that uses human takeover data to incrementally improve self-driving vehicle performance without requiring full retraining from scratch. Evaluated in both simulations and real-world experiments, the approach enables updated driving policies to match or outperform prior versions on dimensions such as social compliance and rare long-tail event handling. The findings suggest that integrating small amounts of human guidance iteratively can address safety gaps stemming from ambiguous traffic laws and uncommon scenarios, supporting broader human-in-the-loop autonomous systems.
- Workforce
- Quality assurance
Research
Challenges and Recommendations in Regulating AI Medical Devices in the European Union: A Scoping Review
Guilherme Semedo, E Valpaços, Liliana Teles et al.
Healthcare · 2026-09-05
This scoping review synthesizes 70 studies (2018–2025) to map the regulatory challenges and recommendations surrounding AI-enabled medical devices in the EU. Analyzing 261 challenges and 113 recommendations, it finds that the overlapping requirements of the MDR, IVDR, and AI Act—combined with a lack of harmonized standards—constitute the most frequently reported problem domain, while transparency is the only area where recommendations outpace identified challenges. The review provides a structured framework to guide policymakers and regulatory scientists working on AI medical device governance in the EU.
- AI policy
- Certifications
Research
Interpretation in the Age of AI: (Re)Negotiating the Legal Language Zone in Professional Correspondence
Daniel Green
International Journal for the Semiotics of Law - Revue internationale de Sémiotique juridique · 2026-09-05
This conceptual paper introduces the 'Legal Language Zone' (LLZ), a framework for analyzing professional legal correspondence along six communicative dimensions including structural cohesion, terminological precision, and strategic indeterminacy. It examines how large language models can improve drafting efficiency and surface-level correctness in legal communication while falling short in context-sensitive interpretation, strategic judgment, and anticipating future interpretive consequences. The paper argues for a collaborative rather than substitutive human–AI model in which legal professionals retain responsibility for ethical judgment and professional accountability. These findings bear directly on questions of workforce roles, professional responsibility, and the governance of AI in legal practice.
- Workforce
- AI policy
Research
Robot risk awareness, zero-sum beliefs, and hotel frontline employee’s self-objectification: the moderating roles of psychological safety and gender
Lujie Hao, Shaofeng Wang
Humanities and Social Sciences Communications · 2026-09-05
This study examines how hotel frontline employees psychologically respond to the growing presence of service robots. Using Social Comparison Theory and Cognitive Appraisal Theory with survey data from 298 employees, the researchers find that greater robot risk awareness increases self-objectification—a form of psychological defense—mediated by zero-sum beliefs about human-robot competition. Psychological safety buffers this effect, while gender shapes the pathway: male employees are more likely to develop zero-sum beliefs, whereas female employees more readily internalize those beliefs as self-objectification. The findings offer practical guidance for creating psychologically safe and gender-sensitive workplaces amid automation.
- Workforce
Research
Fairness and Bias Detection in Artificial Intelligence-Driven Decision-Making in Employment Processes using Machine Learning Model
Francisca Nonyelum Ogwueleka, Uchenna Igboeli, Isa Ibrahim Wadda
Journal of Science Innovation and Technology Research · 2026-09-05
This study evaluated fairness and bias in AI-driven hiring systems by developing machine learning models—Random Forest, Gradient Boosting, and Support Vector Machine (SVM)—trained on 1,000 applicant records with demographic and professional attributes. The SVM model achieved the highest classification accuracy (87.0%) and demonstrated the lowest bias across gender, location, experience, and disability metrics. Feature importance analysis found years of experience to be the most influential hiring factor, while gender influence varied by model. The authors recommend continuous monitoring and regular fairness audits to promote more equitable AI-driven recruitment practices.
- Workforce
- AI policy
Research
Procedural fairness, accountability, and legitimacy in the AI-enabled state: evidence from China
Guifeng Fan, Weihua Lyu
Humanities and Social Sciences Communications · 2026-09-05
This study investigates whether traditional legitimacy mechanisms still shape public acceptance of AI-mediated legal obligations in China. Using a mixed-methods design combining surveys (N=642) and semi-structured interviews (n=41) with PLS-SEM analysis, the researchers find that perceived procedural fairness and AI accountability significantly boost perceived legitimacy of the AI-enabled state, which in turn strongly predicts acceptance of algorithmically enforced duties like traffic regulation and administrative decisions. Trust in state institutions works primarily through legitimacy rather than independently, and qualitative data reveals conditional support alongside concerns about opacity, data misuse, and reduced human oversight. The findings demonstrate that algorithmic governance remains normatively dependent on legitimacy judgments, offering empirical insight into AI authority within China's centralized governance framework.
- AI policy
Research
From Predictive Accuracy to Human-Centric Decision Support: An Operational HCT-ML Framework for Intelligent Transportation Systems
Bappa Muktar
arXiv · 2026-09-05
This paper proposes the Human-Centric and Trustworthy Machine Learning (HCT-ML) framework, a structured decision-support profile for deploying machine learning in intelligent transportation systems (ITS). Rather than inventing new AI principles, it operationalizes six human-centric dimensions—relevance, explainability, fairness, uncertainty, human oversight, and actionability—into concrete evidence requirements, indicators, thresholds, and non-compensatory deployment gates organized around four transport-specific constructs. The framework is benchmarked against major governance standards including the NIST AI RMF, OECD AI Principles, EU AI Act, IEEE 7000-series, and USDOT guidance, and is illustrated with automated-driving safety cases and a municipal road-safety worked example using Montréal open data. The authors explicitly note the framework is not empirically validated but provides a reproducible protocol for future stakeholder studies and field deployments.
- AI policy
- Quality assurance
Research
The intrinsic mechanism of public technology acceptance, risk perception, and pharmaceutical safety trust in the context of AI-driven pharmaceutical regulation
Yan-Ping Jing, Dong-Mei Ding, Liang-Fa Sun
Scientific Reports · 2026-09-05
This study examines how public acceptance of AI in pharmaceutical regulation relates to risk perceptions and trust in drug safety, using a survey of 412 Chinese adults analyzed with structural equation modeling. Grounded in cognitive consistency theory, it finds that higher technology acceptance is significantly associated with lower data privacy, algorithmic error, and accountability risks, which in turn associate with greater pharmaceutical safety trust. Accountability risk showed the strongest associations across both pathways. The authors argue that regulatory authorities should prioritize clear accountability frameworks for AI-assisted regulation to address public concerns.
- AI policy
- Quality assurance
Research
Artificial Intelligence Based Framework for Student Engagement Assessment in Classroom Environments
Munish Saini, Harsh Sharma, Eshan Sengupta
Cognitive Computation · 2026-09-05
AI-TEACH is a multimodal AI framework that automatically detects and quantifies student distraction in classrooms by fusing video and audio streams using YOLO-NAS, ByteTrack, MediaPipe, and a BiLSTM network. Validated on classroom recordings with 200 students, the system achieved 90% accuracy and 92% recall in distraction detection, outperforming unimodal and non-real-time approaches. The framework generates per-incident severity scores and session-wide engagement reports via an instructor dashboard, enabling real-time, data-informed teaching decisions. Its scalability and objectivity over manual observation methods have direct implications for inclusive education and adaptive pedagogy.
- Workforce
- Quality assurance
Research
HYBRID LLM AND DETERMINISTIC SEVERITY ENGINE FOR NIST SP 800-53 COMPLIANCE DECISION SUPPORT
M.S Rakesh Babu Rapolu
INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & TECHNOLOGY · 2026-09-05
This paper presents a hybrid AI system for automating cloud security compliance reviews under NIST SP 800-53 Rev. 5 and FedRAMP High standards. The system pairs a large language model (LLM) for organizing and assessing evidence with a deterministic Python engine that makes final risk classifications across four levels (LOW, MODERATE, HIGH, CRITICAL). Evaluated on 100 synthetic scenarios spanning all 20 NIST control families, the system achieved 92% exact agreement on combined compliance-status and risk outcomes, 98% status accuracy, 100% CRITICAL recall, and a mean audit latency of 7.70 seconds. The approach demonstrates that hybrid LLM-deterministic architectures can meaningfully accelerate and standardize compliance decision-making in regulated cloud environments.
- Quality assurance
- Certifications
Research
Artificial Intelligence in Echocardiography and Point-of-Care Ultrasound: Applications, Clinical Integration, and Future Directions
Fabio Savorgnan, Paola Pilla, Sarah Visokay et al.
Current Pediatrics Reports · 2026-09-05
This narrative review examines how AI is being applied across the echocardiography and point-of-care ultrasound workflow—spanning image acquisition, view recognition, segmentation, automated quantification, disease classification, and structured reporting. The strongest clinical evidence supports automated left ventricular segmentation and ejection fraction estimation, with a randomized trial showing AI-generated ejection fraction assessment was noninferior to sonographer assessment and required fewer cardiologist corrections. The review highlights that AI should currently serve as an augmentative rather than autonomous technology, with clinicians retaining interpretive responsibility, and identifies key gaps including pediatric/congenital heart disease data, cross-device generalizability, and medicolegal accountability. Safe implementation is said to require external validation, bias assessment, uncertainty communication, audit trails, and institutional governance.
- Quality assurance
- Workforce
Research
Performance of vision-language models compared with 252 medical students on text-only and image-based dermatology examinations
Ozan Erdem, Abdurrahim Yilmaz, Ahmet Sait Şahin et al.
Scientific Reports · 2026-09-05
This study benchmarked four vision-language models (ChatGPT-4o, ChatGPT-5, Gemini 2.5 Flash, and Gemini 3 Pro) against 252 fifth-year medical students on ten dermatology clerkship examinations combining text-only and image-based questions. All VLMs substantially outperformed students on text-only components (mean scores above 95 vs. 84.9), but image-based performance was uneven: Gemini 3 Pro and ChatGPT-5 surpassed students while students outperformed the other two models. The findings show that multimodal medical competence in VLMs is inconsistent and model-dependent, and the authors conclude these tools are better suited as complementary rather than standalone resources in dermatology education.
- Quality assurance
- Certifications
Research
Predictive Policing and Artificial Intelligence in India's Criminal Justice System: Constitutional Limits and a Framework for Accountable Use
Harinath Reddy G, Manne Sainath
International Journal of Law Management & Humanities · 2026-09-05
This paper examines the constitutional and regulatory limits of AI-powered predictive policing in India's criminal justice system. Using doctrinal and comparative analysis across Indian constitutional law, U.S. and EU precedents, it argues that place-based analytics may support non-coercive resource allocation under strict safeguards, but person-based risk scores must never form the basis for suspicion, coercive police action, or judicial outcomes. The authors propose a framework including statutory authorization, algorithmic impact assessments, data-quality and equality audits, contestability mechanisms, and independent oversight. The paper matters because it articulates concrete accountability safeguards for AI systems that risk amplifying historical enforcement inequalities into automated forecasts of unequal risk.
- AI policy
- Certifications
Research
Youth Entrepreneurship in the Digital Era: A Systematic Review and Capability Formation-Activation Framework
James Wang
Scholarly review . · 2026-09-04
This systematic review synthesizes 115 peer-reviewed articles on youth entrepreneurship, examining why entrepreneurial intention rarely translates into sustained venture creation. The authors find that access to finance, networks, institutional legitimacy, and low-risk experimentation opportunities are critical mediating factors, and that entrepreneurship education builds relevant skills but its long-term effect on actual venture creation is unclear. Digital platforms and generative AI can lower barriers to entry but may also create platform dependency and over-reliance on AI outputs that young founders cannot critically evaluate. The resulting capability formation-activation framework recommends that programs and tools be judged not just by whether they raise intention or technology use, but by whether they support persistence, adaptation, and responsible decision-making.
- Workforce
- Enterprise
Research
Digital Trade and Data Governance in Africa: Reforming Ghana’s Data Protection Act for Secure Cross-Border Data Flows Under the African Continental Free Trade Area (AFCTFTA) Framework
Joseph Kwaku Asamoah
JOURNAL OF BUSINESS AND AFRICAN ECONOMY · 2026-09-04
This paper examines how Ghana's Data Protection Act (2012) lacks adequate guidance on cross-border data transfers, a gap that poses risks to privacy protection, regulatory compliance, and economic integration under the African Continental Free Trade Area (AfCFTA). Using doctrinal and comparative legal analysis drawing on the ECOWAS Supplementary Act on Personal Data Protection and China's PIPL, the authors propose a hybrid regulatory model featuring presumptive adequacy for regional partners, standard contractual safeguards, and tiered security assessments for high-risk data exports. The reforms are intended to align Ghana's framework with AfCFTA digital trade obligations while safeguarding constitutional privacy rights and supporting Africa's broader digital economy.
- AI policy
Research
Evaluating GenAI ‐produced feedback on undergraduate bioscience essays against good higher education feedback practice
Annabel Court, Nigel Francis, Andrew Shore et al.
FEBS Open Bio · 2026-09-04
This study evaluated ChatGPT4o-generated formative feedback on 30 first-year bioscience essays, assessing alignment with established principles of good higher education feedback practice using a bespoke rubric and mixed-methods analysis. Results showed that GenAI feedback was generally accurate and essay-specific, with strong content focus and feed-forward advice, but was limited in critical guidance and motivational feedback. The authors conclude that while GenAI may offer useful instant formative support, substantial scaffolding by educators is likely needed for it to be effective in real learning contexts.
- Quality assurance
- Workforce
Research
Evaluating performance bias in face-to-BMI vision transformer models across diverse human populations
J C Hoffman, Michael Gurven, Hillard Kaplan et al.
bioRxiv (Cold Spring Harbor Laboratory) · 2026-09-04
This study evaluates performance bias in Vision Transformer (ViT-H/14) models that estimate body mass index from facial photographs, testing generalization across four Indigenous populations (Orang Asli, Ju/'hoansi, Sama, and Tsimane). The researchers found that in-distribution training—using data from the same population being assessed—consistently produced the best accuracy, while training on a broad cross-cultural dataset was the strongest fallback when target-population data was unavailable. The findings highlight that mainstream face-to-BMI models, typically trained on government records, social media, and celebrity photos, fail to represent global morphological diversity and introduce systematic bias against underrepresented populations. The study argues that diverse, globally representative training datasets are essential for health AI tools to generalize reliably across human populations.
- Quality assurance
- AI policy
Research
Exploring the Ethical Boundaries of AI Use in Secondary Vocational Education
Peter Tokoš, Roman Hrmo, Juraj Miština et al.
International Journal of Engineering Pedagogy (iJEP) · 2026-09-04
A survey of 1,745 vocational secondary school students across Slovakia, Czech Republic, Hungary, and Poland finds widespread AI tool use for understanding materials, generating texts, completing assignments, and project work. While many students expressed confidence in evaluating AI-generated content, thematic analysis of open-ended responses revealed academically dishonest uses—such as generating code and writing essays without meaningful cognitive engagement—driven by time efficiency and pragmatic task management. The study concludes that pedagogical approaches must be updated to foster responsible AI use, and that skills like prompt formulation, output verification, and critical evaluation are becoming essential competencies for vocational graduates.
- Workforce
- AI policy
Research
The dependency trap: a systematic review of the contributing factors and consequences of student over-reliance on generative AI in higher education
Chenhui An, Ke Lu, Junfeng Yang
Humanities and Social Sciences Communications · 2026-09-04
This systematic review synthesizes 54 empirical studies (2022–2026) on student over-reliance on generative AI in higher education, identifying four contributing factor categories—psychological factors, capability gaps, environmental stress, and technological pull—and four consequence domains including cognitive impairment, learning efficacy decline, academic integrity risks, and psychosocial impairment. Drawing on Lewin's Field Theory and systems dynamics, the authors argue that these consequences can feed back into the very conditions that drive over-reliance, creating a self-reinforcing 'dependency trap.' The paper recommends that institutions move away from restrictive AI policies toward pedagogical innovation, process-oriented assessment, and holistic psychosocial support as more effective responses.
- Workforce
- AI policy