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
Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Divyanshu Kumar, Nitin Aravind Birur, Tanay Baswa et al.
arXiv · 2026-09-09
This paper presents a black-box red-teaming framework called SAGE-RT for evaluating security risks in agentic AI systems—AI agents that autonomously read inputs, call tools, and act across multiple steps. The framework introduces a seven-domain risk taxonomy and automatically generates 120 adversarial scenarios per domain, validated by LLM judges. Testing across two multi-agent architectures (CrewAI and AutoGen) with four base models revealed severe vulnerabilities: 56.25% average governance risk, 65% privacy risk in multi-agent configurations, and up to 85% agent behavior vulnerabilities. The work highlights that standard single-turn evaluations are inadequate for agentic systems and offers a scalable, access-free path toward safer agent deployments.
- Quality assurance
- AI policy
Research
Watermarks Without Verification: AI Text Watermarking After the EU AI Act
Alexander Nemecek, Vipin Chaudhary, Erman Ayday
arXiv (Cornell University) · 2026-09-09
This paper examines the governance implications of AI text watermarking under Article 50 of the EU AI Act, which took effect August 2, 2026, requiring generative AI providers to mark and make detectable all AI-generated content. The authors argue that the central failure is not watermarking itself but the inability of third parties to verify vendor claims or user objections about quality degradation, identifying information encoding, or robustness. Testing the open-source SynthID-Text implementation on two open-weight models, they find that on prose the watermark's effect is no larger than changing a sampling seed, while on code it costs up to three points of correctness on one model and is undetectable on the other—a limitation of detectability rather than quality. The paper maps these verification gaps to specific institutional requirements: release of matched outputs, configuration disclosure, accredited audits, a shared evaluation protocol, and interoperable detection.
- AI policy
- Quality assurance
Research
The Vibe Shift in Software Engineering: Evaluating AI-Led Conversational Programming for Performance, Cognition, and Responsible Adoption
Sales G. Aribe, Louie Jay S. Labastida
arXiv · 2026-09-09
This study evaluates 'vibe coding,' an AI-led conversational programming approach where developers generate software via natural-language interaction with large language models, comparing it to traditional and AI-assisted coding in a 30-participant mixed-methods experiment. Results show vibe coding reduced task completion time by 27% versus traditional coding and 12% versus AI-assisted coding, but came with lower maintainability indices and higher security vulnerabilities, indicating meaningful quality trade-offs. Usability was rated 'good' (SUS = 71.4) and cognitive workload was moderate (NASA-TLX = 55.5), with qualitative themes highlighting trust calibration, loss of control, and prompt-engineering strategy as key concerns. The authors propose a three-pillar responsible adoption framework emphasizing hybrid human-AI integration, human oversight, and context-aware deployment, positioning vibe coding as productive but requiring critical governance.
- Workforce
- Enterprise
- Quality assurance
Research
Fed-XLM-R: a privacy-preserving federated framework with adapter-scoped differential privacy for mental health triage-level intent classification on resource-constrained edge devices
Karthiga M, Emerson Raja Joseph, Subhash Patil et al.
Frontiers in Digital Health · 2026-09-09
Fed-XLM-R is a federated learning framework for classifying mental health triage-level intent (Anxiety/Depression/Normal) on edge devices while preserving patient privacy. By confining differential privacy noise to lightweight adapter layers (only 0.87% of parameters), the system achieves 92.3% accuracy and an F1 of 0.918 at a strong privacy budget (ε=1.0), within 0.5% of a centralized baseline. INT8 quantization cuts inference latency 3.8-fold on low-end CPUs, and adapter-only communication reduces bandwidth costs by over 40-fold. The paper demonstrates that privacy-preserving, edge-deployable federated NLP for mental health triage is feasible, though it has not yet been validated in real-world clinical or multilingual deployment settings.
- Quality assurance
- AI policy
Research
Adversarial Training for Tabular Credit Scoring: A Multi-Attack Robustness Evaluation in P2P Lending
Gijs A. F. Niewzwaag, Marijn G. S. Veth, Manuele Massei et al.
arXiv (Cornell University) · 2026-09-09
This paper benchmarks adversarial training defenses for machine learning credit-scoring models used in Peer-to-Peer lending, testing three model families against four attack types (FGSM, PGD, Salt-and-Pepper noise, and DeepFool) plus mixed-attack regimes on Lending Club data. Results show that adversarial training strongly improves robustness against the specific attack it was trained on and generalizes reasonably within gradient-based attacks, but transfers poorly to non-gradient corruption. Mixed-attack training provides the most balanced defense across diverse attack types without sacrificing clean-data performance. The findings support adopting multi-attack stress testing as part of credit-model governance rather than relying on single-attack defenses, which overstate real-world resilience.
- AI policy
- Quality assurance
Research
Copyright and AI : The building blocks
Jordan M. Blanke
American Business Law Journal · 2026-09-09
This law review article argues that existing copyright doctrine—built on precedents like Baker v. Selden, Feist, Campbell v. Acuff-Rose, and Warhol v. Goldsmith—already provides the analytical tools needed to resolve AI-related disputes over training data, AI-generated output, originality, and authorship. Rather than calling for a doctrinal overhaul, the article contends that copyright law has historically adapted alongside new technologies and will continue to do so with AI. The analysis covers core concepts including idea/expression, human authorship, substantial similarity, and transformative use as they apply to machine-mediated creativity.
- AI policy
Research
A global framework for artificial intelligence education in medicine: international working group recommendations
Laura C. Rosella, Yindalon Aphinyanaphongs, James Barry et al.
npj Digital Medicine · 2026-09-09
An international working group developed a structured framework for AI education in medicine, organized across seven domains and 24 learning objectives. The authors argue that current AI education is fragmented, overly technical, and disconnected from clinical realities, and that effective training must be clinically grounded, ethically integrated, and implementation-aware across the full training continuum. The framework is intended to guide educators in translating AI competencies into medical practice. This work directly informs how healthcare professionals are certified and trained in AI literacy.
- Certifications
- Workforce
Research
Zero-shot governance
Carlo Perrotta
Journal of Education Policy · 2026-09-09
This paper introduces 'zero-shot governance' to describe scenarios where general-purpose, domain-agnostic generative AI — particularly Large Language Models — intervenes in policy decisions. Using the UK government's discontinued Redbox prototype as a case study, the author analyzes its codebase alongside the political and economic conditions shaping its development to show how off-the-shelf LLMs are being integrated into civil servants' professional workflows. The central finding is that the general-purpose nature of LLMs is a structural, ineliminable feature that can be mitigated but not removed, posing lasting governance challenges. The paper concludes by reflecting on the implications of this dynamic specifically for education policy.
- AI policy
Research
Competing strain and resource pathways linking AI-enabled work systems to employee wellbeing
A. S. Sathish, R. Indradevi
Frontiers in Artificial Intelligence · 2026-09-09
This study of 516 Indian IT employees uses PLS-SEM and the Job Demands–Resources and Conservation of Resources frameworks to examine how AI adoption and digital work intensity affect employee wellbeing through technostress and psychological resilience. Key findings show that digital work intensity positively influences wellbeing, psychological resilience is a significant predictor and mediator of wellbeing, while technostress does not significantly affect wellbeing directly. Perceived organisational support is positively linked to wellbeing but does not moderate the technostress–wellbeing relationship. The paper highlights that AI-enabled work environments create both strain and resource-building processes, with psychological resilience playing a stronger role than technological strain in shaping employee outcomes.
- Workforce
Research
Artificial Intelligence and Human Resource Management in the South African Public Sector: Opportunities, Risks, and Institutional Readiness
Humphrey Motsepe
International Journal of Applied Research in Business and Management · 2026-09-09
This systematic literature review examines how AI technologies are influencing human resource management (HRM) in South Africa's public sector, drawing on peer-reviewed publications, policy documents, and international governance reports from 2020–2026. The study finds that AI offers meaningful potential for administrative efficiency, workforce planning, predictive analytics, automated recruitment screening, and digital performance management, but that adoption is held back by governance gaps, digital skills shortages, ethical risks, and limited infrastructure. The authors conclude that realizing AI's benefits in public-sector HRM will require institutional reforms, investment in digital competencies, and clear governance frameworks, offering strategic guidance for South Africa's ongoing digital transformation efforts.
- Workforce
- AI policy
Research
Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: a five-year, multi-framework study
Gwenolé Quellec, Mathieu Lamard, Sarah Matta et al.
Scientific Reports · 2026-09-09
This five-year, multi-framework study evaluated OphtAI, a CE-marked AI system for automated diabetic retinopathy screening, across four distinct settings—a benchmark dataset, a large-scale US Veterans Affairs study, an external validation using handheld cameras in Finland, and a UK NHS comparative evaluation. The researchers found substantial variability in sensitivity and specificity across settings that differed in imaging devices, patient populations, referral definitions, and handling of ungradable images. The findings demonstrate that AI performance in DR screening is not a fixed property of the system itself but an emergent property of the evaluation framework. The authors call for context-aware validation strategies and more standardized evaluation protocols for clinical AI systems.
- Quality assurance
- Certifications
Research
Generative AI in Clinical Dental Education: Balancing Innovation With Academic Integrity Through Teacher‐Centric Implementation
Khaled E. Ahmed, Michael F. Burrow
Australian Dental Journal · 2026-09-09
This narrative review examines how generative AI is reshaping dental education, focusing on faculty perspectives, assessment redesign, and clinical integration within the Australian context. The authors find widespread gaps in faculty AI knowledge and pedagogical readiness, and argue that traditional assessments are highly vulnerable to AI assistance—raising fundamental questions about what competencies those assessments actually measure. The paper calls for authentic, performance-based evaluation that captures skills AI cannot replicate, such as manual dexterity and patient interaction, alongside structured faculty development and phased implementation. Australian regulatory frameworks are identified as actionable guides for responsible, ethically grounded adoption.
- Certifications
- Quality assurance
Research
Industry Ethics
Simone Casiraghi
arXiv · 2026-09-09
This chapter examines the implications of translating AI ethics into formal technical standards, focusing on the IEEE 7000-2021 standard and its associated CertifAIEd certification system, and their potential alignment with the EU AI Act (AIA). The author argues that while these IEEE tools promise more responsible AI governance, they reduce ethics to an engineering requirement shaped by standardization logic, exposing persistent challenges around compliance, vagueness, representativeness of standard-setting bodies, and procedural transparency. The analysis maps overlaps and gaps between IEEE ethical initiatives and AIA requirements, offering a critical perspective on whether standardized ethics can meaningfully govern AI behavior.
- Certifications
- AI policy
Research
Data scraping for AI model training and data privacy: a comparative analysis of United States and European Union Data Privacy Laws
Olumide Timothy Ajayi, Chukwuemezie Charles Emejuo, Francis Aondongu Wayo et al.
Humanities and Social Sciences Communications · 2026-09-09
This paper conducts a comparative legal analysis of data scraping for AI model training under U.S. and EU data privacy frameworks. It finds that in the U.S., scraping publicly accessible data for AI training is generally permissible, while in the EU such scraping of personal data almost always conflicts with GDPR Article 6's lawful basis requirements, though enforcement has been inconsistent across member states. The authors conclude that data scraping broadly undermines privacy principles and call for reform and reconciliation of existing privacy laws to balance AI innovation needs with data protection.
- AI policy
Research
Dancing with AI: how human-AI interaction affects employee task performance
Y. Wu, Y. Wu, Ping Li et al.
Humanities and Social Sciences Communications · 2026-09-09
This study develops and validates a 19-item measurement scale for human-AI interaction in the workplace, identifying three dimensions: anthropomorphic tool, adaptive trust, and unidirectional emotional connection. Drawing on self-concept theory, the authors find that human-AI interaction is positively associated with employee task performance, with role identity and self-efficacy partially mediating that relationship. The findings offer both a conceptual framework and a validated instrument for understanding how generative AI shapes workforce outcomes.
- Workforce
Research
Disambiguating Adaptivity in AI Research and Regulation
Bartosz Michał Radomski
arXiv · 2026-09-09
This paper argues that the term 'adaptive' in AI is not vague but systematically ambiguous across four distinct research traditions from cognitive science, biology, and AI, which can produce contradictory characterizations of the same system. To manage this ambiguity, the authors propose a four-question disambiguation protocol asking what changes, against what norm, who set that norm, and whether the system can detect its own failure by that norm. The last question carries direct regulatory weight: systems unable to register their own failures cannot self-monitor, so oversight responsibility falls entirely on external actors and autonomous deployment should not receive reduced oversight regardless of measured performance. The work has direct implications for AI regulation and the governance of autonomous systems.
- AI policy
- Certifications
Research
Beyond efficiency: interactional foundations of fairness, accountability and transparency in AI-supported performance evaluation
Md Irfanuzzaman Khan, Robin C. Ladwig
Journal of Enterprise Information Management · 2026-09-09
This survey-based study of 289 Australian managers examines how AI-supported performance evaluation systems affect perceptions of fairness, accountability, and transparency (FAT), and how those perceptions relate to managerial trust. Using PLS-SEM analysis, the authors find that interaction quality, human agency, and perceived humanness positively predict all three FAT dimensions, while transparency shows the strongest association with trust. The research highlights that organizations deploying AI in performance evaluation should preserve human oversight, documented override authority, and credible appeal mechanisms to foster trust. These findings carry direct implications for how enterprises design and govern AI-driven HR processes.
- Workforce
- Enterprise
Research
Governing Privately Supplied AI in Publicly Governed European Health Systems: Procurement, Accountability and Institutional Capacity under the EU AI Act
THEODOROS ZARKOS
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-09
This preprint develops a normative and health-policy framework for how European public health systems can procure privately supplied AI while maintaining genuine institutional accountability. It introduces the concept of an 'accountability–capacity gap'—where formal responsibility stays with public health organizations but critical technical and operational capabilities reside with private vendors—and proposes five safeguards addressing non-transferable public responsibility, informational access, traceability, corrective authority, and exit capacity. The analysis applies these safeguards to the EU AI Act, medical-device regulation, and public-procurement instruments, arguing that regulatory compliance alone is insufficient without practical institutional capacity to investigate, intervene, or switch suppliers. The paper positions public procurement as a proactive governance mechanism to preserve accountability across the full AI lifecycle.
- AI policy
- Certifications
Research
Beyond Automatic Fluency: Human Agency, Quality Risk and Professional Practice in AI-Assisted Machine Translation
Kizito Innocent Lawa
East African Journal of Arts and Social Sciences · 2026-09-09
This paper examines how AI-assisted machine translation—including neural MT and large language models—is reshaping translators' work, competence, and professional agency. While AI can generate fluent output quickly, the authors argue that fluency can mask errors in meaning, terminology, register, and factual accuracy, creating significant quality risks. Using a critical integrative review and documentary policy analysis, the paper proposes the TRACE-MT framework covering task and risk classification, responsible data handling, human oversight, contextual quality assurance, and equity in language coverage—arguing that accountable human judgment must remain decisive rather than being displaced by uncritical automation. The framework addresses translator education, post-editing protocols, procurement, and organizational accountability, with particular attention to low-resource African languages where system coverage does not guarantee contextual reliability.
- Workforce
- Quality assurance
- AI policy
Research
Ethics and the Regulation of Artificial Intelligence
Simone Casiraghi
arXiv · 2026-09-09
This book examines how ethics discourse has been institutionalized within the EU's AI regulatory framework, analyzing three case studies—research ethics committees, advisory expert groups, and standardization bodies—through the lenses of law and science and technology studies. It evaluates AI ethics governance against principles of accountability, transparency, and participation, finding that ethical considerations frequently take a technocratic form that lacks the democratic checks and balances typical of traditional technology regulation. The authors argue for a 'deflation or re-politicization' of ethics in AI governance to correct these legitimacy deficits. The work is directly relevant to policymakers, civil society actors, and legal scholars engaged in AI regulation in the EU and beyond.
- AI policy
Research
Operational digital twins in the built environment: A systematic review of semantic infrastructure and deployment practice
Chady Elias, Isabella Hutchins, Raja R. A. Issa
Journal of Information Technology in Construction · 2026-09-09
This systematic review of 46 operational digital twin deployments in the built environment finds that while meaningful operational capabilities exist—particularly for HVAC, energy, and indoor environmental quality—they have not yet coalesced into portable, transferable architectures. Formal semantic representations appeared in roughly half of reviewed systems, but fewer than a third used them to directly shape analytics or decision-making, and operational outputs rarely flowed into maintenance systems or governed external processes. The authors conclude that machine-readable building context and lifecycle-governed metadata are key constraints on scalability, and argue that semantic infrastructure should be established early in a building's lifecycle to enable AI-ready, site-agnostic operations. The findings matter for enterprise facility management and AI deployment at scale, highlighting a significant maturity gap between demonstrated capabilities and replicable, portfolio-wide adoption.
- Enterprise
- Quality assurance
Research
Human Capital Formation, Labor Market Transformation, and Wage Dynamics in AI-Semiconductor Industrial Zones: Predictive Economic Modeling for the Pax Silica Economic Security Zone
Laszlo Pokorny
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-09
This study builds a predictive economic model to assess whether the Philippine labor force can meet the high-skill demands of the Pax Silica Economic Security Zone, a planned AI and semiconductor hub projected to generate roughly 190,000 jobs and attract USD 10 billion in investment. Using publicly available data and calibrated synthetic microdata, the authors find a 30.9% average skills gap concentrated in engineering roles, a 51.7% conditional semiconductor wage premium, and strong social rates of return on both engineering education (19.9%) and short-cycle vocational training (58.6%). The research concludes that human capital investment is economically justified but must be sequenced ahead of physical infrastructure, and that the Penang model of incremental, training-led upgrading is the most applicable international template. The findings directly bear on workforce development strategy, wage dynamics, and labor market restructuring in an emerging high-tech industrial zone.
- Workforce
- AI policy
Research
The relationship between AI anxiety and work engagement among early childhood teachers: a moderated mediation model of job crafting and perceived organizational support
Haiyan Cui, Yantao Shi, Xueli Hui et al.
Frontiers in Psychology · 2026-09-09
This study surveyed 933 early childhood teachers to examine how AI anxiety affects their work engagement, finding that greater AI anxiety is negatively associated with work engagement. Job crafting (proactively reshaping one's role) mediated this relationship, while perceived organizational support moderated the link between AI anxiety and job crafting. The findings suggest that reducing AI anxiety, strengthening organizational support, and cultivating job crafting behaviors are practical strategies for sustaining teacher engagement as AI becomes more embedded in early childhood education.
- Workforce
Research
Governing medical AI under the EU AI Act: an integrated compliance-by-design framework
Qaiser Khan, Abdul Raffay Saeed
Frontiers in Digital Health · 2026-09-09
This paper examines how multiple overlapping EU regulations—the AI Act, MDR, IVDR, GDPR, and EHDS—jointly govern medical AI systems, arguing that continuous lifecycle compliance is preferable to one-time market-entry assessments. It identifies training and validation datasets as a key regulatory convergence point and analyzes how responsibility is allocated across the general-purpose AI value chain. The authors propose a compliance-by-design framework that translates these intersecting legal requirements into auditable governance artefacts spanning system classification, data governance, human oversight, and post-market monitoring. The framework is illustrated through two contested use cases and offers tailored recommendations for different stakeholder groups, though the authors note it requires future empirical validation.
- AI policy
- Certifications
Research
GENERATIVE ARTIFICIAL INTELLIGENCE IN VOCATIONAL UNDERGRADUATE FINANCE AND ACCOUNTING EDUCATION: APPLICATIONS, RISKS, AND FUTURE DIRECTIONS
Bijie Li, HaoXuan Li
World Journal of Educational Studies · 2026-09-09
This narrative review examines how generative AI tools, particularly large language models, are being integrated into vocational undergraduate finance and accounting education before institutions have fully adapted their curricula or controls. The paper identifies applications such as dialogic tutoring, formative feedback, case simulation, and spreadsheet support, while warning that fluent AI output may contain fabricated facts, outdated rules, or incorrect calculations. The authors propose a responsible integration model that separates AI-free foundational learning from guided and critically evaluated AI use, requiring source verification, calculation checks, and oral or practical defenses. The findings matter for both workforce preparation and quality assurance in vocational programs, as unrestricted AI assistance risks undermining the independent professional judgment students need.
- Workforce
- Quality assurance