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
Consumers with Weaker Applications Are Less Receptive to Algorithmic Evaluation
Qiao Liu, Gerald Häubl, Noah Castelo
Journal of Consumer Research · 2026-08-23
This study examines how consumers react when organizations use algorithms rather than humans to evaluate applications for loans, insurance, and similar services. The key finding is that applicants who believe their application is weak are more deterred by algorithmic evaluation than stronger applicants, because weaker applicants prefer the flexibility and leniency they associate with human evaluators. This asymmetry means that algorithmic evaluation disproportionately discourages weaker applicants from applying at all, with implications for who actually gains access to valued services and opportunities.
- Enterprise
- AI policy
Research
A Comparative Analysis of Artificial Intelligence Regulatory Frameworks in India and European Union
Nimisha Mishra, T.V Rajesh
International Journal of Law Management & Humanities · 2026-08-23
This paper compares the European Union's AI Act (2024) with India's Digital Personal Data Protection Act (2023) to assess the regulatory gap in AI governance between the two jurisdictions. The EU AI Act provides a comprehensive, risk-based legislative framework covering unacceptable, high, limited, and minimal risk categories with strict compliance requirements for high-risk sectors such as healthcare, employment, and law enforcement. India, by contrast, lacks a dedicated AI-specific statute, leaving AI governance fragmented and largely policy-driven, with no formal mechanisms for risk classification, algorithmic accountability, or addressing harms like deepfakes and automated decision-making bias. The paper concludes with recommendations for India to develop a rights-protective, innovation-friendly AI regulatory framework aligned with its constitutional values and socioeconomic context.
- AI policy
Research
Reskilling the Banking Workforce in the Age of Artificial Intelligence: Emerging HRM Priorities and Future Skill Requirements in the Indian BFSI Sector
Avadhesh Vyas, Mr. Virendra Choudhary
IJARCCE · 2026-08-23
This conceptual review examines how AI-driven automation is reshaping skill requirements for workers in India's banking, financial services, and insurance (BFSI) sector, drawing on evidence from over twenty-five peer-reviewed articles and institutional reports including those from the World Economic Forum, Reserve Bank of India, and State Bank of India. The study identifies five critical skill categories gaining importance—AI and digital literacy, data and analytical skills, cybersecurity and risk governance, human-centric skills, and adaptability—and proposes a six-stage HRM reskilling framework covering skill-gap identification, role-based mapping, targeted reskilling, internal mobility, continuous capability assessment, and responsible AI culture. The paper argues that reskilling must blend technical training with ethical judgment and human-oversight capability, and advances a 'Reskill-Redeploy-Retain' strategic model. The findings signal a broader shift from training-oriented HR practices toward adaptive, skill-based talent management in the Indian BFSI sector.
- Workforce
Research
Impacto de la inteligencia artificial en la reconfiguración del empleo: Evidencia empírica en economías avanzadas y Sudamérica
Said Paredes-Castillo, Saylon Cuchiparte-Guamangate, Miguel Sangurima
Multidisciplinary Latin American Journal (MLAJ) · 2026-08-23
This study examines how AI industry development and venture capital investment in AI affect employment across 14 advanced and South American economies from 2013 to 2023, using World Bank and ILO panel data analyzed with dynamic econometric models. The findings indicate that corporate AI and R&D investment via venture capital reduces industrial (manufacturing-sector) jobs while simultaneously increasing total employment in the economy. The authors interpret this as a technology-driven sectoral restructuring—workers shift out of industry into other sectors—rather than a net loss of jobs overall. The evidence spans both advanced and South American economies, making it relevant for understanding AI's labor market effects across different development contexts.
- Workforce
- AI policy
Research
A GOVERNANCE-ORIENTED PROMPT ENGINEERING FRAMEWORK FOR AI-ASSISTED INTERNAL AUDITING
Özden Şentürk
Denetişim · 2026-08-23
This conceptual paper develops a governance-oriented framework for how internal auditors should structure and document their use of large language models (LLMs). It maps 13 audit-specific prompt types to the four phases of the internal audit lifecycle—planning, fieldwork, reporting, and follow-up—treating AI-generated outputs as formal working-paper artifacts subject to human validation and documentation controls. The framework argues that structured prompt engineering can improve repeatability, transparency, and professional skepticism while clarifying that LLM output should serve as intermediate analytical support rather than audit evidence. The contribution is practical and conceptual, reframing prompt engineering as a governable audit competency.
- Quality assurance
- AI policy
Research
Deterministic Governance Is Multi-Dimensional: Beyond Authorization in AI Systems
Edward Meyman
Zenodo (CERN European Organization for Nuclear Research) · 2026-08-23
This paper proposes the Multi-Dimensional Deterministic Governance Model, a vendor-neutral framework for evaluating AI governance architectures in regulated environments such as healthcare, financial services, and defense. It argues that a single authorization gate is insufficient for true governance, and instead specifies five assurance functions—governance ontology, constraint codification, deterministic decision logic, authorization artifact, and execution enforcement—that must all compose at the runtime authorization boundary for an architecture to claim deterministic governance. The framework defines deterministic governance as producing identical verdicts for identical governed state while emitting verifiable artifacts enabling independent third-party replay and verification. It is intended as a diagnostic evaluation tool for enterprise buyers, regulators, CISOs, and governance architects assessing AI governance infrastructure.
- Enterprise
- AI policy
Research
AI for tumour proportion scoring of programmed death‐ligand 1 immunohistochemistry in non‐small cell lung cancer: a review of commercial and non‐commercial tools
Joachim Webers, Sandrina Martens, Pascal Gervois et al.
Histopathology · 2026-08-23
This review examines AI tools designed to score PD-L1 immunohistochemistry tumour proportion scores (TPS) in non-small cell lung cancer, comparing seven commercial and 13 non-commercial tools across six analytical levels. The authors find that evidence is highly heterogeneous—varying in cohort composition, assay-scanner settings, reference standards, and threshold reporting—making direct comparison and universal ranking impossible. Commercial tools more often documented workflow integration and certification status, while non-commercial tools better described model architecture and experimental design. The review concludes that safe routine deployment requires more transparent reporting, prospective multi-centre validation, and ongoing quality monitoring, and recommends local verification over universal rankings.
- Quality assurance
- Certifications
Research
AI-ASSISTED REWRITING AND IDEA LAUNDERING IN ACADEMIA: A PROOF-OF-CONCEPT STUDY FROM THE PERSPECTIVE OF PUBLICATION CONTROLS
Yusuf Mert Velioğlu, Hakan Velioğlu, Selçuk Olum et al.
Denetişim · 2026-08-23
This proof-of-concept study tests whether Turnitin's similarity and AI-detection reports can catch 'idea laundering'—the practice of rewriting another author's core argument through AI assistance without attribution and presenting it as original. Using 36 files derived from 12 social science articles, the study found that naive LLM rewriting dropped average Turnitin similarity from 99.75% to 44.50%, and structured rewriting dropped it further to 19.08%, while independent raters confirmed the underlying ideas were largely preserved across all cases. The authors conclude that current automated publication controls are insufficient to detect idea laundering and call for a multi-layered governance approach combining semantic evaluation, argument traceability, and editorial judgment.
- Quality assurance
- AI policy
Research
Closed-loop AI achieves certifiable engineering design
Tianyi Yu, Chengxing Tao, Haoxuan Shen et al.
arXiv (Cornell University) · 2026-08-22
This paper introduces 'The AI Engineer,' an agentic AI framework that couples large language models to deterministic engineering solvers in a closed loop to autonomously design floating offshore wind structures. The system uses topology optimization and particle swarm optimization against physics-based load cases, and an Automated Reviewer scores each candidate design across five engineering dimensions. Validation by the China Classification Society's Approval in Principle process confirmed that the top AI-generated design met professional standards, outperforming a human-optimized baseline by reducing steel mass and unit capital cost each by 8.1%. This work demonstrates that closed-loop AI with physics-grounded, codified constraints can produce certifiable engineering designs at reduced cost.
- Certifications
- Enterprise
Research
Dissecting Neuro-Symbolic Quality Assurance for Synthetic Oncology Data Generation
Laxmigayathri Challa, Yuhan Zhou, Ana Cleveland et al.
arXiv (Cornell University) · 2026-08-22
This paper investigates how individual quality-assurance components in neuro-symbolic pipelines affect the validity of synthetically generated oncology clinical data. Through three controlled studies, the authors find that schema validation is the primary filter—rejecting roughly 29% of ungated records for schema failures and 20% for invalid staging—while clinical-logic validation is largely redundant and retrieval augmentation effects are strongly model-dependent. The results show that symbolic gating improves clinical validity of synthetic cancer-staging records but does not improve vocabulary richness, and that ontology density should not be used as a proxy for corpus quality. These findings provide concrete, actionable guidance for designing QA pipelines in synthetic medical data generation.
- Quality assurance
Research
HIRA: A Human-in-the-Loop Retrieval-Augmented Cascade for Document Classification in Regulated Industries
Shangxuan Tian, Yanhui Chen, Carlos Queiroz
arXiv (Cornell University) · 2026-08-22
HIRA is a training-free document classification system designed for regulated industries that combines BM25, dense text embeddings, and image-level representations through a cascading pipeline: confident documents are classified by retrieval alone, uncertain ones are routed to a locally hosted large language model verifier, and remaining edge cases go to human reviewers whose corrections are stored as retrieval exemplars. On an 80-class trade-finance corpus of 30,233 documents, HIRA improved Macro-F1 from 0.6218 to 0.8548 while requesting human correction for only 6.4% of documents; on the Tobacco-3482 benchmark it reached 0.9423 Macro-F1, outperforming a zero-shot LLM baseline by 17.4 percentage points. The system addresses practical regulated-industry constraints such as data residency, limited labeled data, and costly model-governance procedures without requiring any model weight updates. These results suggest that selective human feedback combined with retrieval-memory adaptation can substitute for repeated model retraining in long-tail classification settings.
- Enterprise
- Quality assurance
Research
Applicability of AI in Due Diligence Process of Mergers and Acquisitions: An Exploratory Study
Kartik Kumar Satvedi, Sushma Rani
International Journal of Engineering and Management Research · 2026-08-22
This exploratory study examines how liability should be assigned under Indian law when an AI-assisted due diligence tool fails to detect a critical risk in a merger or acquisition. The paper analyzes relevant provisions of the Companies Act 2013, the Indian Contract Act 1872, professional negligence principles, and SEBI's Takeover Code, finding that the existing framework is fragmented and that no rule specifies the level of human oversight required for AI tool outputs to meet the standard of care. Even the EU's Product Liability Directive 2024/2853, often cited as a reform model, is found inadequate because it excludes pure economic loss suffered by corporate claimants. The authors recommend a contract-based AI liability clause for engagement letters, a targeted SEBI circular, and clearer standards from professional regulatory bodies.
- AI policy
- Enterprise
Research
AI-Driven Operational Intelligence and Workforce Reskilling in U.S. Hotels
Salami Abdul Mohammed
Journal of Social Science and Human Research Studies · 2026-08-22
This paper reviews how AI-driven operational intelligence systems—covering revenue management, scheduling, procurement, guest personalization, and business intelligence dashboards—are reshaping skill requirements for hotel managers and staff across U.S. hotel operations. Using a systematic narrative literature review grounded in Human Capital Theory and the Technology-Organization-Environment framework, the authors find that barriers to workforce reskilling are primarily organizational and environmental rather than a shortage of available training tools. The paper proposes a three-level reskilling framework tailored to hotels of different scales and argues that failure to close the analytical skills gap carries consequences not just for individual properties but for national workforce development across an industry employing 2.17 million workers.
- Workforce
Research
Artificial Intelligence and Workforce Preparedness in India: A Human-Capital and Sociotechnical Perspective on Labour-Force Adaptation
Ritika, Manju Dahiya
Journal of Asia Entrepreneurship and Sustainability · 2026-08-22
This study surveys 600 Indian workers across five sectors to assess how prepared the Indian labour force is to adapt to AI-driven employment changes. Using human capital theory and a sociotechnical framework, it finds a moderate mean preparedness score (3.807/5) with workforce preparedness strongly correlated with AI adoption (r=.817) and AI economic impact (r=.810), explaining 61.5% of variance in the regression model. Preparedness varies significantly by gender, age, job level, and experience but not meaningfully by sector. The authors conclude that AI readiness is a jointly produced capability requiring continuous learning, workplace reskilling, AI literacy, and inclusive institutional policy support—particularly important given India's large and educationally diverse working-age population.
- Workforce
- AI policy
Research
Grading‐labour cost analysis of AI ‐assisted versus human‐only diabetic retinopathy screening in two Danish healthcare settings
Lars Grønlykke, S C Wiberg, Alex Carter et al.
Acta Ophthalmologica · 2026-08-22
This cost-minimisation analysis compares AI-assisted versus human-only diabetic retinopathy screening labour costs across two Danish healthcare settings. In tertiary diabetes centres, AI assistance reduced annual grading-labour costs by 33% (€282,192 vs €421,425), while in private ophthalmology practices the reduction was 62% (€569,692 vs €1,502,698). Savings stemmed primarily from reduced ophthalmologist time on cases the deep-learning model classified as DR-negative, and the AI-assisted strategy remained cost-saving across all sensitivity analyses. The study highlights that reimbursement and tariff structures must be aligned with efficiency gains for health systems to fully capture AI-driven savings in routine screening programmes.
- Enterprise
- AI policy
Research
Delays between CE mark and FDA regulatory approval of AI-enabled software for radiology
Yijun Ren, Daniel Windecker, Isaac Shiri et al.
npj Digital Medicine · 2026-08-22
This study examined 239 AI-enabled radiology software products to characterize how approval timelines differ between the EU (CE marking) and the US (FDA clearance). Among devices that received both authorizations, those that obtained CE marking first waited a median of 17.5 months before receiving FDA clearance, compared to just 3.5 months for the reverse sequence. Factors like EU Class IIa classification were associated with faster dual-authorization, while radiograph interpretation software faced longer delays. The authors conclude that a persistent transatlantic regulatory asymmetry exists and call for greater cross-jurisdictional coordination and transparency.
- AI policy
- Certifications
Research
Training on the Utilization of Artificial Intelligence for Optimizing MSME Product Marketing on E-Commerce Platforms
Nurmalitasari, Nurchim, Safina Callistamalva Arindrajaya et al.
International Journal Of Community Service · 2026-08-22
This community service study trained 30 micro, small, and medium enterprise (MSME) representatives in Karanganyar Regency on using AI tools for e-commerce marketing, measuring outcomes across four competency domains via pre- and post-tests on a 0–100 scale. Average scores rose from 52.00 to 78.25—a 50.5% relative improvement—with marketing content creation showing the largest gain (50 to 82) and AI tool use the smallest (55 to 70). The authors note these are descriptive findings only and do not constitute evidence of statistical significance or long-term business impact. The results support practice-oriented AI training as a means of strengthening MSME digital marketing capabilities, while recommending follow-up mentoring to sustain outcomes.
- Workforce
- Enterprise
Research
Beyond velocity: a causal quality- and security-sensitive schedule performance index for AI-assisted software projects
Mohammad Tanhaei
Scientific Reports · 2026-08-22
This paper proposes CQSS-SPI, a new scheduling metric for AI-assisted software projects that corrects the traditional Schedule Performance Index by incorporating quality debt and security penalty factors within a structural causal model. Applied to an eight-sprint retrospective case study of the BioArc hospital information system, the framework revealed that sprints appearing on-schedule or ahead (raw SPI up to 1.127) were actually underperforming once hidden quality and security liabilities were accounted for, with average performance dropping by 10.3 percentage points. The causal model also supports counterfactual analysis of governance choices such as stricter security gating, showing 3.2–4.6 percentage point improvements under stronger alternatives. The authors note that results are from a single system and that multi-domain validation remains future work.
- Quality assurance
- Enterprise
Research
Who Bears the Cost of Honesty? A FAccT Workshop Synthesis and Research Agenda for Equitable AI Disclosure
Runlong Ye, Jessica He, Finola Finn et al.
arXiv (Cornell University) · 2026-08-21
This paper synthesizes findings from a participatory workshop at the 2026 ACM FAccT conference that examined who bears the social and professional costs of disclosing AI use. Using scenario-anchored power mapping and design fiction, participants identified harms including suspicion, stigma, competence penalties, and surveillance that AI disclosure norms can impose, particularly on minoritized groups. Key concerns include workers being penalized in performance evaluations for using accessibility-related AI tools. The paper proposes a diagnostic framework called the 'Cost-of-Honesty Stack' alongside design suggestions and research directions for more equitable AI disclosure.
- Workforce
- AI policy
Research
Generative AI in Brand Activism: Impacts on Consumers’ Negative Affect and Decision Comfort
Zhao Lin, Alexis Yim, Annie Peng Cui
Journal of Consumer Affairs · 2026-08-21
This paper examines how generative AI-created brand activism messages affect consumers emotionally and behaviorally. Across five studies, the researchers find that AI-crafted sociopolitical brand messages trigger greater negative affect and reduce decision comfort compared to human-created messages, primarily because consumers perceive them as less authentic. The effect is weaker among politically conservative consumers and those who tend to anthropomorphize AI. The findings offer practical guidance for brands navigating AI use in value-laden communications.
- Enterprise
Research
Counterfactual, Per-Decision Bias Auditing for Automated Hiring: Localizing and Explaining Disparate Impact in Applicant Tracking Systems
Jay Barach
arXiv (Cornell University) · 2026-08-21
This paper introduces the AI Bias Firewall (AIBF), a method that audits automated applicant tracking systems one hiring decision at a time by neutralizing protected-attribute proxies, re-scoring each decision, and measuring the resulting counterfactual shift. Evaluated on the Adult and COMPAS datasets, AIBF identifies bias-flipped decisions with an AUC of 0.963 and is precise enough that reviewing only 5% of decisions surfaces 55% of harmed candidates, compared to 6% under group-based review. The authors also flag a key limitation: correcting flagged decisions improves but does not fully restore legal parity because so-called merit features still carry residual proxy correlation. The tool is open-sourced under Apache 2.0, making it directly relevant to regulatory demands for auditable hiring systems.
- Workforce
- AI policy
- Certifications
Research
How Brussels reproduces Silicon Valley technosolutionism: Sociotechnical imaginaries in the EU’s regulatory approach to AI (2019‒2025)
Álvaro Oleart, Alejandro Flores Moleon
Internet Policy Review · 2026-08-21
This paper analyzes the sociotechnical imaginaries—shared visions of desirable futures and risks to be avoided—embedded in the EU's major AI regulatory documents from 2019 to 2025, including the 2024 AI Act, the 2025 GPAI Code of Practice, and the 2019 Ethics Guidelines for Trustworthy AI. The authors find a structural tension between public regulatory oversight and Big Tech's growing co-stewardship role, and argue this arrangement reproduces Silicon Valley-style technosolutionism rather than offering a distinctly European democratic alternative. The study highlights risks of regulatory capture and over-reliance on private infrastructure, and calls attention to how responsibilities are distributed between EU institutions and technology platforms. The findings matter for AI governance debates about who legitimately governs AI safety and what democratic alternatives exist to industry-led regulation.
- AI policy
Research
Prediction certification cannot replace explanation certification: a competence envelope for trustworthy AI under compound stress
Nataliya Shakhovska, Ivan Izonin, Stergios-Aristoteles Mitoulis
arXiv (Cornell University) · 2026-08-21
This paper proves that prediction-based AI certification—covering accuracy, calibration, and conformal coverage—is mathematically insufficient to establish trustworthiness. The authors demonstrate a separation theorem showing that a reliable and a compromised model can be indistinguishable under every prediction-side certificate while differing arbitrarily in how they make decisions. They introduce the 'competence envelope,' a framework combining prediction and explanation certification, which empirically reveals failure modes invisible to prediction-only checks across diverse datasets and model classes.
- Certifications
- Quality assurance
Research
Vibe Coding and Web Application Security: A Twin-Prompt Study
Darko Andročec
arXiv (Cornell University) · 2026-08-21
This paper examines whether adding explicit security requirements to natural-language prompts improves the security of AI-generated web applications. Across six web applications each generated in a baseline and a security-aware variant by the same agentic coding assistant, the security-aware prompts produced fewer confirmed vulnerabilities (24 versus 51) and eliminated all Critical and High severity findings. The authors caution that the corpus is small and results are descriptive rather than statistically validated, positioning this as a preliminary study. The findings matter for quality assurance practitioners and developers relying on AI coding tools, suggesting prompt design meaningfully affects security outcomes.
- Quality assurance
- Enterprise
Research
Clinical decision support in nasopharyngeal carcinoma: comparative evaluation of large reasoning and language models
Lihong Wang, Luxun Wu, Feng Jiang et al.
npj Digital Medicine · 2026-08-21
This study evaluated five AI models—three large language models (LLMs) and two large reasoning models (LRMs)—on 50 open-ended clinical questions covering nasopharyngeal carcinoma (NPC) management, scored by radiation oncologists. LRMs (Grok 3 Think and Deepseek-R1) outperformed LLMs overall, with Grok 3 achieving the highest accuracy (84.0%) and relevance (91.6%), while all models showed limitations including hallucinations and outdated content. The findings suggest LRMs have potential as assistive clinical decision-support tools but require rigorous validation before deployment in practice.
- Quality assurance
- Enterprise