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
Coverage-Driven Verification for Safety-by-Design in AI-Based Collision Avoidance Systems
Thomas Stefani, Johann Maximilian Christensen, Elena Hoemann et al.
arXiv (Cornell University) · 2026-08-21
This paper proposes a structured engineering method for assessing whether the data distributions used to train and verify AI/ML-based collision avoidance systems adequately represent their intended operational design domain (ODD), in line with EASA learning assurance requirements. The authors evaluate statistical measures—finding the chi-squared test unsuitable for large datasets—and adopt Kullback–Leibler divergence and Cramér's V as quantitative representativeness metrics. The method is demonstrated on experimental simulation data from Horizontal and Vertical Collision Avoidance Systems (HCAS/VCAS), showing how distribution comparison can support safety assurance for AI in aviation. The work contributes toward a systematic Safety-by-Design AI engineering process aligned with emerging EASA guidance.
- Certifications
- Quality assurance
Research
CertVLA: Certified Defense against Physical Visual Attacks for Vision-Language-Action Models
Hui Lu, Zhijie Peng, Yuqi Lin et al.
arXiv (Cornell University) · 2026-08-21
CertVLA introduces a certified defense framework for Vision-Language-Action (VLA) models against physical visual attacks such as localized patches and texture perturbations. The system uses calibrated covering masks to ensure at least one prediction remains attack-free, then conjoins query-level decisions across a full closed-loop rollout to provide a provable action certificate. The authors prove that any rollout certified by CertVLA executes actions consistent with clean, attack-free predictions, and that this consistency guarantee can further imply task success under dual-mask correctness conditions. Experiments in simulation and real-world settings validate both the empirical and certified effectiveness of the approach.
- Certifications
- Quality assurance
Research
Can Legal AI Know When It Is Wrong? And Do Students Know When It Is?
Angel Mary John, Vipin Singh, Jerrin Thomas Panachakel
arXiv (Cornell University) · 2026-08-21
This paper audits three major LLMs—ChatGPT, Meta AI, and Perplexity AI—on 60 Indian Contract Act legal cases, finding that all models frequently deliver incorrect verdicts with near-maximum confidence, a phenomenon the authors call 'inertia of confidence.' Meta AI showed the highest High-Confidence Error Rate (31.7%), meaning nearly a third of its wrong answers came with dangerous certainty. A survey of 380 Indian law students revealed that most learned to verify AI outputs only after encountering hallucinated citations, and 71.1% had received no formal training on ethical AI use despite 81.6% knowing that submitting fabricated cases could constitute contempt of court. The authors argue this gap between AI overconfidence and students' reactive verification habits creates systemic professional negligence risks, calling for adversarial legal research pedagogy and source-grounded verification architectures.
- Workforce
- AI policy
Research
The Rising Cost of Trust: Practitioners' Trust Signals, Controls, and Responses in the Software Supply Chain
Ranindya Paramitha, Siri Paidipalli, Laurie Williams et al.
arXiv (Cornell University) · 2026-08-21
This study examines how software practitioners navigate trust in the software supply chain, including the emerging threat of AI-generated dependencies, through semi-structured interviews with 38 industry and open-source practitioners. Using thematic analysis grounded in social-science trust concepts, the researchers find that trust is eroding and becoming increasingly costly, with aware practitioners accumulating verification controls. To manage these rising costs, practitioners automate verification, delegate trust decisions to designated 'guardians,' or contemplate exiting the supply chain altogether. The findings offer a conceptual vocabulary—signals, guardians, system trust—intended to inform future interventions for a healthier software supply chain.
- Quality assurance
- Enterprise
Research
Large Language Models at the Intersection of Software Engineering and Software Security:An Evidence-Centered Structured Survey and Research Agenda
Wei Lin, Tao Zhou, Zhaofei Xie et al.
arXiv (Cornell University) · 2026-08-21
This structured survey synthesizes research on Large Language Models applied to both software engineering and software security tasks, covering work through May 2026. It finds that execution feedback and repository access improve engineering task completion but do not alone establish security, while vulnerability-classification scores rarely establish deployable correctness. The authors introduce an assurance framework separating functional correctness, security, operational reliability, evidence provenance, and agent authority, and identify recurring validity threats such as weak test oracles, data leakage, and proxy-only security checks. The paper concludes that LLM capability should be judged as an assurance case supported by task-appropriate evidence rather than a single benchmark score, and proposes a minimum reporting protocol and research agenda for jointly secure-and-functional evaluation.
- Quality assurance
- Certifications
Research
HelixCross – Design and development of an AI-powered cross-examination simulation for forensic expert testimony training
Mark Barash, Bruce Budowle
Forensic Science International Synergy · 2026-08-21
HelixCross is an alpha-stage AI tool that uses a prompt-engineered large language model with retrieval-augmented generation to simulate defense counsel cross-examination and provide structured coaching for forensic expert witnesses. It targets recurring problems in DNA testimony—such as likelihood-ratio misinterpretation, probabilistic-genotyping assumptions, and unwarranted certainty—through a phase-based curriculum, a synthetic DNA case, and a ten-dimension coaching rubric. The paper describes its architecture and training assets but is transparent that evaluation so far is limited to informal functional testing without controlled outcome data, and that platform-dependency risks have already emerged. It represents a scalable alternative to the current informal, mentor-dependent model of testimony preparation, though a formal evaluation phase is needed to demonstrate its effectiveness.
- Workforce
- Certifications
Research
Chat First, Worry Later: Understanding Individuals' Privacy Perceptions Using ChatGPT in a Work Context
Christoph Nirschl, Magdalena Glas, Gerhard Messmann et al.
arXiv (Cornell University) · 2026-08-21
This study surveyed 224 workers across multiple employment sectors who use ChatGPT in their jobs, examining how privacy concerns, ChatGPT proficiency, and organizational GenAI policies interact to shape actual usage behavior. The findings show that organizational policies are significantly associated with higher privacy-related proficiency, but overall proficiency remains low. Higher privacy concerns reduce both how often workers use ChatGPT and how diversely they apply it, particularly in organizations lacking formal GenAI policies.
- Workforce
- AI policy
Research
Beyond Sentiment: Context-Aware Emotion Detection in Guest Feedback for Intelligent Hospitality Systems
Srinivasan
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This paper introduces a context-aware emotion detection framework for analyzing hotel guest feedback, going beyond binary positive/negative sentiment to classify eight distinct emotions (joy, trust, fear, surprise, sadness, disgust, anger, and anticipation) across 50,000 hotel reviews. Using transformer-based language models with contextual embeddings, the system achieves 87.3% accuracy and outperforms general-purpose models like RoBERTa, DeBERTa-v3, and DistilBERT on micro-F1 and macro-F1 metrics after domain-adapted fine-tuning. A key finding is that 34% of apparently positive reviews contain underlying negative emotions, demonstrating the inadequacy of conventional sentiment analysis for hospitality use cases. The paper also addresses practical deployment requirements including latency, monitoring, and governance, making it directly relevant to enterprise hospitality AI systems.
- Enterprise
- Quality assurance
Research
Capable but Not Deployable: Institutional Constraints on AI Exposure in Finance
Claes Backman, Christos A. Makridis
CESifo · 2026-08-21
This paper argues that standard AI exposure metrics—which ask language models to rate how technically feasible it is for AI to perform O*NET tasks—miss a critical layer in regulated industries: institutional deployability. Studying 2,199 tasks across 99 finance-and-insurance occupations and scoring them with eight frontier models, the authors find that adding finance-specific regulatory context reduces AI exposure scores by roughly one-fifth compared to bare technical feasibility scores, with the largest gaps in regulated, client-facing roles. Crucially, when they map these scores to publicly traded firms, only the institutional 'markdown' predicts stock returns: firms with high regulatory friction underperform those with low friction by about 25 percentage points in cumulative abnormal returns over three years post-ChatGPT, while technical feasibility scores alone predict nothing. The findings suggest that for regulated industries, deployable exposure—accounting for compliance, supervision, and accountability requirements—is a more relevant measure of AI's economic impact than raw capability.
- Enterprise
- AI policy
Research
AI-assisted, instructor-supervised grading and feedback in higher education: Design and evaluation of an end-to-end pipeline
Leonardo Franco Cruz, Miguel Mira da Silva, Henrique S. Mamede
Computers and Education Open · 2026-08-21
This paper designs and evaluates a GPT-4o-based pipeline that automates rubric-linked grading and formative feedback for a Master's-level course, with instructors retaining official grading authority. Across 362 submissions from 66 students, the AI matched the instructor's grade within 0.5 points on a five-point scale in 83% of cases (MAE = 0.31), though chance-corrected agreement was only moderate (ICC = 0.49). About 80% of surveyed students rated the AI feedback useful, and the system cost EUR 0.55 per student per assignment with feedback delivered within 24 hours. The findings support LLMs as scalable, low-cost supplements to instructor evaluation for formative assessment, while cautioning against replacing human graders given moderate agreement levels.
- Quality assurance
- Workforce
Research
An AI-Driven Privacy-Preserving Framework for Automated Legal Hold, Data Preservation, and Enterprise Data Compliance
Akhilesh Gundaboina
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This paper presents PALHI (Privacy-Aware Legal Hold Intelligence), an AI framework that automates legal hold and data compliance processes for enterprise organizations managing electronically stored information. Using transformer-based semantic classification, named-entity recognition, federated learning, and differential privacy, the system achieves 92.40% accuracy and a 93.18% legal-hold detection rate on an enterprise email and document corpus, with only a 2.16% accuracy loss under differential privacy constraints. The framework replaces largely manual, rule-based compliance workflows with an auditable, scalable system that reduces compliance overhead while preserving data confidentiality. This matters for enterprises facing growing pressure from privacy regulations and litigation-driven data preservation requirements.
- Enterprise
- AI policy
Research
No Judgment Without a Reason: Counterfactual Receipts for Versioned AI Evaluators
Ye Chen, Weining Zhang
arXiv (Cornell University) · 2026-08-21
This paper addresses a critical flaw in AI evaluators: they often produce correct labels through flawed or unjustified reasoning, which is dangerous in agentic systems that gate actions or supply training feedback. The authors formalize evaluator accountability through a framework of grounds, norms, and authority, introduce 'judgment receipts' as minimal source-replacement sets that explain verdict changes, and release ReasonBench—a benchmark covering 19,520 cases. Experiments show that while models like Qwen3-1.7B achieve high standard accuracy, robustness collapses under meaning-preserving permutations and multi-source updates, revealing that standard accuracy alone is insufficient for trustworthy evaluator auditing.
- Quality assurance
- Certifications
Research
Deep Learning-Based Detection of Deceptive UI Patterns in Web Applications Using Convolutional Neural Networks
Intisar Hadi Abbas
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This paper presents a CNN-based system that automatically detects and classifies deceptive UI patterns (dark patterns) in web application screenshots across six deception categories. Trained and evaluated on a curated dataset of 4,800 annotated screenshots, the proposed model achieves a macro-averaged F1-score of 0.914 and AUC of 0.971, outperforming baselines including VGG16, ResNet50, EfficientNetB3, SVM, and Random Forest. Grad-CAM visualizations confirm the model focuses on semantically meaningful elements such as pre-ticked checkboxes and obfuscated pricing. The authors position the framework as a foundation for scalable browser-based auditing tools and regulatory compliance pipelines targeting deceptive design practices.
- AI policy
- Quality assurance
Research
Cybersecurity Talent Gap Prediction Models for Financial Infrastructure Protection
Pratik Mahajan
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This paper proposes a predictive modeling framework for forecasting cybersecurity talent shortages in financial market infrastructures (FMIs) before gaps become critical. The ensemble model—combining time-series baselines, gradient-boosted trees, and sequence models—draws on threat intelligence, vacancy aging, incident telemetry, HR data, and the NIST NICE workforce taxonomy to score staffing shortfalls role by role. On a synthetic 6-month holdout dataset mirroring an FMI environment, the model reduced shortage-forecasting error by 23.8% over a seasonal naive baseline and achieved an F1-score of 0.842 and ROC-AUC of 0.93 for classifying critical shortages. The work matters because it shifts cybersecurity workforce planning in financial institutions from reactive vacancy-filling to proactive, data-driven hiring, reskilling, and sourcing decisions, and it delivers a reference architecture and deployment roadmap for practical adoption.
- Workforce
- Enterprise
Research
A Policy-Aligned Agentic RAG Framework for Risk-Aware Decision Support in Enterprise Customer Relationship Management
Chitrapradha Ganesan
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This paper proposes PAL-CRM-RAG, a multi-layer agentic retrieval-augmented generation framework designed to embed policy validation, risk-aware escalation, and decision auditability directly into enterprise CRM systems. Evaluated on a synthetic corpus of 20,000 CRM tickets and 30 policy documents, the framework achieves 100% policy compliance and a perfect escalation F1 of 1.000, compared to 0% compliance for a non-RAG baseline, while risk classification reaches 95.1% accuracy. Ablation results confirm that removing the policy validator drops compliance to zero and removing the risk classifier eliminates escalation capability entirely. The findings demonstrate that generative AI can be operationalized responsibly in customer service environments where AI recommendations directly affect customer rights and privacy-sensitive decisions.
- Enterprise
- AI policy
- Quality assurance
Research
A Systems Engineering Framework for Real-Time User Error Prediction and Adaptive User Experience Control Abstract
Mykola Savenko
Journal of Engineering Management and Systems Engineering · 2026-08-21
This paper presents a systems engineering framework that integrates behavioural sensing, machine learning-based error prediction, and adaptive user experience (UX) interventions into a single closed-loop architecture. A four-week controlled crossover experiment with 84 users and over 161,000 interaction events found that a transformer model achieved an AUC of 0.941 in predicting user errors in real time, while model-triggered interventions reduced task-level error rates from 0.280 to 0.160 and cut task completion time from 145.2 s to 117.8 s. Usability and experience scores (SUS, UEQ, NPS) improved substantially and perceived workload (NASA-TLX) dropped by 11.8 points, all with end-to-end system latency under 100 ms. The findings demonstrate that predictive error control can be embedded in digital systems without disrupting real-time interaction, offering a validated basis for designing adaptive interfaces.
- Enterprise
- Quality assurance
Research
Algorithmic Personalization in Digital Marketing and Personal Data Protection: Legal Limits of Consent, Profiling, and Targeted Advertising
Gladys Gelves Suárez
Journal of Intelligent Decision Making and Information Science · 2026-08-21
This legal analysis examines how EU regulations—including GDPR, the ePrivacy Directive, the Digital Services Act, and the AI Act—govern algorithmic personalization and targeted advertising in digital marketing. Drawing on regulatory guidance from the EDPB and CJEU case law, the article finds that consent-based frameworks suffer from structural deficiencies such as information asymmetry, consent fatigue, and commodification of personal data, while profiling rules remain insufficiently constrained due to ambiguities in automated decision-making and weak explainability requirements. The authors propose a rights-based regulatory framework emphasizing purpose limitation, data minimization, algorithmic transparency, and human oversight as alternatives to the notice-and-consent paradigm. The findings are relevant to ongoing policy debates about balancing digital marketing innovation with fundamental rights protection.
- AI policy
- Enterprise
Research
From Deepfake Detection Research to Public Policy: A Survey on Generalization Challenges in Synthetic Media for the Brazilian Electoral Context
Thauan de Souza Tavares Da Silva, Diego Addan Gonçalves, David Menotti
Journal on Interactive Systems · 2026-08-21
This survey examines why deepfake detection systems struggle to perform reliably outside controlled lab settings, with a focus on Brazil's electoral environment. The authors frame the problem as socio-technical rather than purely algorithmic, highlighting gaps in generalization due to distribution shifts, cultural variability, and adversarial adaptation. The paper connects these technical shortcomings to concrete policy implications for regulation, platform governance, and automated decision-making in Brazilian electoral contexts. It argues that addressing synthetic media threats requires aligning detection research with trust, interpretability, and institutional accountability.
- AI policy
- Quality assurance
Research
Understanding Time Availability and Format Preferences for AI Professional Development in Health Professions Education
Angie Garcia, Bethany Grubb, Yulia Piller et al.
Journal of CME · 2026-08-21
This cross-sectional survey of 175 health professions educators found that more than half lacked formal AI training, yet most wanted practical, flexible professional development on AI literacy and responsible use. Participants preferred short on-demand videos and in-person sessions of 30 minutes to one hour per week, with real-world examples and curriculum-integrated projects as top priorities. The findings suggest that AI professional development programs for health educators should be concise, applied, and institutionally supported, with a strong emphasis on ethical practice and critical evaluation of AI-generated outputs.
- Workforce
- Certifications
Research
Supplementing Korean artificial intelligence ethical principles: A five-year update
Un-Kon Lee
PLoS ONE · 2026-08-21
This study synthesizes 37 prior works to extract 17 AI ethical principles and evaluate how their prioritization has shifted before and after 2020. Using Importance-Performance Analysis, the authors find that principles like transparency, robustness, and prohibition of infringement should be continuously maintained, while responsibility, wellbeing, human rights, controllability, and sustainability need greater emphasis. The study proposes expanding Korea's existing 10 AI ethical principles to 14 and calls for urgent governance frameworks addressing privacy, safety, and data management. The findings provide a structured foundation for developing responsible, trustworthy, and safe AI systems at a national policy level.
- AI policy
Research
INTEGRATING MACHINE LEARNING WITH OCCUPATIONAL INCIDENT ANALYTICS: EMERGING FRAMEWORKS FOR NATIONWIDE WORKPLACE SAFETY ENHANCEMENT
Ayomide Solomon Alao
Magna Scientia Advanced Research and Reviews · 2026-08-21
This narrative review synthesizes U.S.-focused research (2021–2026) on machine learning applications in occupational safety, covering incident classification, text processing of accident narratives, computer vision, and return-to-work forecasting. The authors find that while ML models have reached high computational performance, virtually no published studies measure prospective reductions in actual workplace injury or illness rates, revealing a profound gap between algorithmic precision and real worker safety outcomes. A key structural barrier identified is severe data fragmentation across federal enforcement registries, statistical surveys, sector-specific databases, and state workers' compensation systems. The paper proposes an integrated nationwide framework featuring federated data interoperability, risk-calibrated algorithmic standards, mandatory prospective evaluation protocols, and a phased move toward binding administrative regulation.
- Workforce
- AI policy
Research
Governance and data readiness as prerequisites for digital twin adoption in small and midsize cities
Alence Poudel, Emily Moore
Discover Cities · 2026-08-21
This paper proposes the 'Foundation First' framework, a four-phase readiness pathway designed to help small and midsize cities responsibly adopt Digital Twin (DT) and AI technologies in local government. Rather than treating governance, data quality, cybersecurity, and public trust as afterthoughts, the framework positions them as prerequisites for deployment, sequencing phase advancement through demonstrable data integrity, institutional alignment, audit practices, and governance controls. Developed via structured narrative synthesis and illustrated through the City of Sugar Land, Texas, the framework also maps phased municipal readiness to selected UN Sustainable Development Goals. The work reframes DT adoption as a governance- and data-centered challenge rather than a purely technical one, offering practical guidance for municipal leaders assessing their actual preparedness for advanced urban intelligence systems.
- AI policy
- Enterprise
Research
A systematic review of human vs machine intelligence and ethical tensions in human resource management
Karen Saad Zakaria, Mohammad Nisar Khattak
Discover Artificial Intelligence · 2026-08-21
This systematic review of 84 peer-reviewed articles (2018–2024) maps how AI is being applied across core HR functions—particularly recruitment, selection, and performance management—and identifies nine major ethical issues including bias, transparency, and accountability. The authors develop an integrative governance framework drawing on maturity model, legitimacy, and accountability theories to guide responsible AI use in HRM. The study provides actionable guidance for HR professionals, AI developers, and policymakers on implementing ethical AI governance. It advances the field by shifting focus from AI adoption to structured oversight of algorithmic decision-making in workforce contexts.
- Workforce
- AI policy
Research
Public awareness, trust, perceived usefulness, and willingness to use ChatGPT for healthcare communication among adults in the Asir Region, Saudi Arabia: a cross-sectional study
Geetha Kandasamy, Khalid Orayj, Rayah Asiri et al.
Frontiers in Public Health · 2026-08-21
This cross-sectional study of 429 adults in the Asir Region of Saudi Arabia found that 70% had prior ChatGPT experience and 61.77% were willing to use it for healthcare communication, with higher awareness (β=0.41) and trust (β=0.36) positively predicting willingness while risk perception negatively predicted it (β=−0.12). Awareness was moderate overall but higher among younger and more educated participants, and perceived risks included misinformation, privacy concerns, and overreliance. The findings suggest that improving AI literacy and building trust are key prerequisites for responsible public adoption of generative AI tools in healthcare settings, alongside regulatory oversight to address safety concerns.
- AI policy
- Workforce
Research
Perceptions of artificial intelligence in university student affairs management: a quantitative study at Panzhihua University, China
BAI CHUAN WANG, Yang Guo, Muhammad Asghar Khan
Frontiers in Education · 2026-08-21
A quantitative survey of 386 students, faculty, and staff at Panzhihua University finds that 91.5% are aware of the domestic generative AI tool DeepSeek, and respondents rate AI most favorably for administrative functions like student information management and service delivery, while expressing lower confidence in AI for mental health support. The top concerns are data privacy, system transparency, and ethical risks. The authors recommend a phased AI adoption strategy that prioritizes administrative use cases while maintaining human-AI collaboration for sensitive services like mental health, alongside strict data governance frameworks. Limitations include single-university sampling and reliability issues with composite measurement scales.
- AI policy
- Enterprise