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
Occupational AI Substitution Risk, Skill Mismatch, and Individual Income Inequality: Evidence from China
Juanjuan Gan, Mingyi Yan, Rui Sun
Emerging Markets Finance and Trade · 2026-08-25
Using China Family Panel Studies data matched with occupation-level AI substitution risk and O*NET skill-demand measures, this study finds that AI substitution risk has a dual distributional effect on income inequality in China: it compresses income differences within occupations (reducing within-occupation inequality) while widening gaps across occupational groups (increasing between-occupation and overall inequality). The mechanism behind this pattern appears to involve skill mismatch—AI substitution risk is linked to lower skill shortages but higher skill surplus, suggesting technological pressure leads to human-capital underutilization. Heterogeneity analysis shows older and flexible workers are disproportionately harmed by these inequality-enhancing effects, highlighting uneven vulnerability across the labor force.
- Workforce
- AI policy
Research
Tripartite Accountability in AI-Driven Workforce Displacement: A Stakeholder and Legal Perspective
Tobi Nwulu, Omoseni O. Adepoju
Journal of Business and Digital Innovation · 2026-08-25
This paper argues that AI-driven workforce displacement creates a 'responsibility vacuum' in which no single actor—employer, worker, or government—can adequately respond alone. Drawing on case studies from the United States, Germany, Japan, and Singapore and applying Stakeholder and Human Capital Theories, the authors propose a tripartite accountability framework in which all three parties discharge coordinated, simultaneous obligations: employers extend proactive duties beyond legal compliance, workers adapt skills and pursue collective advocacy, and governments close the gap between AI adoption speed and regulatory response. The framework offers structured guidance for employers, policymakers, and worker organizations navigating AI-driven labor market transitions.
- Workforce
- AI policy
Research
Artificial Intelligence Tools and Monetary Policy Analysis and Forecasting in the Central Bank of Nigeria
Elizabeth I. Uzoechie
INTERNATIONAL JOURNAL OF SOCIAL SCIENCES AND MANAGEMENT RESEARCH · 2026-08-25
This study investigates how AI tools—including machine learning, predictive analytics, natural language processing, and sentiment analysis—have affected monetary policy analysis and forecasting at the Central Bank of Nigeria (CBN) from 2015 to 2025. Using a mixed-methods design with 247 valid survey responses and thematic analysis of interviews, the study finds a statistically significant positive relationship (r = 0.607, p < 0.05) between AI adoption and forecasting effectiveness, with over 92% of respondents agreeing that AI tools are extensively applied. Qualitative evidence confirms improvements in forecasting accuracy, speed, and responsiveness, though infrastructural limitations and skill gaps remain barriers. The authors recommend sustained investment in infrastructure and human capital to fully realize AI's benefits in central bank policy functions.
- AI policy
- Workforce
Research
Towards AI-Augmented Public Audit Systems: A Policy and Implementation Roadmap for Us Government Agencies and Multilateral Organizations
Fobellah Abetoh Nyiawung
JOURNAL OF HUMANITIES AND SOCIAL POLICY · 2026-08-25
This paper presents a policy and implementation roadmap for integrating AI technologies—including machine learning, natural language processing, and network analytics—into public sector audit and accountability functions within US government agencies and multilateral organizations. The authors introduce an original conceptual framework called the AI-Augmented Audit Continuum (AIAC) and propose a structured three-phase implementation approach spanning 24 to 48 months. The roadmap addresses technical infrastructure, workforce transformation, governance and ethics, and policy standards, drawing on existing federal and international AI audit implementations aligned with GAO, OMB, and INTOSAI frameworks. The paper concludes with specific recommendations on procurement, workforce development, standards alignment, and interagency coordination to enable responsible AI adoption across the federal audit ecosystem.
- AI policy
- Workforce
- Enterprise
Research
From motorized projectors to generative AI: cultural conflicts over labor automation in Hollywood
Caitlin Petre, Julia Ticona
AI & Society · 2026-08-25
This paper examines historical cases of labor automation in Hollywood—motorized projectors, computer-based editing software, and screenwriting software—to argue that how a technology is understood as 'automating' is a culturally constructed narrative rather than a purely technical determination. Through content analysis of newspapers and trade publications, the authors identify distinct narrative conflicts in each case: projectionists debated professional skill, editors faced accelerated workflows and managerial control, and screenwriters defended barriers to entry. The findings are used to contextualize contemporary debates over generative AI in Hollywood, challenging the notion that its implications for creative labor are entirely unprecedented. The paper matters for workforce policy because it reveals recurring patterns in how workers, management, and the public contest automation narratives.
- Workforce
- AI policy
Research
AI Managerialism in Higher Education: Algorithmic Governance, Shared Authority, and the Ethics of Institutional Control
Sun Tianzi, Mark Joseph D. Pastor
American Journal of Education and Technology · 2026-08-25
This paper introduces the concept of 'AI managerialism' to describe how algorithmic systems are reshaping governance, decision-making, and institutional control in higher education. Through a critical narrative review and cross-case analysis of three cases—the Ofqual algorithm controversy, Purdue's Course Signals system, and the University of Sydney's generative AI response—the study finds that AI tools can improve efficiency and evidence-informed decisions while also creating risks of opacity, surveillance, and centralized authority. In response, the authors propose an Ethical AI Governance Framework built on five principles: mission alignment, transparency, participatory governance, equity auditing, and bounded scope. The framework offers practical guidance for university leaders and policymakers seeking to align AI adoption with academic values and shared governance.
- AI policy
- Enterprise
Research
Artificial Intelligence Readiness for Compliance Risk Management: A Qualitative Case Study of a Vietnamese Maritime Logistics Firm
Võ Thị Thu Hồng, Vũ Kiều Sa
International Journal of Innovative Science and Research Technology (IJISRT) · 2026-08-25
This qualitative case study examines how a Vietnamese maritime logistics firm (TRA-SAS) can build organizational readiness to deploy AI for compliance risk management. Using interviews, internal documents, and a risk-assessment matrix, the study identifies eight material compliance risks—with customs compliance ranked highest—and finds the firm sits at an uneven digital maturity level (M2–M3), where software is more advanced than data standardization and system integration. High-value AI use cases include document checking, contract analysis, and regulatory change monitoring, but all depend on stronger data governance and human oversight. The paper proposes a risk-first, readiness-gated, human-in-the-loop implementation pathway rather than treating AI adoption as a simple technology purchase.
- Enterprise
- AI policy
Research
From Ideation to Growth: The Impact of Generative AI on Entrepreneurial Processes in Qatar
Maha Ibrahim AlNassr, Arshad Jamal, Syed Rizwan Shahid Pirzada
Journal of Business and Digital Innovation · 2026-08-25
A quantitative survey of 106 Qatari start-up leaders finds that perceived usefulness is the strongest driver of generative AI adoption, with adoption positively and significantly linked to innovation outcomes such as idea generation and rapid prototyping. Growth performance benefits are weaker, constrained by contextual factors like market size, funding ecosystems, and regulatory frameworks. The study extends the Technology Acceptance Model to include ethical legitimacy and institutional readiness, and offers practical guidance for entrepreneurs on phased adoption and hybrid human-AI models, as well as policy recommendations around regulation, capacity-building, and trust.
- Enterprise
- AI policy
Research
Pelatihan dan Pendampingan C-Agile Program dalam Mengoptimalkan Adaptabilitas Karier Mahasiswa UQISA Australia
Olievia Prabandini Mulyana, Umi Anugerah Izzati, Ira Khairani Panjaitan et al.
JURPIKAT (Jurnal Pengabdian Kepada Masyarakat) · 2026-08-25
This community service study tested the C-Agile Program—a structured intervention using an experiential action-learning cycle (Model 3C: Compass, Capability, Connection)—on 20 Indonesian students at the University of Queensland Indonesian Students Association (UQISA) in Brisbane, Australia, to address AI-driven career anxiety. Using a pre-post lagged evaluation design with the Career Adapt-Abilities Scale (CAAS) over four weeks, results showed a statistically significant increase in career adaptability scores (from M=96.65 to M=110.45, p<0.001), with the proportion of students in the high-adaptability category rising from 20% to 75%. The findings suggest that structured career-agility training can measurably restore career agency among Generation Z students facing AI-related labor market disruption, with implications for workforce readiness and alignment with SDGs 4 and 8.
- Workforce
Research
The evolution of organisational AI readiness toward an orchestration capability
Krzysztof Jonak, Andrzej Wodecki
Discover Artificial Intelligence · 2026-08-25
This qualitative study of fifteen senior practitioners finds that as AI moves from pilots to enterprise-wide deployment, foundational IT infrastructure and data maturity have become baseline prerequisites rather than competitive differentiators, while governance, accountability, process discipline, and behavioural alignment now drive effective AI deployment. Drawing on socio-technical systems and dynamic capabilities theories, the authors introduce the construct of 'Organisational AI Orchestration Capability'—the dynamic capacity to integrate, govern, and operationalise heterogeneous AI resources into reliable enterprise outcomes. The work reframes AI readiness from a static resource checklist to an evolving organisational capability and offers managers a diagnostic logic for sequencing AI investments and avoiding premature scaling. These findings are directly relevant to enterprises navigating generative AI and AI-as-a-service adoption.
- Enterprise
Research
AI-enabled integrated employment ecosystem for socially vulnerable groups: a multiple case study and design-research approach
Bahl-Geun Roh
Frontiers in Sociology · 2026-08-25
This paper examines how AI can enable inclusive employment for older adults and persons with disabilities, using case studies of Testworks, Kakao, and Microsoft alongside a design-research approach. It identifies three AI-driven mechanisms—strength matching, education-to-employment pathways, and multi-stakeholder collaboration—and finds that none of the studied organizations fully integrates both vulnerable groups within a single system. From this gap, the authors propose a three-stage circular ecosystem model (Education–Matching–Sustainability) supported by a triple-layered revenue architecture, intended as a transferable framework for societies facing demographic and digital transitions. The work reframes AI as a tool for social inclusion rather than displacement, offering both a theoretical contribution and a practical roadmap for addressing compounded labor-market exclusion.
- Workforce
- AI policy
Research
Oncologists' knowledge, attitudes and needs about artificial intelligence in clinical oncology in Luxembourg in 2026: a national cross-sectional survey (AICO study)
Dominic Kaddu-Mulindwa, Xianqing Mao, Caroline Duhem et al.
Frontiers in Digital Health · 2026-08-25
This national cross-sectional survey of oncologists in Luxembourg found that AI adoption is already widespread—all 25 respondents had used large language models—yet 88% had received no formal AI training and the majority identified lack of validated tools and regulatory uncertainty as top barriers. Most physicians believed they would bear primary legal responsibility for AI-related errors, and 76% reported patients bringing AI-generated medical information to consultations, illustrating a significant 'implementation–governance gap.' The findings underscore the urgent need for structured AI education, validated clinical tools, and clearer regulatory frameworks to enable safe and ethical integration of AI into oncology practice.
- Workforce
- AI policy
Research
Mechanisms underlying university students' perceived AI threat: the mediating roles of perceived controllability and perceived effort–reward imbalance
Yue Tong Liang, Juan Du
Frontiers in Psychology · 2026-08-25
This cross-sectional study of 548 Chinese university students in the humanities and social sciences examines why students feel threatened by AI in educational assessment and employment contexts. Using structural equation modeling, it finds that perceived controllability negatively predicts AI threat while perceived effort–reward imbalance positively predicts it, with both serving as mediating pathways through which factors like AI anxiety, AI self-efficacy, growth mindset, and role overload shape threat perceptions. Role overload had the largest total effect on perceived AI threat, and growth mindset influenced threat only indirectly through both mediators. The findings support the case for AI literacy education, psychological adaptation programs, and career development guidance in universities to help students manage AI-related occupational concerns.
- Workforce
- AI policy
Research
Quantifying System-Level Harms from AI Adoption in Complex Sociotechnical Systems
Paul Vautravers, Oliver Chalkley, Gabriel Downer et al.
arXiv · 2026-08-24
This paper proposes a framework for evaluating AI risks at the system level rather than at the model level, bridging the gap between observed AI behaviour and real-world harm in complex sociotechnical systems such as Critical National Infrastructure. The authors apply their framework to the UK's Real Time Gross Settlement (RTGS) system, using Systems Theoretic Process Analysis (STPA) to derive AI-driven loss scenarios and focusing on adversarial manipulation of LLM-based trading as a case study. Component-level experiments show that simple adversarial inputs cause measurable behavioural shifts when AI recommendations are followed, and mapping these shifts onto a financial contagion model reveals increased bank failures and a lower threshold for cascading disruption, especially under widespread or monopolistic AI adoption. The work argues for evidence-based and anticipatory governance of AI in high-stakes systems, providing a traceable pathway from model behaviour to systemic outcomes.
- AI policy
- Quality assurance
Research
Names Can Hurt: Spotting Slopsquatting Risks Caused by Package Name Hallucinations in Local Coding LLMs
Akash Raj, Sargam Sahu
arXiv · 2026-08-24
This paper addresses 'slopsquatting,' a supply chain attack where adversaries pre-register hallucinated Python package names that AI coding assistants fabricate. The authors build a two-layer detector—combining a deterministic PyPI existence check with a Random Forest classifier trained on package name and metadata features—embedded in a LangGraph retry pipeline that escalates to a stronger fallback model on failure. Across 300 curated prompts, the system produces hallucination-free code 76% of the time, with intra-model retries and cross-model fallback recovering substantial additional cases. A user study of 24 participants reported mean satisfaction of 4.4 out of 5 and adoption intent from 21 of 24, suggesting practical viability for protecting software supply chains from LLM-induced package hallucinations.
- Quality assurance
- Enterprise
Research
Granite.Trust Policy Tools: Shareable, Actionable Policies for Generative AI Applications
Nathalie Baracaldo, Nicolas Mello, Kush R. Varshney et al.
arXiv (Cornell University) · 2026-08-24
This paper introduces Granite.Trust Policy Tools, a framework for specifying and enforcing safety policies in generative AI applications. The core contribution is an Actionable Policy schema—a YAML-based format that defines what model responses can and cannot contain—paired with a synthetic data generation pipeline that produces policy-aligned training data for model alignment and testing. Together, these tools allow organizations to write policies once and apply them across the full GenAI application lifecycle, from model alignment through runtime monitoring. The work addresses a gap where traditional access control policy formats fail to capture content-based constraints needed for generative AI contexts.
- AI policy
- Enterprise
Research
Beyond the Mandate: A Systematic Security Analysis of the Agent Payments Protocol (AP2)
Avital Aviv, Parth A. Gandh, Ron Bitton et al.
arXiv · 2026-08-24
This paper presents a systematic security analysis of Google's Agent Payments Protocol v0.2 (AP2), which allows LLM-driven shopping agents to authorize and execute payments on behalf of users. The authors find that while AP2's signed Checkout and Payment Mandates protect transaction data after signing, the pre-authorization phase—including Agent-to-Agent (A2A) messages and Model Context Protocol (MCP) tool calls—remains unprotected and exploitable. Using the MAESTRO threat-modeling framework and the Artificial Intelligence Vulnerability Scoring System (AIVSS), they catalog 48 threats across five attack families, identify eight reaching the High-risk band, and validate these with proof-of-concept demonstrations across five deployment architectures. The core finding is that valid mandate signatures alone cannot guarantee that an agent-mediated transaction reflects the user's actual intent when pre-authorization context has been manipulated—a critical concern for enterprise and policy stakeholders deploying agentic payment systems.
- Enterprise
- AI policy
Research
A Formal Methodological Framework for Auditing Robustness and Fidelity in Explainable AI: From Application to Trust Certification
Rosa Elysabeth Ralinirina, Jean Christian Ralaivao, Niaiko Michaël Ralaivao et al.
arXiv (Cornell University) · 2026-08-24
This paper develops a formal auditing protocol for post-hoc explainability methods (SHAP and LIME) that measures two key properties: robustness (stability of explanations under small input perturbations) and fidelity (whether highlighted features actually drive model predictions), combining them into a single Trust Score. The protocol is validated on a real-world malnutrition dataset from Madagascar (83 features, 253 records, 4 classes) across three classifiers and two explainers. Critically, the study finds that high predictive accuracy (AUC above 0.99) does not guarantee trustworthy explanations — models can produce degenerate or uninformative explanations, and fidelity scores lose discriminative power under overfitting. The authors argue that auditing XAI outputs is not optional but necessary, especially when explanations inform high-stakes decisions in sensitive domains.
- Certifications
- Quality assurance
Research
LUCAID: Agentic Multimodal AI for Lung Cancer Precision Pathology
Marie-Lisa Eich, Kai Standvoss, Timo Milbich et al.
arXiv · 2026-08-24
LUCAID is an agentic, multimodal AI system designed for precision lung cancer pathology that integrates nine diagnostic modules covering the full routine workflow—including quality control, tumor detection, histological subtyping, biomarker scoring (PD-L1, MET, TROP-2), and automated report generation. Evaluated against large-scale expert annotations, the modules achieved F1 scores of 0.82–0.95, and in prospective clinical validation LUCAID reached 93.0% concordance with an expert-panel reference standard, outperforming five experienced thoracic pathologists who achieved 68.3–81.1% concordance. The results suggest that agentic AI can reduce interobserver variability and extend expert-level diagnostic performance across the complex, multi-step assessments required for precision oncology treatment decisions.
- Quality assurance
- Enterprise
Research
When Youth Enter The Chat: An Epistemic Shift in the Validation of LLM-Based Measures of Student Talk
Liliana Santos-Deonizio, James Malamut, Ramón Martínez et al.
arXiv · 2026-08-24
This paper investigates the use of large language models (LLMs) to measure aspects of student classroom discourse—such as talk moves, collaboration, and equity of voice—and argues that current validation practices (e.g., comparison against adult expert annotations and F1 scores) are insufficient, especially for racially and linguistically marginalized youth. Through a case study of multilingual 8th-grade math students, the researchers employed ethnographic methods including participant observations, interviews, focus groups, and member checks to re-contextualize student conversations. Findings show systematic misalignments between students' own interpretations of their math talk and the LLM-based measures, with students actively contesting both the classifications and the coding schemes used. The paper concludes that meaningful and equitable AI-based measures of student talk require sharing epistemic authority with youth themselves, not just adult researchers or automated systems.
- Quality assurance
- AI policy
Research
TrustShiftProbe: Characterizing, Benchmarking, and Defending Staged Trust Attacks on MCP Servers
Mehrdad Rostamzadeh, Sidhant Narula, Mohammad Ghasemigol et al.
arXiv · 2026-08-24
TrustShiftProbe identifies and evaluates a novel class of server-side attacks on the Model Context Protocol (MCP), the standard interface linking Large Language Model agents to external tool backends. The attack, called TrustShift, works by having a compromised MCP server behave legitimately during an initial conditioning phase before switching to an adversarial payload once a trust threshold is reached, making it invisible to static pre-deployment analysis. The paper introduces a benchmark with nine attack variants across three execution mechanisms and adversarial objectives, finding that TrustShift achieves a 69.5% mean attack success rate across frontier models; the authors' proposed runtime defense, SHIELD, reduces this to 42.7% by auditing server payloads against behavioral baselines learned during clean trust windows. This work matters because it exposes a fundamental trust assumption in agentic AI deployments where the server endpoint itself is the adversary, rather than user prompts or network transport.
- Quality assurance
- Enterprise
Research
EG-ARSA: An Expert-Grounded Open Model for Visual Road Safety Auditing in Low-Resource Settings
Md Thamed Bin Zaman Chowdhury, Moazzem Hossain
arXiv · 2026-08-24
EG-ARSA introduces Expert-Grounded Distillation (EGD), an AI framework that compresses road safety expertise from a large vision-language model into a compact 8-billion-parameter student model for scalable visual road safety auditing. The teacher model is first calibrated against authoritative field audits until it reaches substantial agreement with expert risk assessments (Cohen's kappa = 0.74), then used to generate structured training data for the student via Low-Rank Adaptation. The paper also releases BD-ARSA, the first open expert-grounded Bangladeshi road safety audit dataset with 21,947 image-audit records, and shows that the compact student model outperforms both its larger 31-billion-parameter teacher and Gemini-2.5-Flash in blind expert evaluation. This work demonstrates a cost-effective path to proactive road safety auditing in low- and middle-income countries where qualified auditors and crash records are scarce.
- Workforce
- AI policy
Research
Automata from Agent Traces: Failure and Next-Step Prediction
Seonglae Cho, Franklin Cardenoso Fernandez, Umar Mohammed et al.
arXiv · 2026-08-24
This paper introduces a method that compresses large collections of LLM agent execution traces into compact finite-state machines (FSMs) to make agent behavior interpretable and monitorable. Across twelve public datasets, the resulting FSMs are small (7–43 states), reproduce held-out data with fitness ≥0.997, and build in milliseconds, providing a shared structural substrate for both next-step and failure prediction. For next-step prediction, FSM-state context outperforms Agent Workflow Memory on every ground-truth-matched dataset; for failure prediction, per-state features achieve held-out AUROC up to 0.94, and an online monitor can rank failing runs above passing ones from partial traces to trigger early stopping. Because behavioral topology appears shaped more by the deployment harness than by the underlying LLM, the approach offers a model-agnostic primitive for safety auditing and runtime monitoring of deployed AI agents.
- Quality assurance
- AI policy
Research
How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles
Shang Wu, Catarina G Belem, Shuyuan Fu et al.
arXiv · 2026-08-24
This controlled experiment examined how on-demand AI assistance during logic-puzzle solving affects participants' skill development over time. The study found that lower-cost AI access led to more frequent AI use, and participants who relied on AI assistance subsequently performed worse once the AI was removed, with their unassisted ability overestimated by their AI-assisted performance. Using a Bayesian latent ability model, the researchers show that greater independent problem-solving effort during the AI-access phase was associated with larger gains in latent ability, suggesting that AI assistance can substitute for—and thereby weaken—independent reasoning and skill acquisition. These findings raise important concerns about how AI tools are deployed in learning and work contexts where long-term skill development matters.
- Workforce
Research
Confidently Wrong, Silently So: Auditing Undetectable Failures of a Deployed On-Device Language Model
Shashwat Pandey, Satwik Pandey, Suresh Raghu
arXiv · 2026-08-24
This paper audits an on-device language model deployed on hundreds of millions of devices and finds it fails in ways that are both frequent and invisible to users. The model confabulates on 69% of false-premise questions while refusing 18% of benign inputs—a 'task-asymmetric miscalibration'—and its self-reported confidence is essentially non-discriminative (AUROC 0.47), meaning correct and incorrect outputs look identical on the surface (classifier AUROC only 0.55). Because no cheap single-generation signal reliably flags failures, users and resource-constrained developers have no practical way to detect when the model is wrong at inference time. The authors show that a black-box consistency wrapper can largely recover reliability (confabulation rate drops from 75% to 3%), and they release a model-agnostic audit protocol and evaluation infrastructure for independent oversight of deployed models.
- Quality assurance
- AI policy