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.
5221 items
Research
Causal Evidentiary Governance for High-Risk Machine Learning Systems
Samah Kareem, Barış Çeliktaş
arXiv · 2026-09-01
This paper introduces Causal Evidentiary Governance (CEG), a framework for auditing high-risk machine learning systems—such as those used in credit, hiring, and resource distribution—by replacing standard observational fairness metrics with causal analysis. CEG requires regulated institutions to commit to a versioned directed acyclic graph (DAG) that separates allowable from disallowed causal pathways, and introduces a Causal Harm Rate metric to measure prediction variation attributable to disallowed pathways. Each decision is cryptographically bound to the published DAG via a signed Decision-Evidence Packet, with Merkle tree inclusion proofs enabling efficient audit verification. Empirical validation using synthetic credit applicants and the German Credit dataset shows that the Causal Harm Rate detects injected causal harms more clearly than demographic parity or equalized odds, and that associational metrics can substantially understate pathway-specific harm—directly relevant to compliance with regulations such as the EU AI Act and GDPR.
- AI policy
- Certifications
Research
Data-Driven Persona-Conditioned Agents for A/B Test Simulation
Ziyad Benomar, Weronika Łajewska, Leonardo Perelli et al.
arXiv · 2026-09-01
This paper proposes a simulation framework that uses large language model (LLM) agents conditioned on data-driven personas—built from anonymized real user behavioral signals—to predict the outcomes of A/B tests before they are run with live traffic. The authors systematically evaluate question design formats, persona data sources, population diversity trade-offs, and subsampling strategies across a benchmark of 40 A/B tests, achieving 0.75–0.90 directional accuracy depending on the metric type. The results suggest that data-driven persona-conditioned agents can serve as a fast, low-cost pre-screening tool, reducing the engineering effort and time required to run full live experiments.
- Enterprise
- Quality assurance
Research
Disclosure-Gated User Simulation for Companion-Agent Evaluation
Yao Liu, Yu He
arXiv · 2026-09-01
This paper addresses a well-known flaw in LLM-based user simulation for evaluating AI companion agents: simulated users are overly cooperative, allowing systems to score well simply by asking many questions rather than by genuinely earning user trust. The authors introduce a 'disclosure gate' mechanism that conditions how much information a simulated user reveals based on the companion agent's behavior, structured as a five-level ladder collapsed into three observable depth layers. They train a user simulator against this specification and demonstrate that it is the only candidate meeting both their acceptance criteria—order-preserving rankings and scale-stable absolute scores—achieving a leaderboard correlation of 0.993 with the original benchmark simulator. The work also reveals that naively prompting a frontier model as a simulator inflates all scores without visibly disturbing rankings, a subtle bias that rank-only checks would miss.
- Quality assurance
Research
Staged Linguistic Seeding: Grounded Query Expansion for Verified-Unit QA in AI Contact Centers
Hyeonseop Yoon, Jeong-Eun Park
arXiv · 2026-09-01
This paper presents Staged Linguistic Seeding (SLS), a system for quality-assurance in AI contact centers that answers customer queries only from a closed set of verified QA units — returning answers verbatim or routing to clarification, abstention, or human handoff. Rather than generating answers freely, the system enriches a retrieval index offline by having humans author slot recipes that a language model (gpt-4.1-mini) expands into variants, filtered by a human gate. On held-out queries from two industrial domains, SLS improves hybrid retrieval accuracy (R@1) to 0.881 and 0.930 — gains of +0.27 and +0.34 — and outperforms the doc2query baseline by +0.20 and +0.32 at the same generation budget. Critically, the verified-unit approach eliminates unsupported content from responses (reducing it from 7–13% to approximately 0%), directly addressing a key quality-assurance concern in deployed contact center AI.
- Quality assurance
- Enterprise
Research
Probabilistic Model Checking of Autoregressive Neural Sequence Models
Helge Spieker, Dennis Gross, Arnaud Gotlieb
arXiv · 2026-09-01
This paper introduces a probabilistic model checking pipeline for evaluating autoregressive neural sequence models beyond standard test-set accuracy. It extracts a discrete-time Markov chain (DTMC) from token-by-token generation, verifies formal PCTL specifications using the PRISM model checker, and produces certified probability intervals on constraint satisfaction via a soundness theorem guaranteeing conservative under-approximation. Applied to a GPT-2 process-planning model and a SMILES molecular generator, the pipeline reveals probability mass on constraint-violating outputs that greedy decoding hides but sampling can reach, and quantifies gaps between structural completeness and chemical validity that accuracy metrics cannot report. This matters for quality assurance of deployed AI systems, offering formal guarantees and counterexample traces where pass/fail accuracy scores are insufficient.
- Quality assurance
- Certifications
Research
Ctrl-F-Resist. Practices, Challenges, and Technical Needs of Civil Society Organizations Monitoring the Far-Right Online
Elisabeth Steffen, Helena Mihaljević
arXiv · 2026-09-01
This paper examines how civil society organizations (CSOs) in Germany monitor far-right activity online, finding that monitoring remains largely manual due to a lack of tailored tools, with enhanced search capabilities identified as the most pressing technical need. Based on qualitative interviews with 15 practitioners from 12 CSOs, the study reveals significant barriers to adopting AI tools—including legal uncertainty, limited platform access, and chronic underfunding—and introduces the concept of the 'manual labor trap' to explain why these organizations remain locked into labor-intensive, low-capacity arrangements. While participants were open to AI-assisted features like media processing and content discovery, many were skeptical of automated classification due to concerns about trust, legal usability, and professional credibility. The authors present a conceptual monitoring workflow and an open-source Telegram monitoring prototype, alongside design, policy, and research recommendations to better support CSOs in this role.
- Workforce
- AI policy
Research
Effective Interventions Against AI-Enhanced Scams
Kyle Fredrickson
arXiv · 2026-09-01
This paper examines how AI is changing the economics of scamming operations, noting that global scam losses reached an estimated $442 billion in 2025 and that U.S. reported losses grew by nearly 400% between 2020 and 2025. The author builds a model of scam profitability to identify which interventions remain effective in an AI-driven scamming environment. The analysis finds that three factors—reporting rate, centralization of reporting, and report accuracy—multiply together in their effect on expected victims per scam channel, simultaneously reducing revenue and increasing costs for scammers. The paper concludes that even modest improvements in reporting rates targeting high-value scam infrastructure could substantially undermine the profitability of AI-enhanced scam operations.
- AI policy
Research
Self-Reports Are Not Verification: Environment-Grounded Auditing of LLM Operators in Evolutionary Search
Enrong Pan, Ryan Zhou, Ting Hu
arXiv · 2026-09-01
This paper introduces an environment-grounded auditing method for large language model (LLM) agents by having a model run an evolutionary search on the word-guessing game Contexto, where every guess receives an exact rank as ground-truth feedback. Across 200 runs and 12,249 self-reports, the authors find that all three tested assumptions about self-reporting fail: models overstate success by factors of 4.8 to 9.3, stated confidence is poorly calibrated, inherited rationales provide negligible benefit, and fitness-based selection does not improve report accuracy. The findings demonstrate that LLM self-reports are unreliable as monitoring signals and should be treated as unverified claims rather than evidence of trustworthy behavior. This has direct implications for how AI systems are audited and monitored in deployment settings.
- Quality assurance
- AI policy
Research
Are We There Yet? Assessing Computer-Use Agents for Blind Users' Accessible Interaction with Desktop Applications
Satwik Ram Kodandaram, Monalika Padma Reddy, Xiaojun Bi et al.
arXiv · 2026-09-01
This paper investigates whether computer-use agents (CUAs) can meaningfully support blind screen-reader users in real-world desktop workflows. The researchers conducted a three-week diary study with 8 blind participants using OLLA, a screen-reader-accessible CUA prototype, collecting 1,258 commands across 12 applications. Testing GPT-5 and four other models on the same command traces, the best-performing model (GPT-5) achieved only a 52.5% success rate, with failures attributed to grounding, planning, constraint-tracking, and termination errors. The findings suggest that current CUAs fall well short of reliable accessibility support for blind users, pointing to specific technical and user-experience gaps that must be addressed.
- Workforce
- Enterprise
Research
RecalibrateGPT: AI Fatigue Resilient Conversational Interfaces
Nikhil Wani
arXiv · 2026-09-01
RecalibrateGPT is a conversational interface system designed to reduce 'AI fatigue' — the cognitive burden caused by repetitive type-read-retype loops in LLM interactions. The system introduces five cross-turn operators (Anchor, Replay, Delta, Scope, and Steer), each targeting a distinct fatigue type identified in a formative study: retyping, scanning, decision paralysis, and context drift. A follow-up quantitative pilot study with 12 advanced LLM users found that RecalibrateGPT reduced perceived cognitive workload by half (NASA-TLX = 2.7) while achieving high usability (SUS = 86.5). The work argues that AI fatigue is an interaction-flow cost that interface design can address, not merely a model-quality problem.
- Workforce
- Enterprise
Research
Profiles of Primary School Teachers’ Engagement with Generative AI: The Role of Risk, Trust, and Professional Development Needs
Elisabeth Fuchs, Barbara Lenzgeiger
Education Sciences · 2026-09-01
This survey of 295 primary school teachers identifies three distinct profiles of engagement with generative AI—Skeptical Rejecters, Interested but Uncertain, and Trusting Learners—based on perceived risk, trust, and professional development needs. The study finds a general state of cautious openness, with moderate-to-high perceived risks and relatively low trust across the sample. A notable intention-usage gap emerged, where intended AI use consistently exceeded actual implementation, and the highest professional development needs were reported by the Interested but Uncertain group rather than the most skeptical teachers. The findings underscore the heterogeneity in teacher readiness and call for professional development that builds both AI competencies and informed professional judgment.
- Workforce
- AI policy
Research
Artificial Intelligence in Family Nursing: An Evidence‐Informed Policy Framework for Safe and Equitable Integration
Junko Honda, Sakiko Itoh, Kazumi Kubota
International Nursing Review · 2026-09-01
This paper proposes a five-domain policy framework for integrating AI into family nursing safely and equitably, covering ethical data governance, workforce development, interdisciplinary co-design, adaptive regulation, and sustainable funding. Drawing on a structured narrative synthesis of academic and policy literature, it identifies AI opportunities—such as earlier risk identification, tailored education, remote monitoring, and family-level outcome evaluation—alongside risks including bias, privacy concerns, digital exclusion, and over-reliance on automated recommendations. Japan is used as an illustrative context given its aging population and community-based care model. The framework calls on policymakers, regulators, and educators to clarify accountability, embed AI literacy in nursing education, and ensure reimbursement for nurse-led technology-enabled care.
- AI policy
- Workforce
Research
Explainable cross-layer machine learning for enterprise WLAN auditing
Victor Stoynov
International Journal on Information Technologies and Security · 2026-09-01
This paper introduces an explainable cross-layer machine learning system called CrossAudit that audits enterprise and campus wireless networks by combining radio, MAC-layer, security, and mobility signals to detect anomalies and classify their root causes. Evaluated on a Python simulator across three enterprise sites and 69,120 audit windows, the drift-aware system achieves an anomaly F1-score of 0.770 and an eight-state macro-F1 of 0.855, outperforming a network-only baseline while reducing false-positive alarms. The system also generates evidence-supported recommendations for operator review, making its findings interpretable for network administrators. These results are directly relevant to enterprise network operations, where automated and explainable auditing can improve reliability and reduce manual oversight burden.
- Enterprise
- Quality assurance
Research
When Does Generative AI Adoption Pay Off? Ethical Governance as a Dynamic Capability in SMEs
Tahereh Hasani
Information & Management · 2026-09-01
This study examines when generative AI (GenAI) adoption translates into better innovation performance for small and medium-sized enterprises (SMEs), using survey data from 218 Canadian managers analyzed via structural equation modeling. It finds that ethical AI governance—specifically transparency/explainability, privacy/data protection, and accountability—significantly amplifies the innovation payoff from GenAI adoption, while overly rigid fairness controls may slow early experimentation in resource-constrained SMEs. Green-innovation orientation is the strongest positive driver of adoption, whereas carbon-footprint pressure and economic disruption reduce it. The paper reframes ethical AI governance as a dynamic capability rather than compliance overhead, with downstream benefits including greater adaptability, operational efficiency, and revenue growth.
- Enterprise
- AI policy
Research
The Kazakhstan Context: AI Tools for Employment of Persons with Disabilities
Ramiro Martínez, Christopher Mutseekwa
Economic Consultant · 2026-09-01
This study analyzes AI-based tools for improving employment outcomes among persons with disabilities (PWDs) in Kazakhstan, where roughly 376,000 working-age PWDs exist but only about 30% are employed. The authors find that the gap between unemployed PWDs (~265,000) and available quota or voluntary vacancies (~68,700) is severe, and identify four AI tool categories—assistive technologies, intelligent job-matching, workplace adaptation, and career guidance—as practical interventions. The paper also flags risks such as algorithmic discrimination and gender bias, and proposes a policy roadmap including a dedicated PWD module in Kazakhstan's Digital Employment Service and localized AI applications. The authors conclude that deploying AI for inclusive employment is a strategic economic necessity rather than an optional technological upgrade.
- Workforce
- AI policy
Research
When a Human‐in‐the‐Loop Amplifies the Risk of Misalignment: Considerations for Research Ethics Oversight
Erin Taylor
Ethics & Human Research · 2026-09-01
This paper argues that human-in-the-loop (HITL) approaches, commonly proposed to reduce misalignment risks when using large language models in research ethics oversight, can paradoxically amplify those risks. Using protocol triage as a case study, the author shows how reliance on imperfect proxies for ethical principles creates a proxy-target gap, and that human reviewers' involvement introduces vulnerabilities such as hallucination, over-reliance, reward hacking, and sycophancy. The paper identifies mitigation techniques—including retrieval grounding, calibration drills, structured prompting, and adversarial debate—and concludes that effective oversight requires ethics committee reviewers with specialized competencies to recognize model vulnerabilities and strategically deploy mitigations.
- Quality assurance
- Certifications
Research
Advances in online cost-effectiveness analysis models and AI-assisted decision support tools: A mixed-methods study with implications for China
Jiayi Zhang, Zehui Liu, Furong Tian et al.
Pharmacoeconomics and Policy · 2026-09-01
This mixed-methods study reviewed 29 publications and 29 online cost-effectiveness analysis (CEA) tools to assess how pharmacoeconomic platforms and AI-assisted decision support could improve drug reimbursement negotiations under China's National Reimbursement Drug List (NRDL). The findings show that existing platforms have advanced accessibility, interactivity, and workflow integration for economic evaluations, while AI-assisted tools can improve evidence extraction and model construction efficiency—though their outputs still require manual review. Key challenges include insufficient local data adaptation, inconsistent parameter traceability, limited real-world reimbursement evidence, and poorly defined regulatory usability. The authors conclude that platforms must incorporate locally relevant data, standardized submission-ready outputs, and accountability mechanisms for AI-assisted outputs to be useful in China's HTA and reimbursement policy processes.
- AI policy
- Quality assurance
Research
A comprehensive narrative review of artificial intelligence applications in drug discovery and pharmaceutical practice
Asef Raj, Mehedi Hasan, Sumaiya Binta Shahadat
Journal of Pharmaceutical Analysis · 2026-09-01
This narrative review examines how AI techniques such as machine learning and neural networks are being applied across drug discovery and pharmacy practice. In research settings, AI accelerates identification of drug candidates by predicting physicochemical properties, optimizing pharmacokinetic and pharmacodynamic profiles, and supporting early toxicity assessment. In clinical settings, AI automates prescription verification, assists dose selection, detects adverse drug reactions in near real time, and reduces medication errors. The authors identify key barriers to broader adoption including biased training data, opaque 'black box' models, fragmented regulatory frameworks, and shortages of trained professionals, concluding that long-term impact depends on resolving regulatory and ethical issues through collaboration among researchers, clinicians, and policy makers.
- Enterprise
- AI policy
- Quality assurance
Research
A Taxonomy of Digital Technology Configurations in Manufacturing Enterprises: An Unsupervised Profiling and Longitudinal Trajectory Analysis
Guilei Tan, Rozaini Rosli
Economics & Business Management · 2026-09-01
This study develops an empirical taxonomy of digital technology adoption among 44,103 firm-year observations from Chinese A-share listed manufacturing enterprises (2000–2025) using K-Means clustering on text-mined disclosures across AI, Big Data Analytics, Cloud Computing, Blockchain, and Advanced Digital Technology pillars. It identifies four archetypes—Traditional Laggards (64.40%), Application-Centric Followers (20.71%), Cloud and Algorithmic Integrators (11.56%), and Full-Stack Frontier Pioneers (3.33%)—that differ significantly in enterprise scale, capital leverage, board independence, and state ownership. Longitudinal analysis reveals a decisive shift after 2015, with traditional laggards contracting and cloud-integrated configurations expanding rapidly. The taxonomy offers practitioners and policymakers a benchmarking tool to calibrate digital investment portfolios.
- Enterprise
- AI policy
Research
AI-Driven Intelligence Layer 4 SME-to-B2B Marketplace Optimization: The AISLE Framework
R. Geetha, P. Selvi, R. Manju et al.
International Journal of Computer Information Systems and Industrial Management Applications · 2026-09-01
This paper presents the AISLE framework, an AI-driven optimization layer designed to address sourcing, demand forecasting, and pricing inefficiencies for small and medium enterprises (SMEs) in B2B digital marketplaces. Evaluated across 47 SMEs in India and Southeast Asia over six months, the framework demonstrated ML-driven supplier matching reducing search time by up to 68%, order forecasting accuracy improving by 28.7%, and gross margin gains of 12–18%. The system is built around lightweight, federated, and API-first architectures suited to resource-constrained environments, extending AI capabilities to micro-SMEs previously excluded from mainstream digital commerce platforms. The authors argue the findings have practical implications for platform architects, policymakers, and practitioners in emerging economies.
- Enterprise
- AI policy
Research
Research on the Interaction between Creator and Intelligence-Generated Content (AIGC) Painting Tools from the Perspective of Technological Mediation Theory
Qi Zhang, Yi Ding, Xiaomeng Hu
Computers in Human Behavior Reports · 2026-09-01
Drawing on semi-structured interviews with 24 visual-creative practitioners and grounded theory methodology, this study examines how AI-generated content (AIGC) painting tools mediate creative workflows. The findings show these tools are used primarily in early and intermediate creative stages—supporting ideation, reference generation, and iterative exploration—but redistribute rather than eliminate labor by introducing new tasks such as prompt crafting, output comparison, and revision. Creators retain agency through aesthetic judgment and direction-setting, while their evaluations of the tools vary by professional context, proficiency, and concerns about copyright, platform dependence, and shifting labor expectations.
- Workforce
Research
Data centres and the new strategic infrastructure governance: From cloud ESG to compute sovereignty
Vlado Vivoda, Ron Janjua, Danilo Borja
The Extractive Industries and Society · 2026-09-01
This paper argues that data centres have become strategic infrastructure whose governance frameworks have not kept pace with their systemic importance, especially as AI drives surging demand for electricity, water, land, and critical materials. The authors develop the CLOUDS framework—covering Coherence, Lifecycle materiality, Operational locality, Utility burden, Digital sovereignty, and Social licence—to diagnose how data centres shape grids, water systems, supply chains, and security architectures. Drawing on cases from Chile, Australia, Mexico, Ireland, and Singapore, the paper identifies two core findings: regulatory incoherence can emerge even without capture when no single institution can govern cross-sector trade-offs, and physical hosting of data centres does not guarantee a nation's sovereign access to compute or higher-value AI workloads. The policy implication is that governments should pursue strategic infrastructure bargaining to secure both system capacity and control over how that capacity is converted into compute.
- AI policy
Research
Megawatts, but few permanent jobs? An agent-assisted evidence review of the data-centre construction cliff
M. Reza Hosseini, Abdolmajid Erfani, Ania Khodabakhshian et al.
Smart and Sustainable Built Environment · 2026-09-01
This paper synthesizes quantitative evidence on how many jobs data-centre construction and operation actually create per megawatt of capacity, finding a sharp contrast between roughly 0.7–2.0 peak construction workers/MW and only 0.15–0.35 permanent operational FTEs/MW. The study warns that gross input-output multiplier totals describe a fundamentally different object from realised local permanent employment, and that job claims vary by an order of magnitude partly because they conflate phase, denominator, boundary, and evidence type. The authors recommend that policymakers and practitioners require employment claims to explicitly state these parameters before approving incentive packages, and propose a six-layer employment boundary framework as a measurement discipline for future evidence collection.
- Workforce
- AI policy
Research
Governing artificial intelligence in cancer care: Perspectives and recommendations from patient organizations and the MULTIR consortium
Isabelle Budin-Ljøsne, Alessandro Boni, Lise Boussemart et al.
Journal of Cancer Policy · 2026-09-01
This study gathered perspectives from 18 patient organization representatives—spanning melanoma, bladder, and lung cancer communities—to co-develop recommendations for governing AI in oncology. Through semi-structured interviews and validation webinars, participants identified six core recommendations targeting developers, policymakers, regulators, and researchers, covering transparent communication, healthcare professional training, demonstrable clinical benefits, secure data infrastructure, ethical/regulatory frameworks, and sustained patient involvement. The findings are intended to guide equitable and patient-centered integration of AI in cancer care, with particular relevance to European digital health and data governance policy.
- AI policy
Research
AI‐Related Hiring Expansion and Within‐City Occupational Demand Reallocation in China: Evidence From Listed‐Firm Job Postings
Zhaoming Sun, Yawen Zhai, X. B. Tian
Growth and Change · 2026-09-01
Using job-posting data from Chinese listed firms (2016–2024), this study finds that cities with higher AI-related hiring demand show a selective concentration of vacancies in occupations closely linked to AI, rather than a broad distribution across all roles. AI-related hiring expansion is also associated with greater demand for digitally complementary occupations, and these effects are amplified in cities with stronger higher-education capacity and more developed innovation ecosystems. The findings reveal a growing within-city divergence in occupational demand driven by AI adoption, with implications for how cities design talent-development systems and manage labor market inequality.
- Workforce