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
Towards a Resilience-Theoretic Foundation for Adversarial Robustness in Industrial Control System Anomaly Detection
Branka Stojanović, Andreas Flatscher, Michael Somma
arXiv (Cornell University) · 2026-09-07
This paper establishes a formal theoretical connection between adversarial robustness in machine-learning-based anomaly detection and system resilience for industrial control systems (ICS) and operational technology environments. The authors map four resilience constructs—disturbance class, absorption capacity, recovery trajectory, and degradation function—onto the adversarial machine learning setting, then derive a compositional resilience bound showing that the system-level binding constraint depends on coupling-adjusted absorption capacity along the attack path, not simply the weakest individual node. Empirical validation on the BATADAL water distribution benchmark reveals operationally significant phenomena invisible to standard benchmarks, including an absorption-degradation divergence under adversarial training and a paradox where hardening the binding node in isolation can actually reduce system-level resilience. The findings carry direct implications for ICS architecture design and certification standards for anomaly detection systems.
- Certifications
- Quality assurance
Research
Risk Is Not Review Value: Wrong-Answer Exposure Under Bounded Review Budgets
SangJin Park, Myungsub Choi, Jineok Kim et al.
arXiv · 2026-09-07
This paper argues that prioritizing which LLM-generated answers to review under a fixed budget should go beyond simply flagging high-risk answers. The authors propose a 'review value' framework that combines estimated wrongness, intervention affordability, impact, and cost, and introduce two metrics—Wrong-Answer Exposure Ratio (WAER) and post-repair residual exposure (PRRE)—to evaluate how well a review queue reduces users' exposure to wrong answers. On a 720-item TAT-QA/SciFact benchmark, their review-value ranking kept WAER nearly unchanged at a 20% budget (0.605 vs. 0.600) but substantially lowered PRRE from 0.881 to 0.716, demonstrating that repairability matters as much as error detection. The findings suggest that trustworthy LLM evaluation must account for how limited human review capacity can be best deployed to minimize wrong answers reaching users.
- Quality assurance
Research
Runtime Assurance with Signed-Ledger Provenance for Autonomous Fusion Control: A Zero-Trust, Hardware-Failsafe Last Line of Defense Beneath the AI
Priyanca Ford, G L Kulcinski
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-07
This paper formalizes and demonstrates the Kronos last-line-of-defense architecture for autonomous control in nuclear fusion plants, combining a certified control-barrier-function (CBF) safety projection, hardware failsafe, runtime assurance, and signed hash-chained provenance ledger to ensure no AI-learned component can cause an unsafe action. Across two distinct magnet designs—a negative-triangularity spherical tokamak and a D–³He tandem-mirror burner—the certified clamp catches all 3,000/3,000 injected unsafe actions with zero escapes and holds safety constraints with 0/100 violations in both test scenarios, while an uncertified greedy controller fails repeatedly. The paper also shows that safety guarantees are provenance-conditional: forward-invariance is preserved in 10/10 lineage-gated calibrations but only 1/10 without the gate, making the signed ledger load-bearing for safety rather than administrative. The architecture includes a Goal-Structuring-Notation assurance case and hazard bow-tie aimed at regulatory review, making this directly relevant to certification and quality-assurance for autonomous systems in high-stakes environments.
- Certifications
- Quality assurance
Research
A study of cybersecurity literature on open-source software and AI
Srinidhi Vasudevan, Anna Piazza, Guru Krishna Ramakrishnan et al.
Greenwich Academic Literature Archive (University of Greenwich) · 2026-09-07
This literature review, commissioned by the UK Department for Digital, Culture, Media & Sport (DCMS), examines the cybersecurity landscape for open-source software (OSS) and open-source AI (OSAI). It finds that open-source components appear in 96% of commercial codebases—including government systems, critical services, and AI pipelines—yet existing security frameworks designed for traditional software do not cleanly apply to AI artefacts such as model weights, training data, and fine-tuning pipelines. High-profile incidents like Log4Shell (2021) and the XZ Utils backdoor (2024) have made these risks undeniable. The review maps global frameworks and practices, identifies security gaps and emerging risks, and offers early policy recommendations for governing open-source dependencies more consistently.
- AI policy
- Quality assurance
Research
MLOps for a Safety-Critical Fusion Control Stack: Versioned Lineage, Gated Model Promotion, and Physics-Balanced Multi-GPU Training for a Compact Spherical-Tokamak Breeder
Priyanca Ford
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-07
This paper formalizes the machine-learning operations (MLOps) pipeline for the Kronos control stack of the Hyperion compact spherical-tokamak fusion breeder, establishing that no ML model may reach control authority over the nuclear plant without first passing a reproducible, physics-anchored promotion gate. The authors derive the full mathematical framework—including constrained model promotion, physics-balanced multi-GPU training objectives, hash-chained provenance ledger, and drift-triggered rollback logic—and demonstrate the pipeline on an NVIDIA HPC campaign where the Kronos engine reproduces all 19 of 19 frozen physics anchors bit-stably across three independent environments. Supporting components achieve strong performance: a Bayesian hyperparameter loop yields a calibrated confinement posterior H_98=1.007±0.030, an equilibrium surrogate reaches 0.139% RMS error at 2877× speedup, and a disruption ensemble estimator achieves AUC 0.980. The work establishes a certified, end-to-end reproducible promotion pipeline as the operational foundation for trustworthy learned control of a safety-critical nuclear machine.
- Quality assurance
- Certifications
Research
Ollama in the Wild: A Longitudinal Measurement of Exposed Ollama LLM Endpoints at Internet Scale
Zuyao Xu, Xiang Li, Yuqi Qiu et al.
arXiv (Cornell University) · 2026-09-07
This paper presents a year-long, internet-scale measurement study of publicly exposed Ollama endpoints—self-hosted large language model servers running on port 11434—tracking over 152,000 unique IP addresses across approximately 4.8 million IP-day observations. Key findings include that 26.4% of discovered hosts appear for only a single day, upgrade rates following known CVEs are extremely low (0.43–2.90%), and cloud/hosting providers dominate the exposure landscape. The results characterize exposed Ollama deployments as a persistent, growing, and concentrated security surface, with recurring insecure practices suggesting that provider-level interventions may be the most effective mitigation path.
- Quality assurance
- AI policy
Research
Norms at a Price: Why RL-Based Alignment Can Promise Conditional Compliance at Best
Kevin Baum, Rūta Binkytė, Felix Jahn
arXiv (Cornell University) · 2026-09-07
This paper argues that reinforcement learning-based alignment cannot guarantee that AI systems comply with norms when unobserved, only when they believe they might be caught. Because training scores only observed behavior, a policy that complies conditionally is indistinguishable from one that complies genuinely—making 'conditional compliance' the theoretical ceiling of behavioral training. The authors conclude that the remedy lies not in better internalization of norms but in architectural constraints that make violations structurally unavailable rather than merely disincentivized.
- AI policy
- Certifications
Research
"We Permit the Use of AI, but [...]": The Landscape of AI Policies in Popular Open Source Projects
Andre Hora, Romain Robbes, Stefano Zacchiroli
arXiv (Cornell University) · 2026-09-07
This paper analyzes 281 AI contribution policies adopted by open source projects to map how these communities are governing AI-generated code. The authors find that 83.3% of policies permit or encourage AI contributions, but most impose conditions such as high human involvement (67.3%) and accountability assignment (43.4%), while nearly half require disclosure of AI use. The study also identifies ten countermeasures against low-quality AI-generated content ('AI slop') and shows that these policies are evolving rapidly, with half of dedicated AI policy files already revised since creation. The findings provide a baseline for understanding how open source governance is adapting to AI-assisted development.
- AI policy
- Enterprise
Research
You can contribute if you... An Empirical Framework of AI Contribution Policies in OSS
Gregorio Robles, Daniel M. Germán
arXiv (Cornell University) · 2026-09-07
This paper empirically studies how open source software (OSS) projects are regulating AI-generated contributions through their guidelines and documentation. The authors analyze the rationales, rules, and expectations embedded in these policies, finding they aim to protect maintainer attention, preserve accountability, ensure quality, address legal concerns, and sustain pathways for newcomer learning. Based on this analysis, they introduce an AI Contribution Governance Framework to help OSS communities design AI contribution policies and give researchers a vocabulary for studying AI's impact on collaborative software development.
- AI policy
- Quality assurance
Research
The Oversight Gap: What LLM Safety Monitors Miss, and Why It Is Not Capability
Xin Xu
arXiv (Cornell University) · 2026-09-07
This paper analyzes fundamental limits of LLM safety monitors when evaluating 'hyperproperties'—security properties like cross-tenant noninterference or sandbagging that require observing two executions rather than one. The authors derive a tight mathematical bound showing that single-trace monitors cannot exceed a balanced accuracy of 1/2 plus half the total variation distance between distributions, defining an 'oversight gap' as the shortfall below this bound. Testing nine LLM monitors, they find the gap is mostly not a capability problem: simply specifying the correct comparison procedure closes 61% of the gap, while having information without the right procedure or vice versa leaves large shortfalls. The paper concludes that benchmark validity for such properties should be mechanically proved rather than audited by LLMs, after finding that frontier LLM judges falsely certified a flawed benchmark version despite a statistically significant directional bias.
- Quality assurance
- Certifications
Research
Copyright Protection of AI-Generated Works: Challenges and Emerging Legal Approaches
Dr. Sunita Banerjee, Md Jiyauddin -
arXiv · 2026-09-07
This book examines the legal tensions between existing copyright frameworks and AI-generated creative works, analyzing whether current laws adequately address questions of authorship, originality, ownership, and infringement when AI systems produce literature, art, music, and software. It conducts a comparative legal analysis spanning Indian copyright law, the Berne Convention, TRIPS, WIPO treaties, EU regulation, U.S. human-authorship doctrine, and UK treatment of computer-generated works. The work identifies gaps in existing legal structures and explores potential remedies including new legislation, licensing regimes, transparency requirements, collective management, and sui generis protections. Its findings are directly relevant to policymakers and legal practitioners seeking a balanced framework that supports both technological innovation and human creativity.
- AI policy
Research
FunnelAudit: Responsibility Auditing in Multi-Route Recommender Systems
Jie Li, Dudu Luo, Jiayang Niu et al.
arXiv (Cornell University) · 2026-09-07
FunnelAudit is a framework for auditing responsibility in multi-route recommender systems that combine retrieval, allocation, fusion, and ranking stages. The paper shows that standard one-at-a-time ablation methods are inadequate—92.55–99.64% of responsible incident-control pairs require a non-empty contingency, meaning single-control ablation recovers only 0.36–7.45% of truly responsible controls. Across nearly 260,000 user-target incidents from three real datasets, the framework identifies which system controls are pivotal for disputed recommendations and generates verifiable certificates with paired serving executions. The findings demonstrate that explicit serving semantics and checkable audit witnesses are essential for meaningful accountability in complex recommender pipelines.
- Quality assurance
- AI policy
Research
Effects of AI-generated feedback on l2 writing: a PRISMA 2020 systematic review of performance, revision, and engagement outcomes
Kristian Burhan
JRTI (Jurnal Riset Tindakan Indonesia) · 2026-09-07
This PRISMA 2020 systematic review synthesized 20 peer-reviewed empirical studies (2023–2025) examining generative AI feedback on second-language English writing in EFL/ESL contexts. The evidence shows GenAI feedback most consistently supports local accuracy, lexical development, revision activity, and learner engagement, while effects on syntactic complexity, higher-order composing, and long-term retention remain mixed or conditional. The review supports a complementary model where AI handles rapid, high-volume local feedback and teachers address global structure, contextual appropriateness, and affective support. Key limitations include short intervention periods, small samples, and limited longitudinal evidence, with calls for preregistered, more rigorous future designs.
- Workforce
- Quality assurance
Research
How do interdisciplinary pathways shape gender diversity in ICT? Evidence from a quantitative study
Irene Finocchi
Technology in Society · 2026-09-07
This quantitative study examines whether interdisciplinary academic programs in ICT attract more women than traditional computer science programs, using enrollment data across all Italian universities at both bachelor and master levels over multiple academic years. The analysis finds that interdisciplinary programs blending computer science with fields like economics, management, and finance enroll on average 16% more female students than traditional programs, with fields like AI and cybersecurity seeing uplifts exceeding 25 percentage points. The findings suggest that expanding interdisciplinary pathways in higher education could be an effective intervention for reducing gender gaps in ICT.
- Workforce
- AI policy
Research
A Systematic Review on AI and Employee Trust: Examining Transparency, Fairness and Human Oversight in HR Decision Making
Sujit Kumar Mahapatro
International Journal of Computer Information Systems and Industrial Management Applications · 2026-09-07
This systematic review synthesizes 105 studies (drawn from 1,883 records) on how AI adoption in human resource management affects employee trust, focusing on transparency, fairness, and human oversight. Key findings show no clear direct link between algorithmic transparency and employee trust, but procedural and distributive fairness act as mediating factors in that relationship. The review also identifies context-specific barriers in India—including fragmented regulation, digital illiteracy, and cultural factors—that shape how employees respond to AI-driven HR decisions. The authors conclude that building trust in AI HR tools is a sociotechnical challenge requiring interdisciplinary approaches and tailored policy recommendations for HR practitioners, policymakers, and AI developers.
- AI policy
- Workforce
Research
Enhancing the Adaptability of Female-Led Business Households in the Mekong Delta Amidst the Transition to Vietnam’s New Private Economic Development Framework: A Multi-Dimensional Analysis of Policy Absorption, Digitalization, and Sustainable Accounting
Hanh Thi Pham, Trang Phung Thi Quynh, A. Nguyen et al.
International Journal of Sustainable Business Management and Accounting · 2026-09-07
This study examines how female-led business households (FBHs) in Vietnam's Mekong Delta are adapting to sweeping policy reforms—including Resolution 198/2025/QH15 and the abolition of the presumptive tax regime by 2026—that require millions of informal businesses to formalize and digitalize. Using a mixed-methods design with 300 surveyed business owners, 24 interviews, and four case studies, the authors find that adaptation is constrained by time poverty, limited credit access, and uneven digital infrastructure, while AI-enabled accounting tools show promise but face adoption barriers including limited competence and data privacy concerns. The Vietnam Women's Union emerges as a key institutional intermediary improving technology uptake and formalization. The paper concludes that a coordinated 'Support Ecosystem'—combining differentiated policy deadlines, simplified AI-driven accounting, reformed credit access, and union intermediation—is needed to turn the 2026 mandate into an opportunity for women's economic empowerment.
- AI policy
- Workforce
Research
Challenges in Developing Human Resources to Meet the Demands of the Digital Economy: A Multi-Faceted Analysis from the Perspectives of HR Managers and Training Institutions in Vietnam
Do Thi Dinh
IIARD INTERNATIONAL JOURNAL OF ECONOMICS AND BUSINESS MANAGEMENT · 2026-09-07
This qualitative study from Vietnam examines the 'digital skills gap' by interviewing 18 senior HR and training managers at leading enterprises and universities. Findings reveal a systemic mismatch: enterprises need workers with both technical skills (AI/ML, DevOps) and digital-environment soft skills, while training institutions struggle with slow curriculum updates, resource constraints, and superficial industry partnerships. The study frames the gap as a systemic failure rather than any single party's fault, proposing practical remedies such as Work-Integrated Learning, Joint Innovation Labs, and a national Open Digital Skills Framework. The research matters because it clarifies root causes of workforce-readiness shortfalls in a growing digital economy and offers actionable policy and institutional models to align labor supply with demand.
- Workforce
- AI policy
Research
Buying AI, buying compliance? Artificial intelligence investment and corporate fraud
Sifei Li, Hui Zhang, Tianyu Gao
International Review of Economics & Finance · 2026-09-07
Using Chinese A-share listed companies from 2010 to 2023, this study finds that corporate investment in artificial intelligence reduces the likelihood of fraud by lowering agency costs and improving internal control quality. The fraud-reduction effect is stronger in firms with higher AI asset productivity and in regions with lower marketization levels, with AI software assets identified as the primary driver. Additionally, AI investment reduces information disclosure violations more than operational violations and lowers debt financing costs by inhibiting fraudulent activity.
- Enterprise
- Quality assurance
Research
AI, human labor, and the task frontier: automation, complementarity, and the net effect of Artificial Intelligence on employment
Isam Atoba, Mohamed Amine Korchi
International Review of Applied Finance Economics and Management · 2026-09-07
This paper extends the Acemoglu-Restrepo task-based framework to distinguish between automation-type AI (robots, RPA) and complementarity-type AI (text mining, NLP, machine learning), finding that the latter is being adopted 2.6 times faster across 27 EU member states. Using Eurostat data covering 59 NUTS-2 regions, the authors show that regions exposed to automation AI lose jobs (β = −5.19) while regions exposed to complementarity AI regain them (β = +2.37), yielding a slightly positive net employment effect (+0.04 percentage points) compared to a typical robot-driven loss of −0.37 percentage points. The findings suggest that the net labor market impact of AI depends critically on whether the technology replaces or augments human judgment, with policy and firm-level adoption choices determining which type dominates.
- Workforce
- AI policy
Research
EU policy on artificial intelligence integration in manufacturing
Т. В. Полозова, Mykhailo Salikhov
SCIENTIA FRUCTUOSA · 2026-09-07
This study examines how EU artificial intelligence policy from 2018 to 2025 has driven AI adoption in manufacturing, using normative-doctrinal, comparative, and economic-statistical methods. The share of enterprises using at least one AI technology rose from 6.93% in 2021 to 17.27% in 2025, with the highest growth in pharmaceuticals, computer products, chemicals, metallurgy, and refined petroleum. The authors argue that EU regulatory frameworks provide a governance model—spanning policy framing, regulatory mechanisms, and evaluation cycles—that countries like Ukraine can adapt for their own AI agendas.
- AI policy
- Enterprise
Research
Data-driven real estate valuation, part II: data, uncertainty and governance for reliable automated valuation models in practice
Miroslav Despotović, Wolfgang A. Brunauer, David Koch
Journal of Property Investment and Finance · 2026-09-07
This review paper examines what is required for automated valuation models (AVMs) used in real estate to be reliable, fair, and production-ready in high-stakes settings. It finds that data quality sets the upper bound on AVM performance more than model complexity, and that calibrated prediction intervals and periodic backtesting are prerequisites for trustworthy use. The paper also identifies ESG signals and synthetic data as promising but governance-sensitive expansions to AVM inputs, and concludes that hybrid human-AVM architectures are the most defensible operating model under current regulatory expectations. The findings carry direct implications for AVM providers, institutional users, and regulators responsible for model risk management.
- Enterprise
- Quality assurance
- AI policy
Research
Automation, employment and economic outcomes across European industries
Orkhan Mammadov
Economics of Innovation and New Technology · 2026-09-07
This study uses a panel dataset of 32 European countries and 18 industries from 1994 to 2019 to examine how industrial robot adoption and AI exposure relate to employment, wages, output, and gross value added. Robot density is consistently associated with employment reductions in routine-intensive activities, while higher AI exposure is linked to smaller estimated employment losses from robots, consistent with task reallocation in knowledge-intensive settings. Effects vary across sectors and between Western and Central/Eastern European economies, highlighting the context-dependent nature of automation's labor market impacts and the need for tailored policy responses.
- Workforce
- AI policy
Research
HARMONIZING GLOBAL AI GOVERNANCE: A COMPARATIVE STUDY OF CORPORATE COMPLIANCE FRAMEWORKS UNDER THE EU AI ACT V. EMERGING US AND ASIAN REGULATORY MECHANISMS
NANDANA B S
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-07
This paper analyzes how multinational firms must navigate incompatible AI regulatory regimes—principally the EU AI Act (Regulation 2024/1689) alongside emerging US and Asian frameworks—arguing that genuine harmonization at the level of binding obligation remains largely a fiction. The authors find that convergence is occurring not through treaties but via the extraterritorial reach of EU rules (the 'Brussels effect') and private technical standards, producing a stratified compliance market where the most demanding jurisdiction sets the operational floor. For India, the paper recommends a calibrated interoperability strategy built on existing digital public infrastructure rather than wholesale adoption of either EU or US approaches. The work is grounded in comparative regulatory analysis and has direct relevance for enterprise compliance planning across jurisdictions.
- Enterprise
- AI policy
Research
HARMONIZING GLOBAL AI GOVERNANCE: A COMPARATIVE STUDY OF CORPORATE COMPLIANCE FRAMEWORKS UNDER THE EU AI ACT V. EMERGING US AND ASIAN REGULATORY MECHANISMS
NANDANA B S
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-07
This paper examines how multinational firms must navigate incompatible AI regulatory regimes across the EU, US, and Asia, using the EU AI Act (Regulation 2024/1689) as the central reference point. It argues that 'harmonization' across jurisdictions is largely illusory at the level of binding law, with real convergence driven instead by the extraterritorial reach of EU regulation (the 'Brussels effect') and private technical standards rather than treaties. The paper concludes that a stratified compliance market is emerging where the most demanding jurisdiction sets the operational floor, and recommends that India pursue a calibrated interoperability strategy grounded in its existing digital public infrastructure rather than simply copying EU or US approaches.
- AI policy
- Enterprise
Research
ARTIFICIAL INTELLIGENCE-ENABLED REGTECH FOR KYC, FINANCIAL COMPLIANCE, AND REGULATORY REPORTING: APPLICATIONS, RISKS, AND FUTURE DIRECTIONS
Yun Li, HongZhe Zhang
World journal of economics and business research. · 2026-09-07
This narrative review examines how AI-enabled regulatory technology (RegTech) can support financial institutions in know-your-customer (KYC) procedures, compliance monitoring, investigations, and regulatory reporting by combining machine learning, graph analytics, natural language processing, and workflow automation. The paper finds that while these tools can improve data reconciliation, alert prioritization, and automated reporting, their effectiveness is constrained by incomplete labels, class imbalance, concept drift, opaque models, and privacy risks. The authors argue that responsible deployment requires layered architectures with traceable controls, human oversight of consequential decisions, and full auditability of data transformations and model versions. AI-enabled RegTech should be treated as governed decision support rather than an autonomous substitute for legal interpretation or regulatory accountability.
- Enterprise
- AI policy