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
When Optimization Becomes Manipulation: Defending Generative Search against Malicious Generative Engine Optimization
Haozhang Li, Yangguang Shao, Xinjie Lin et al.
arXiv · 2026-09-02
This paper addresses the problem of Generative Engine Optimization (GEO), where web documents are rewritten to manipulate generative search engines into citing them. The authors propose GEO Defender, a two-stage defense system consisting of a Shield Reranker and Training-Free Shield Generation (TFSG) that requires no fine-tuning of the target LLM. Experiments across five LLMs and seven GEO attacks show GEO Defender reduces average attack success rate from 50.32% to 6.20% while retaining 94.12% of benign-evidence use and preserving answer quality. This matters for the integrity of AI-powered search systems, which are increasingly used to surface and synthesize information at scale.
- Quality assurance
- Enterprise
Research
Examining the Vulnerability of Multi-Agent Medical Systems to Human Interventions for Clinical Reasoning
Benjamin C Liu, Dillon Mehta, Rishi Malhotra et al.
arXiv · 2026-09-02
This study examines how human interventions at 'fault points'—moments when AI agent reasoning is most vulnerable—affect the diagnostic accuracy of multi-agent medical systems. Using the MedQA dataset with simulated doctor-patient conversations, the researchers found that correct interventions improved baseline diagnostic accuracy by up to 40%, while incorrect or bias-related interventions degraded performance by up to 6% and increased diagnostic drift and uncertainty. The work also revealed that simulated agents exhibit cognitive biases analogous to those seen in real-world clinical practice, such as premature closure and susceptibility to misleading cues. These findings suggest that strategically identifying and guiding fault points could be a practical mechanism for improving diagnostic robustness in AI-driven clinical reasoning systems.
- Quality assurance
- Workforce
Research
FUSE: An Evaluating Framework for Dangerous Capabilities of LLMs
Zhengyi Jin, Ru Zhang, Xiao Chen et al.
arXiv · 2026-09-02
FUSE introduces a modular evaluation framework that measures the dangerous capabilities of large language models (LLMs) across three independent dimensions: Knowledge (what a model knows about hazards), Defense (how well it refuses harmful requests), and Harm (how damaging its outputs are when it does comply). Applied to 12 commercial LLMs using a chemical-biological module and a cyber pilot, the framework finds that models with similar knowledge levels can differ sharply in refusal resilience, and that strong refusal behavior does not guarantee less harmful outputs when compliance occurs. Critically, a temporal analysis shows dangerous capability has not monotonically declined over time—newer models deepen domain knowledge while only partially improving defenses—meaning scaling and alignment progress do not uniformly reduce safety risk. The framework provides standardized, reproducible capability profiles intended to support governance and oversight of frontier AI systems.
- AI policy
- Certifications
Research
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction
Jaehoon Jeong, Jay-Yoon Lee
arXiv · 2026-09-02
This paper addresses 'Guilty Bias' in Legal Judgment Prediction (LJP) models, where training data skewed toward prosecutorial perspectives and guilty outcomes causes models to blindly accept guilt. The authors introduce OBJECTION, an inference-time pipeline that embeds an Adversarial Lawyer Agent into a three-step legal reasoning process (offense, unlawfulness, culpability) to inject defense arguments and challenge presumptions of guilt. Evaluated on a new dataset of 3,400 real-world innocent cases, OBJECTION reduces the False Guilty Rate from 82.93% to 16.69% compared to the state-of-the-art baseline. The work represents meaningful progress toward aligning Legal AI systems with the presumption of innocence.
- Quality assurance
- AI policy
Research
AI agents reshape consensus formation in human groups
Lin Chen, Ziyi Liu, Xia Hu et al.
arXiv · 2026-09-02
This study examines how varying proportions of LLM agents in mixed human-AI groups affect consensus formation in a collaborative communication task. The researchers find three regimes: low agent proportions allow humans to lead consensus, intermediate proportions disrupt convergence, and high proportions produce strong consensus driven by agent-led conventions. Critically, human-led and agent-led consensus differ in character—human norms are more concrete and grounded in real-world analogies, while agent-led norms are more abstract and geometrically segmented. The findings suggest that AI composition, agent proportion, and transparency are important design variables for human-AI collaborative systems, with implications for how group norms and collective behavior emerge.
- Workforce
- AI policy
Research
Disease Burden over Skin Tone: Decomposing the Dermatology-AI Generalization Gap
Nirajan Kunwor, Sanjaya Poudel, Quoc-Huy Trinh et al.
arXiv · 2026-09-02
This paper investigates why dermatology AI models fail to generalize in resource-constrained settings, separating the effects of skin-tone underrepresentation from disease-distribution shift. Testing a cancer-trained ResNet-50 baseline, two dermatology foundation models (DermLIP and MONET), and a general-purpose vision model (DINOv3) on two datasets (DDI and SCIN), the authors find that disease-distribution shift is the dominant factor: the cancer baseline drops from 0.62 to 0.21 balanced accuracy on unfamiliar clinical conditions, while skin-tone gaps are smaller (0.10–0.18) and inconsistent. Representation analysis shows that cancer-specialized features poorly cluster unfamiliar conditions (kNN purity lift +0.06 over chance), whereas dermatology-pretrained features retain stronger transferable structure (+0.23). Critically, starting from dermatology foundation models, roughly ten labeled examples per clinical category recovers most attainable performance, suggesting lightweight adaptation as a practical path forward.
- Quality assurance
- AI policy
Research
READY or Not: Reliable Enterprise Agent Deployment
Veronica Chatrath, Bryan Zhu, Jingxuan Fan et al.
arXiv · 2026-09-02
This paper introduces READY (Reliable Enterprise Agent Deployment), a framework that evaluates AI agents not just on task accuracy but on whether they can meet a specified reliability level under human oversight at acceptable cost. Rather than asking how well an agent performs autonomously, READY asks under what conditions and at what cost reliable deployment is achievable, selecting the minimum-cost oversight policy that satisfies a given reliability target and statistically validating it on held-out cases. A clinical-audit case study with 16 agent systems and 750 cases demonstrates that two agents with nearly identical autonomous accuracy (72.8% vs. 72.5%) require meaningfully different human review rates (39.2% vs. 29.6%) to hit the same 76% reliability target — a difference invisible to standard benchmarks. The framework provides a principled, evidence-based basis for comparing agent systems, setting oversight requirements, and making deployment decisions in enterprise workflows.
- Enterprise
- Quality assurance
Research
Beyond Outcome Gaps: Process-Aware Fairness Diagnosis for LLM-based Multi-Agent Decision Systems
Yiran Zhao, Lu Zhou, Liming Fang et al.
arXiv · 2026-09-02
This paper presents SCOPED-Hiring, a fairness diagnosis pipeline for LLM-based multi-agent hiring systems that looks beyond final hire rates to examine the decision-making process itself. By constructing controlled resume variants and logging over 311K structured decision trajectories, the system reveals that balanced overall hiring outcomes can mask hidden unfairness within the process—such as career gaps triggering suspicion, proxy cues shaping qualification judgments, and identity cues causing unequal investigation of candidates. Targeted repairs guided by this process-level diagnosis reduced the total layered burden of unfairness by 72.3% while shifting the overall hire rate by only 1.86 percentage points, demonstrating that process-aware auditing enables more effective and precise fairness interventions than outcome-based audits alone.
- AI policy
- Enterprise
Research
Monitoring Web Agents Without Internal Signals: Observable Trajectories and Key-Step Supervision
Sitong Pan, Yipeng Shen, Yilin Lu et al.
arXiv · 2026-09-02
This paper investigates how to monitor web-browsing AI agents for risk of failure when internal model signals (like token probabilities) are unavailable. The researchers develop two types of observable trajectory features—'Macro' features summarizing agent-environment behavior across steps, and 'Micro' features measuring consistency of intentions and actions via repeated black-box queries—and introduce a 'key-step boundary' labeling method that identifies the first critical uncorrected error leading to failure. Tested across the WebArena-Lite and Online Mind2Web benchmarks with five different AI backbones, these observable signals match the performance of internal-signal baselines and support early intervention and transfer to new website categories. The findings matter for quality assurance of AI agents deployed in real-world web environments, where operators often lack access to model internals.
- Quality assurance
Research
Privacy Washing: Detecting Internal Contradictions in Privacy Policies
Thomas Brackin
arXiv · 2026-09-02
This paper introduces 'privacy washing' — the phenomenon where privacy policies contain internal contradictions that undermine stated commitments — and develops a four-stage automated pipeline using large language model panels to detect them. Applied to two corpora of website privacy policies (123 from 2026 and 115 from 2015), the pipeline finds panel-confirmed contradictions in 12.2% and 36.5% of companies respectively, with third-party data sharing being the dominant contradiction category in both sets. The results are partially reproduced in a stability re-run using entirely different models, suggesting the findings are not artifacts of specific LLM choices, though the authors caution that panel verdicts are not validated against human expert judgment and prevalence figures represent lower bounds. The work matters for policy and quality-assurance because it provides a scalable method to identify potentially misleading privacy disclosures that affect consumer trust and regulatory compliance.
- AI policy
- Quality assurance
Research
Knowing Is Not Enough: Information Retrievability as a Precondition to Effective LLM Oversight
Xinyu Fu, Narayan Ramasubbu, Dennis Galletta
arXiv · 2026-09-02
This paper investigates why workers often fail to catch errors in large language model outputs, arguing the problem lies not just in capability or motivation but in whether oversight-relevant information is mentally accessible at the moment of review. Across two randomized lab-in-the-field experiments with 640 customer-facing employees, the researchers find that having users generate their own explanations during onboarding improves error detection and strengthens recall of verification-relevant reasoning, while periodic retrieval cues help sustain detection accuracy under routine LLM use. The study identifies 'information retrievability' as a distinct precondition for effective human oversight and offers practical, lightweight interventions—self-explanations and daily cues—to keep human oversight resilient as LLM use becomes habitual in organizational settings.
- Workforce
- Enterprise
Research
Plataformas digitais e trabalho: limites constitucionais da gestão algorítmica
Elinaldo Santana Santos Júnior, Gabriel Leal Rezende, Kennedy Vieira Rodrigues
Revista edUCA - Revista Multidisciplinar da Faculdade Católica Paulista · 2026-09-02
This paper examines the constitutional limits that Brazil's 1988 Federal Constitution places on algorithmic management by digital platforms. Through qualitative, bibliographic, and documentary legal-dogmatic research, the authors find that while algorithms promote innovation and efficiency, they can intensify power asymmetries and reduce transparency in decisions affecting workers. The study concludes that free enterprise must be reconciled with the constitutional principles of human labor valorization, dignity, and social justice, requiring standards for transparency, human oversight, and contestability of algorithmic decisions.
- Workforce
- AI policy
Research
The Effects of Artificial Intelligence on Middle Management Skills: A Systematic Literature Review
Semih Sancar, Alper Camcı
Journal of Business Administration and Social Studies · 2026-09-02
This systematic literature review examines how AI integration is reshaping the role and required competencies of middle managers in organizations. Analyzing 39 articles published between 2018 and 2025 via the PRISMA protocol, the study finds that AI-driven automation is displacing routine supervisory tasks and pushing middle managers toward data-driven decision-making, human–AI collaboration design, and psychological empowerment. The findings suggest a fundamental competency shift requiring higher-order cognitive, analytical, and social skills from this managerial tier.
- Workforce
- Enterprise
Research
Generative AI and Technical Debt in Software Development: Productivity Gains versus Code Quality and Security Risk
Rishu Roy, Vishal Khasgiwala, Vanita Joshi
International Journal of Computer Information Systems and Industrial Management Applications · 2026-09-02
This empirical study of 412 software professionals, 1,240 code commits, and 22 practitioner interviews finds that generative AI coding assistants (e.g., GitHub Copilot, ChatGPT, Amazon CodeWhisperer) significantly boost short-term developer productivity (β = 0.48, p < .001), but without adequate governance, AI-assisted code carries 37% higher technical-debt density and 41% higher security-vulnerability density than human-written code. Productivity gains are also lower for experienced developers when accounting for maintenance burden (β = 0.06, p > .05). The study identifies governance maturity as the key moderating factor and proposes a three-dimensional governance model covering velocity, maintainability, and security to manage what the authors reframe as a socio-technical risk rather than a purely technical one.
- Quality assurance
- Enterprise
Research
AEGIS: An Auditable Evidence-Governed Interface for Cost-Aware AI Harness Selection in Finance
Xiaozhen Wang, Francois Buet-Golfouse
INRIA a CCSD electronic archive server · 2026-09-02
AEGIS is a framework for selecting among AI model 'harnesses' in bank onboarding by jointly considering task information, policy, evidence, budget, deadlines, and runtime health. The paper proves that ignoring governance information (evidence and approvals) leads to quantifiable expected regret, and demonstrates empirically across 1,000 borrowers that governance-conditioned routing reduces normalized loss by roughly 0.099 compared to task-only routing, and cuts operations loss by 33.2%. Across 629 matched cells in four tool-use suites, AEGIS dominates suite-fixed approaches on safety, attack success, latency, and call counts. The framework uses LP duality to produce auditable supporting prices and signed evidence packages to certify lineage, making it directly relevant to compliance and governance in financial services.
- Enterprise
- AI policy
Research
Governing artificial intelligence in a developing economy: lessons from the EU, South Korea, and Japan for Thailand’s regulatory design
Suphawatchara Malanond, Pongsarun Boonyopakorn
Cogent Social Sciences · 2026-09-02
This article compares AI regulatory approaches from the EU, South Korea, and Japan to inform governance design for Thailand as a developing economy. Using institutional economics (North, Coase, Williamson) and policy transfer theory, it evaluates three models across dimensions including legal certainty, rights protection, innovation compatibility, and institutional feasibility. The authors argue Thailand should adopt a context-sensitive hybrid framework combining voluntary principles, targeted binding rules for high-risk AI, and an eventual consolidated framework law. The work contributes to literature on AI governance in developing economies by showing how comparative regulatory analysis can balance rights protection with innovation and international alignment with domestic fit.
- AI policy
Research
Accountability, Integrity: AI Policy in Public Libraries
Kathryn FitzGerald, Benjamin Charles Germain Lee
Public Library Quarterly · 2026-09-02
This study conducted an environmental scan of over 200 American and Canadian public library websites to identify and analyze AI policy statements, finding 32 such policies. The analysis reveals that all identified policies reference at least two library values, with privacy, security, and responsibility cited most frequently. The findings highlight how public libraries are navigating AI adoption while attempting to uphold core institutional values, filling a gap in published research on this topic.
- AI policy
Research
Creative agency on generative AI music platforms: Reframing artistic identity and the creative subject
Ugur Baloglu
Opus uluslararası toplum araştırmaları dergisi · 2026-09-02
This study uses critical discourse analysis to examine how four generative AI music platforms (Suno, Udio, AIVA, and Soundraw) construct user agency and creative identity through their interface texts, marketing materials, and terms of use. Three key tendencies were found: a tension between attributed creative agency and contractual ownership limits, the use of metaphors (magic, tool, partnership) to obscure algorithmic operations, and a 'democratization' rhetoric that may devalue professional musical expertise. The study also finds that none of the platforms disclosed specific training datasets, nor any consent, attribution, or compensation mechanisms for rights holders, raising significant concerns about transparency and professional musicians' standing in the industry.
- Workforce
- AI policy
Research
The sustainable AI development contradiction: Chinese policy on sustainability and the costs of AI development
Levi Checketts
AI & Society · 2026-09-02
This review paper examines the ecological costs of AI development specifically in the context of China, identifying three interconnected environmental problems: AI's carbon emissions, water consumption, and harm from hardware production and disposal. The author argues that treating 'AI for sustainability' as a solution to climate change constitutes a 'moral hazard,' as it may discourage governments from taking immediate action toward negative carbon emissions. Drawing on World Meteorological Organization forecasts and the political philosophy of the Communist Party of China, the paper concludes that China's stated values require a cautious, limited approach to AI development oriented toward ecological sustainability.
- AI policy
Research
Guidance for laboratory implementation, governance and continuous assurance of artificial intelligence in histopathology
Emad A. Rakha, Jelle Wesseling, Anikó Kovács et al.
Archiv für Pathologische Anatomie und Physiologie und für Klinische Medicin · 2026-09-02
This paper proposes a practical governance framework for pathology laboratories implementing commercially approved AI systems in routine histopathology practice. Rather than focusing on AI development or regulatory approval, it addresses the local verification, ongoing quality assurance, human oversight, and risk-proportionate governance strategies needed after a system enters clinical use. The framework distinguishes different AI application types—diagnostic, biomarker, workflow, and generative—and outlines tailored implementation and monitoring strategies for each throughout the AI operational lifecycle. It is intended to support laboratories, accreditation bodies, and healthcare providers in safely and sustainably integrating AI while maintaining diagnostic quality and patient safety.
- Quality assurance
- Certifications
Research
Leakage-Aware Cross-Dataset Evaluation of Prompt Injection Detection Using Classical Machine Learning and Transformer Models
Oğuzhan KİLİM
Yalvaç Akademi Dergisi · 2026-09-02
This paper introduces a leakage-aware cross-dataset evaluation framework for detecting prompt injection attacks against Large Language Model-based systems. The authors cleaned a dataset of over 426,000 records (removing duplicates, conflicting labels, and overlapping prompts with the external test set) and evaluated classical ML models (SGD Logistic, Linear SVM) and Transformer models (DistilBERT, DeBERTa-v3-small) both internally and on an independent 5,000-example test set. While internal performance was near-perfect (0.997–1.000), external performance dropped significantly, with the best model (DeBERTa-v3-small) achieving only 0.7360 accuracy and 0.7119 attack sensitivity. The findings demonstrate that internal validation alone is insufficient for reliable prompt injection detection, and that independent external validation, leakage checks, and domain shift analysis are essential for trustworthy model evaluation.
- Quality assurance
- AI policy
Research
Responsible Artificial Intelligence in Courts: A Four‐Test Framework
Kwan Yiu Cheng
AI & Innovation · 2026-09-02
This paper proposes a four-test framework for evaluating whether AI systems used in courts meet legal, procedural, and institutional standards for responsible deployment. The four tests—Law Accuracy, Judge Control, Litigant Contestability, and Public Transparency—must all be satisfied simultaneously, with accurate legal training data serving as the foundational requirement enabling the other safeguards. The framework addresses risks such as AI outputs based on outdated law, erosion of judicial authority, inability of litigants to challenge AI-influenced decisions, and lack of public awareness. The work is directly relevant to governance and regulatory frameworks governing AI in high-stakes public institutions.
- AI policy
- Certifications
Research
An Adaptive Dual-Loop Artificial Intelligence Framework for Integrated Disease Surveillance and Health Workforce Learning in Low-Resource Health Systems
Kenneth Goga Riany, Agnes Linus Muthoni, Marcellah Onsomu et al.
Medinformatics · 2026-09-02
This paper presents and evaluates an Adaptive Dual-Loop AI Framework that simultaneously addresses disease surveillance and health workforce training in low-resource settings, tested across 24 facilities and 312 health workers in Kenya over 10 months. The surveillance loop achieved an AUC of 0.94 versus 0.81 for ARIMA and 0.72 for threshold baselines, delivering a mean 3.6-day early-warning lead and cutting outbreak-response time from 12.9 to 6.2 days. The workforce learning loop raised competency scores from 55.0% to 75.4% and reporting completeness from 62% to 92%. The findings demonstrate that coupling predictive epidemiological intelligence with embedded adaptive learning yields compounding gains neither component achieves independently, offering a practical model for strengthening public-health capacity where resources are scarce.
- Workforce
- AI policy
Research
Authentication status and AI triage concordance among care seekers in a US health system
Bilal A. Naved, Quintan M Slott, Dr. Adeel Malik et al.
npj Health Systems · 2026-09-02
This prospective cohort study of 6,772 users of an AI symptom-checker (triage tool) deployed across a large US health system found that alignment between a user's pre-stated care intent and the AI's recommendation strongly predicts downstream digital engagement: call-to-action interaction was roughly twice as high when the two matched. Authentication status also mattered—unauthenticated users were far more likely to be escalated by the AI, while 42% of authenticated portal users seeking an office visit were redirected. Non-engagers reported they would seek care off-platform, raising continuity-of-care concerns. The findings highlight how AI triage design and user-experience choices shape whether patients follow through within a health system's digital ecosystem.
- Enterprise
- Quality assurance
Research
EAP Teachers’Agency in Flux and Transformation: A Longitudinal Study in the GenAI Context
Can Chen
International Journal of TESOL Studies · 2026-09-02
This 24-month longitudinal case study follows two EAP teachers—one novice, one experienced—at a Sino-British university navigating curriculum cuts and rapid GenAI integration. The study finds that professional agency is jointly shaped by emotions and identity: the experienced teacher reframed anxiety into adaptive growth, while the novice exhibited strategic compliance alongside internal resistance due to emotional turbulence and identity fragmentation. The authors propose an 'Emotion-Identity-Agency Nexus in Flux' model showing how macro forces like neoliberal KPIs and GenAI disruption filter through teachers' emotional and identity resources to produce adaptation, compliance, or resistance. The findings call for emotional scaffolding for novice teachers and differentiated performance evaluation policies in English-medium instruction contexts.
- Workforce
- AI policy