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
Automated Testing of LLM-Based Post Hoc Explainers Using Model Checking as an Oracle
Dennis Gross, Helge Spieker
arXiv · 2026-08-31
This paper tackles the challenge of verifying whether large language model explanations of decision-making policies are actually correct. The authors use probabilistic model checking as an automatic oracle to grade LLM answers against exact reference results, and they develop a taxonomy of query categories to systematically generate and prioritize test cases. Testing across seven MDP environments reveals significant performance gaps: a reasoning model passes 85% of test cases, a mid-size model passes 70%, and a 1B model falls below random baseline. The findings provide concrete evidence about how trustworthy LLM-generated explanations are, with direct implications for quality assurance of AI systems deployed in explainability roles.
- Quality assurance
Research
Collapsibility of Performance Metrics in Clinical Predictive AI
João Matos, Ben Van Calster, Richard D. Riley et al.
arXiv · 2026-08-31
This paper examines whether 15 commonly used performance metrics in clinical predictive AI are 'collapsible' — meaning the overall population metric equals the weighted average of subgroup metrics. The authors show that five metrics, including the widely used AUC (c-statistic), calibration intercept, calibration slope, expected calibration error, and Nagelkerke R², are non-collapsible, while ten others such as Brier score, F1-score, and net benefit are collapsible. Non-collapsibility means that overall and subgroup performance values can differ spuriously even when no real disparity exists, potentially misleading AI fairness evaluations. The authors argue that explicitly reporting collapsibility properties is essential for transparent and interpretable fairness assessments of clinical AI models.
- Quality assurance
- AI policy
Research
CHASE: How Content Ecosystems Are Reshaped When Ranking Is the Only Target
Qianwen Gao, Zichang Su, Yiwen Hou et al.
arXiv · 2026-08-31
This paper introduces CHASE, a simulation framework that studies what happens to content ecosystems when creators repeatedly optimize documents to rank highly in LLM-based retrieval systems. Running 20 rounds of iterative rewriting across six domains, the study finds that quality-ranking alignment consistently declines (mean Spearman's rho change of -0.068), meaning documents that score well on ranking signals become progressively less aligned with independently judged quality. A control condition confirms the effect is driven by adaptation to ranking incentives rather than rewriting alone. The findings matter for AI-powered search and content platforms because they show that optimizing purely for LLM ranking can degrade overall content quality and reshape creator incentives at a population level.
- Quality assurance
- Enterprise
Research
AI Morbidity and Mortality: A Framework for Clinical AI Failure Review
Paulius Mui, Dean F. Sittig, Steve Labkoff et al.
arXiv · 2026-08-31
This paper proposes 'AI Morbidity and Mortality' (AI M&M), a structured, blameless framework for reviewing individual clinical AI failures and near-misses — modeled on traditional medical morbidity and mortality conferences. The framework classifies each failure event across four linked dimensions (Trigger, Mechanism, Clinical Pathway, and Corrective Action) and is demonstrated through five illustrative outpatient medication and clinical decision-support cases, with two independent reviewers reaching agreement across all 20 axis-level classifications. It is designed to complement — not replace — existing model monitoring, patient safety reporting, and regulatory oversight by converting individual AI-in-workflow failures into actionable institutional learning. The work matters because it addresses a significant gap: neither aggregate model monitoring nor traditional safety reporting is well-suited to explaining how risk emerges from interactions among AI systems, clinicians, workflows, and institutional controls.
- Quality assurance
- AI policy
Research
Co-Annotator: Expert-Distilled ViT and VLM for Visual and Documentation Guidance in Age-Related Macular Degeneration
Ziheng "Leo" Li, Benjamin Freeman, Akshay Raman et al.
arXiv · 2026-08-31
Co-Annotator is an AI system that distills expert gaze and dictation into two complementary guidance tools for ophthalmology: a gaze-aligned Vision Transformer that highlights fixation-aligned areas of interest in retinal OCT images, and an ontology-bounded vision-language model that pre-fills editable biomarker summaries for age-related macular degeneration. In controlled studies with ophthalmology residents across two academic institutions, combining both modalities simultaneously increased correct diagnoses per minute by 40% and reduced comment editing time by 67%, without compromising diagnostic accuracy. The AOI guidance produced lasting perceptual efficiency gains even after guidance was removed, while VLM guidance more than doubled biomarker documentation breadth. The work demonstrates that expert-distilled multimodal AI can simultaneously address visual search overhead and documentation burden in real clinical workflows.
- Workforce
- Quality assurance
Research
Generating Workflow DAGs from Natural Language with Non-Reasoning LLMs
Anand Iyer, Bhanu Khetharpal, Srinivas Upadhya et al.
arXiv · 2026-08-31
This paper tackles the challenge of converting natural-language routing rules written by business administrators into executable workflow graphs (DAGs) for enterprise contact centers. The authors identify an 'emission-density bottleneck' — models accurately select correct graph nodes but increasingly misconfigure attributes and Boolean logic as graph complexity grows — and address it with a neuro-symbolic decomposition approach that offloads combinatorial graph construction to a deterministic compiler guided by a compact intermediate representation. On a 635-rule synthetic benchmark, the full system achieves approximately 89% LLM-judge validity, ~90% exact-match condition accuracy, and 99–100% valid JSON, while using roughly half the prompt tokens of a monolithic approach; on GPT-5.3-chat, judge validity improves by 24 percentage points, reaching statistical equivalence to a reasoning model without its cost. The work demonstrates that non-reasoning LLMs can meet production-relevant quality thresholds for enterprise workflow automation when paired with structured compilation techniques.
- Enterprise
Research
The Differential Reasoning Router: Operationalizing Cost-Aware LLM Annotation in E-commerce
Cheng Lyu, Jingyue Zhang, Vinny DeGenova et al.
arXiv · 2026-08-31
The Differential Reasoning Router (DRR) is a cost-aware framework for using large language models to annotate structured product data in e-commerce, designed specifically for the 'cold-start' problem where labeled data is scarce and the value of expensive reasoning models is unknown. DRR estimates separate success probabilities for a standard model and a reasoning model at both the sample and business-rule levels, routing easy cases directly, reserving reasoning for cases where it is likely to improve outcomes, and escalating ambiguous cases to human reviewers. In a production e-commerce deployment, the framework matches the accuracy of the strongest confidence-based router while achieving more than 60% savings in reasoning-token costs. This matters for enterprise AI adoption because it offers a principled path from human-heavy annotation at launch toward high-confidence automated workflows, reducing operational costs without sacrificing accuracy.
- Enterprise
- Quality assurance
Research
SIR: Self-improving Red-teaming for Compute Use Agents
Chen Xiong, Zhiyuan He, Pin-Yu Chen et al.
arXiv · 2026-08-31
SIR introduces a black-box indirect prompt injection (IPI) attack framework targeting computer use agents (CUAs)—vision-language models that control a real operating system via mouse, keyboard, and terminal. The system composes stealthy injections from reusable plain-language principles and uses an iterative feedback loop to diagnose failed attacks and distill new bypass strategies. Tested against three frontier CUAs, SIR raises attack success rates dramatically, for example from 4% to 24% on Claude Opus 4.8 and from 0% to 28% on Gemini 3.5 Flash, while benign tasks still complete; attack principles also transfer across model architectures. These findings reveal that fixed, hand-written safety benchmarks may significantly underestimate the risk posed by adaptive adversaries targeting deployed AI agents.
- Quality assurance
- AI policy
Research
GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space
Boqi Chen, Xudong Liu, Yunke Ao et al.
arXiv · 2026-08-31
GPAgentBench-2K introduces a new benchmark for evaluating large language models as primary-care clinical agents, built from expert-validated real-world GP encounter records and framed as a Constrained Markov Decision Process (CMDP) with six foundational clinical actions and safety-informed abstention. Evaluating 16 state-of-the-art LLMs reveals that performance degrades significantly as the clinical action space scales, and that even the most accurate frontier models violate safety constraints in over half of high-risk cases—a 'clinical quality-safety gap.' A reinforcement learning baseline using Constrained Group Relative Policy Optimization (C-GRPO) improves over unconstrained methods but still falls well short of clinically acceptable safety standards. The findings highlight critical limitations in deploying current LLMs as autonomous clinical agents and set a reference point for future research aimed at safer clinical AI.
- Quality assurance
- AI policy
Research
A Discrepancy Evaluation of Workforce Capacity and Emerging Skills: A Utah Case Study
Lendel Narine, Paul Hill, Andree Walker-Bravo
Administrative Sciences · 2026-08-31
This Utah case study surveyed 226 hiring managers to measure gaps between how important employers rate workforce competencies and how satisfied they are with employee proficiency in those areas. Using the Ranked Discrepancy Model, the study found the largest gaps in time management, critical thinking, conflict resolution, ethical judgment, and verbal communication. Looking ahead, employers flagged AI and machine learning as the most disruptive emerging technology and ranked communication, collaboration, analytical thinking, and lifelong learning as the competencies that will matter most over the next decade. The findings are intended to guide recruitment strategies, educational program design, and workforce development policy.
- Workforce
- AI policy
Research
Prohibiting Political Deepfakes in Election Campaigns Based on the Principles of Honest and Fair Elections
Fahmi Nur Wicaksono
Adagium. · 2026-08-31
This normative legal study examines regulatory gaps in Indonesia's handling of AI-generated political deepfakes during election campaigns. Analyzing existing laws—including the Election Law and the Electronic Information and Transactions (ITE) Law—the authors find current frameworks lack specific mechanisms to counter AI-based identity manipulation that threatens electoral integrity and voter autonomy. The paper proposes a proportionate legal framework comprising explicit prohibitions on political identity falsification, mandatory AI content labeling, rapid content removal procedures, and coordinated oversight among Indonesia's electoral and digital regulatory bodies. The proposed model aims to protect voter intent while preserving legitimate expression such as political satire.
- AI policy
Research
Do green innovation and resilience mediate the relationship between AI-enabled dynamic capabilities and sustainable supply chain performance?
Kexing Li, Glauco De Vita, Mahdi Bashiri et al.
Cleaner Logistics and Supply Chain · 2026-08-31
Drawing on survey data from 634 managers across 423 Chinese construction firms, this study finds that AI-enabled dynamic capabilities improve sustainable supply chain performance by boosting firms' ability to sense, seize, and reconfigure resources. Both green innovation and supply chain resilience significantly mediate this relationship: green innovation proactively improves environmental performance while resilience enables reactive adaptation to disruptions. The findings suggest construction firms should strategically invest in AI-enabled dynamic capabilities to meet regulatory requirements and build competitive, sustainable supply chains.
- Enterprise
- AI policy
Research
Humanoid Robotics Adoption and Corporate Investment Efficiency
Anqi Wang, Yanbo Wang, Xin Li et al.
Journal of Global Information Management · 2026-08-31
Using a panel of 412 Chinese A-share listed firms from 2015 to 2022, this study finds that humanoid robotics adoption (HRA) is significantly negatively correlated with investment inefficiency (β = -0.312), meaning firms that adopt humanoid robotics tend to allocate capital more efficiently. A composite corporate governance index positively moderates this relationship (interaction β = -0.148), with non-state-owned firms and those with greater board independence benefiting most. The findings extend agency theory to intelligent manufacturing and offer direct implications for China's 15th Five-Year Plan, suggesting that governance quality is a key condition for robotics adoption to translate into efficient investment rather than wasted capital.
- Enterprise
- AI policy
Research
Are You Qualified, ChatGPT? Examining Clinical Skills and Competencies of ChatGPT in Delivering Systemic Interventions
Gizem Erdem, Nilüfer Kafesçioğlu, Gökçenay Başer
Journal of Marital and Family Therapy · 2026-08-31
This study evaluated ChatGPT's clinical and cultural competence in delivering systemic (couples/family) therapy interventions aligned with the AAMFT Code of Ethics. Twelve experienced supervisors, blind to the AI's identity, rated the model as only moderately competent, comparable to an early-career clinician with fewer than 500 hours of experience, with notable weaknesses in pacing, joining, and cultural attunement. The findings highlight concrete limitations of AI-delivered therapy and reinforce the irreplaceable role of human judgment and intuition in clinical practice.
- Quality assurance
- Certifications
Research
Real-Time Audience Sentiment Analytics for Adaptive and Personalized Media Broadcasting
Asep Sutarman, Fandi Ahmad, Thomas Sumarsan Goh
Bridging of Emerging AI and Media Broadcasting (BEAM) · 2026-08-31
This paper presents a real-time audience sentiment analytics framework for adaptive media broadcasting that fuses textual comments, interaction behavior, and temporal engagement signals. A multimodal model combining a transformer-based text encoder, behavioral feature network, temporal attention, and a contextual bandit engine achieved 89.6% accuracy and 88.9% macro-F1, with end-to-end latency of 184 ms processing ~1,420 events per second. In simulated adaptive broadcasting, sentiment-aware personalization improved click-through rate by 15.9%, viewing duration by 12.8%, and audience satisfaction by 10.6% over static broadcasting. The framework incorporates fairness constraints, confidence thresholds, and privacy-preserving aggregation, positioning the system as a decision-support tool rather than an autonomous editorial authority.
- Enterprise
Research
Competing Framings: Variability in AI‐Generated Health Policy Guidance and Its Implications for Global Nursing
Amina Silva, Monica Lino
Nursing Inquiry · 2026-08-31
This paper examines how seven large language models, developed in different socio-technical and geopolitical contexts, respond to a standardized prompt about the global nursing shortage. Through comparative qualitative analysis, the researchers identified four competing policy logics—workforce, efficiency, equity, and mobility—suggesting that AI outputs reflect the institutional environments of their developers rather than neutral, universal knowledge. Each model presented its framing with authority while failing to acknowledge its own situatedness. The authors argue that AI-generated health policy guidance should be treated as situated artifacts and offer a framework to help nurses, educators, and policymakers engage with these tools more critically and safely.
- AI policy
- Workforce
Research
Ethnographies of Human‐AI Collaborations: What Arrangements of Expertise Emerge?
Netta Avnoon
Sociology Compass · 2026-08-31
This review paper maps four patterns of human-AI collaboration identified through ethnographic and sociological research: near-complete AI replacement of human expertise, exploitative algorithmic management (including platform and gig economies), productive co-creation of new expert roles and markets, and outright expert dismissal of AI tools. Drawing on sociology of expertise, management studies, and science and technology studies (STS), the author argues that understanding AI's societal role is inseparable from understanding how expert work is defined and reorganized. The review is relevant to workforce dynamics because it systematically documents how AI reshapes, displaces, surveils, or is ignored by human workers across a range of professional and labor contexts.
- Workforce
Research
Algorithmic Governance And Predictive Artificial Intelligence In Hospital Workforce Management: A Narrative Review Of Burnout Prevention Strategies
Ufuk Burak Karcıoğlu
Kastamonu Üniversitesi Sağlık Bilimleri Fakültesi Dergisi · 2026-08-31
This narrative review examines how predictive AI and algorithmic governance could help address organizational drivers of healthcare worker burnout in hospital settings. The authors find that AI-enabled workforce management may improve workload predictability, reduce administrative burdens, and support workforce planning, but caution that current evidence does not support a definitive causal claim that predictive AI prevents burnout. The review emphasizes that these technologies should be embedded within broader organizational strategies and robust governance frameworks, and calls for longitudinal empirical studies across diverse healthcare systems and professional groups to evaluate both benefits and unintended consequences.
- Workforce
Research
AI-Led or Human-Led? Disclosure of the CSR Decision-Maker, Motive Attribution, and Perceived CSR Authenticity
Keonyoung Park, Dongqing Xu, Jiamin Xie
Journal of Public Relations Research · 2026-08-31
This experiment (n=517 U.S. adults) tested how disclosing an AI system versus a human team as the CSR decision-maker affects public perceptions. Participants assigned to the AI-led condition reported lower competence, fairness, transparency, public-serving motive attributions, and perceived CSR authenticity, even though the described initiative was identical. Capability judgments predicted motive attributions, which in turn predicted perceived authenticity, and individual attitudes toward AI moderated these effects. The findings matter for enterprise communicators and policy discussions around mandatory AI disclosure in corporate communications.
- Enterprise
- AI policy
Research
THE IMPACT OF ARTIFICIAL INTELLIGENCE ON JOB DESCRIPTIONS: EVOLVING SKILLS AND HR CHALLENGES
ABD RAHMAN AHMAD, HAIRUL RIZAD MD SAPRY, ALAA S SALAM et al.
Quantum Journal of Social Sciences and Humanities · 2026-08-31
This systematic review of 52 peer-reviewed studies (2014–2024) examines how AI is reshaping formal job descriptions across occupations. The findings show a decline in routine task specifications and a rise in higher-order cognitive, interpersonal, adaptive, AI literacy, and ethical-judgement competencies as central expectations. The study further finds that these shifts are straining traditional HR processes such as job analysis, workforce planning, and talent acquisition. The authors propose an evidence-based competency framework to guide organizational adaptation as AI tools become more prevalent in work environments.
- Workforce
- Enterprise
Research
Software Engineer Competency Framework in the Era of Generative AI: A Literature Review
Muhamad Anggun Novembra
Jurnal Serumpun Teknik Informatika · 2026-08-31
This literature review synthesizes existing research into a three-pillar competency model for software engineers working alongside generative AI tools such as GitHub Copilot and ChatGPT. The model covers foundational technical skills augmented by AI tool literacy and prompt engineering, cognitive-analytical abilities like critical code review and AI-output verification, and meta-skills including AI governance and ethical judgment—all differentiated across junior, mid-level, and senior career stages. The review also identifies a notable research gap: competency development at the senior engineering tier is substantially understudied compared to earlier career stages. The findings carry direct practical implications for software organizations and educational institutions redesigning workforce development and training pathways in the generative AI era.
- Workforce
Research
Knowing its name, not its nature: word association mapping of student AI cognition and evidence-based micro-credential design in Turkish higher education
İsmail Şan, Hanife Gülhan Orhan Karsak
International Journal of Educational Technology in Higher Education · 2026-08-31
This study surveyed 436 Turkish undergraduate students across four public universities using Word Association Tests to map how students cognitively represent artificial intelligence. Findings show students hold instrumentally fragmented mental models dominated by utilitarian concepts like convenience and speed, while algorithmic transparency, ethical governance, and technical mechanisms are largely absent from their associative networks. Based on these diagnostic results, the authors propose an Ethical AI and Workforce Readiness micro-credential framework intended to bridge the gap between students' functional use of AI tools and deeper conceptual mastery. The work offers a replicable, data-driven methodology for needs-assessment-driven credential design in higher education contexts.
- Workforce
- Certifications
Research
Domain-Specific Contexts versus General Enterprise Contexts: Effects on AI-Driven Agile Requirements
Ajay Roy
International Journal of Computer Applications · 2026-08-31
This study compares two AI tools—Atlassian Rovo and Microsoft Copilot 365—for generating Agile user stories across ten scenarios ranging from basic CRUD operations to complex financial logic. Rovo, which draws on structured development artifacts in Jira and Confluence, outperformed Copilot on INVEST fit (4.6 vs. 3.6), technical accuracy (4.6 vs. 3.1), and reduced human revision time by 60% (3.2 vs. 8.0 minutes). The authors attribute the gap to context quality: structured artifacts reduce AI hallucinations, while unstructured communication logs from Outlook and Teams introduce noise. The findings suggest that domain-specific enterprise AI contexts yield meaningfully better requirements engineering outcomes than general enterprise contexts.
- Enterprise
- Quality assurance
Research
Passenger acceptance of AI-controlled flight operations: Trust and risk perspectives
Cemal Durmuşçelebi
Havacılık ve Uzay Çalışmaları Dergisi · 2026-08-31
This study examines passenger willingness to fly in AI-controlled aircraft using a technology acceptance model tested on 258 participants via structural equation modeling. Results show that trust is the central driver, significantly increasing perceived usefulness and reducing perceived risk, while perceived risk in turn lowers perceived usefulness. Perceived usefulness emerges as the strongest predictor of intention to fly, and perceived ease of use plays no significant role—suggesting that for unfamiliar, high-risk technologies, trust and risk perception matter more than usability. The findings offer practical guidance for developers, airlines, and policymakers seeking to build public acceptance of autonomous aviation.
- AI policy
- Enterprise
Research
The perceived role of consciousness in moral status attributions
Rony Hirschhorn, Liad Mudrik, Niccolò Negro
PsyArXiv (OSF Preprints) · 2026-08-31
This large preregistered study (N=1,493) examined whether people treat consciousness as a necessary or sufficient condition for granting moral status to entities such as animals and AI. The findings show that while consciousness was strongly associated with moral status, participants both attributed moral status to non-conscious entities and denied it to conscious ones, meaning consciousness alone neither guarantees nor is required for moral consideration in public perception. Respondents also clustered into distinct decision-making profiles shaped by individual background characteristics, not just properties of the entity being judged. The authors argue these results have direct implications for how policy debates around animal welfare and AI regulation should be framed.
- AI policy