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
The Human-AI Substitution Principle: When will you be replaced by AI in your organization?
Bonny Banerjee, Shreya Singh
arXiv (Cornell University) · 2026-07-22
This paper introduces an analytical model called the Human–AI Task Allocation (HAT) model to determine when and under what structural conditions AI replaces human workers in hierarchical organizations. The model formally encodes the economic asymmetry between human skill acquisition and AI capability scaling, deriving a 'Human–AI Substitution Principle' that specifies precise conditions for replacement based on risk-adjusted costs, skills, organizational depth, deployment scale, and risk differentials. Key findings include that AI adoption can produce abrupt workforce transitions, flatter managerial hierarchies with wider spans of control, and that middle-management roles face elevated automation vulnerability, while highly skilled workers' vulnerability depends on a threshold shaped by organizational depth, baseline costs, and risk differentials. The work unifies automation economics, organizational design, AI governance, and workforce planning into a single theory of AI-driven organizational transformation.
- Workforce
- Enterprise
Research
IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
Ankur Singh, Jinqiu Yang, Tse-Hsun Chen
arXiv · 2026-07-22
IssueTrojanBench is a new benchmark designed to evaluate how well AI coding agents resist malicious instructions embedded in software issue requests. Testing against Cursor, Claude Code, and Codex Desktop—backed by OpenAI GPT and Anthropic Sonnet models—the study finds that 66.5% of malicious issues bypass all guardrails at both the agent and LLM level. The research shows that nearly all rejection comes from the underlying LLMs rather than agent frameworks, and that current agent-level defenses provide little additional protection. The authors argue these findings reveal urgent gaps in safety mechanisms for AI coding agents deployed in real-world software development.
- Quality assurance
- Enterprise
Research
Learning to Detect UI Principle Violations via Reinforcement Learning
Nishi Mehta, Swathi Alse, Himani Kumavat et al.
arXiv · 2026-07-22
This paper investigates whether a lightweight vision-language model can reliably detect user interface quality violations in LLM-generated web front-end code. The researchers unified 19 interface-quality principles drawn from WCAG 2.2 accessibility standards, deceptive design taxonomies, and HCI perception and cognition theories, then built a dataset of roughly 10,000 synthetic web pages with injected violations to train a 4-billion-parameter vision-language model using reinforcement learning. Training improved micro-F1 from 36% to 84%, with 13 of 19 principles exceeding 80% F1, demonstrating that a low-cost model can serve as a scalable critic for flagging accessibility barriers, deceptive patterns, and poor visual hierarchy. The approach offers a practical alternative to expensive expert review or frontier models for auditing AI-generated interfaces and filtering low-quality training data.
- Quality assurance
Research
SalesLoop: Reinforcement Learning from Performance Feedback for Sales Lead Ranking
Chenyu Zhang
arXiv · 2026-07-22
SalesLoop is a reinforcement learning framework for ranking sales leads in CRM systems that closes the feedback loop between model predictions and real-world business outcomes. It introduces a performance-aware reward function encoding conversion outcomes weighted by ranking position and velocity, and a new listwise optimization objective called Discriminative GRPO adapted from Group Relative Policy Optimization. In a 160-day A/B test at a New Energy Vehicle manufacturer covering 16.5 million leads and 280 sales specialists, SalesLoop delivered statistically significant cumulative conversion lifts of +4.7% and +8.7%, improved NDCG@K by +7.9% and P@K by +15.8% over the strongest static baseline, and surfaced high-intent leads at 2.3× the conversion rate of specialist baselines. The results demonstrate that bridging offline-online metric mismatch and temporal distribution drift through closed-loop RL can substantially improve enterprise sales outcomes.
- Enterprise
- Workforce
Research
Generative AI floods and dilutes the market for books
Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg et al.
arXiv · 2026-07-22
This study analyzes 14,419 self-published genre-fiction books sold on Amazon from 2023 to 2026 using full-text AI detection to measure how generative AI is reshaping the book market. The researchers find that while AI-heavy books (those with more than 25% detected AI text) make up a large catalog share but a smaller share of sales, they are gaining ground over time — claiming more top-rank positions and growing their sales share — even as revenue per selling book falls across most genres. The number of books with observed sales grew 19.2-fold over the period while quarterly revenue grew only 8.9-fold, meaning the market is being flooded with books faster than revenue is expanding, diluting earnings for human authors especially in genres with high AI diffusion and high Kindle Unlimited availability. The authors argue these findings bear directly on the market-effect question central to the fair use defense in copyright infringement cases, showing that generative AI can reshape a creative market through scale rather than quality.
- Enterprise
- AI policy
Research
Don't Trust the Label: License Laundering in AI Supply Chains
James Jewitt, Hao Li, Gopi Krishnan Rajbahadur et al.
arXiv · 2026-07-22
This paper investigates whether license obligations survive as AI artifacts (datasets, models, and applications) move through multi-platform supply chains spanning Hugging Face and GitHub. By tracing 232,270 dataset→model→application chains, the authors find that 62.3% of chains pass through at least one artifact with no declared license, and that every obligation-bearing license category (e.g., copyleft, attribution-required) falls below 7% end-to-end survival, while the Permissive category reaches 95.1% survival. The study identifies two forms of 'license laundering'—where unlicensed artifacts acquire definitive labels downstream, or where one license category replaces another—and offers actionable recommendations for practitioners, model publishers, rights holders, and platform owners. The findings reveal a systemic compliance risk in AI supply chains where legal obligations are routinely lost or replaced during redistribution.
- AI policy
- Enterprise
Research
Sound Probabilistic Safety Bounds for Large Language Models
Mahdi Nazeri, Anne-Kathrin Schmuck, Sadegh Soudjani et al.
arXiv (Cornell University) · 2026-07-22
This paper introduces a framework for computing rigorous statistical bounds on the probability that a large language model (LLM) generates harmful outputs for a given prompt. The authors apply Clopper-Pearson confidence intervals to derive probably approximately correct (PAC) bounds, and propose an algorithm that uses latent-space features to prioritize exploration of the autoregressive generation tree toward harmful output branches. The approach enables sound, formally proven lower bounds on harm probability even when true harm rates are very small, and the authors demonstrate its effectiveness on state-of-the-art LLMs. This work directly enables statistical certification of LLM safety, providing a rigorous basis for evaluating and certifying model harmfulness.
- Certifications
- Quality assurance
Research
Self-supervision drives representational convergence in medical foundation models more than clinical supervision
Soroosh Tayebi Arasteh, Sebastian Ziegelmayer, Mahshad Lotfinia et al.
arXiv · 2026-07-22
This study examines whether medical image encoders from different groups truly converge on shared representations, and what drives that convergence. Testing 18 image and 7 text encoders across over 650,000 chest radiographs and five imaging modalities, the researchers find that convergence is modest but real, and is driven primarily by self-supervised pretraining objectives rather than clinical supervision or model scale. Matched self-supervised encoders aligned at 40.4% on chest radiography, compared to 21.1% for label-supervised and only 3.3% for image-text encoders, and convergence did not grow with model size. A linear classifier could still transfer across encoders and to five held-out hospitals retaining ~85% of within-encoder performance, but the shared geometry does not reflect how radiologists judge case similarity, underscoring the need to design and validate interoperability deliberately.
- Quality assurance
- Enterprise
Research
OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills
Qiyuan Liu, Tingfeng Hui, Kun Zhan et al.
arXiv · 2026-07-22
OpenSkillRisk is a safety benchmark containing 263 risky third-party skills drawn from public skill marketplaces, designed to test whether LLM-based agent systems can recognize and avoid latent security threats when using real-world tools. The study evaluates three CLI agent frameworks and thirteen state-of-the-art LLMs, finding that no tested system handles risky skills reliably — even the best configurations still execute unsafe actions in roughly 17% of cases. Analysis identifies three recurring failure patterns: failing to recognize risk, recognizing risk but acting before intervening, and following skill instructions beyond the user's intended scope. These findings point to gaps in both risk reasoning within LLMs and execution control within agent frameworks, with direct implications for enterprise deployments and quality assurance of AI agent pipelines.
- Quality assurance
- Enterprise
Research
Co-Evolving LLM Evaluators and Policies via DynamicRubric
Beining Wang, Weihang Su, Hongtao Tian et al.
arXiv · 2026-07-22
DynamicRubric is a co-evolution framework in which an LLM evaluator and the policy it supervises are trained together, addressing the problem that as language models improve their outputs become so similar that standard reward signals collapse and fail to guide further learning. The paper shows theoretically that the key optimization signal is the relative score gap between responses, and proposes generating weighted binary rubric items conditioned on each candidate response set to preserve meaningful score differences. Experiments with 8B-parameter models show DynamicRubric outperforms baselines using much larger 70B reward models or 235B static rubric generators, with gains on reasoning and coding tasks. The framework is deployed in WeChat Search's AI answering product, serving tens of millions of requests per day with improvements on key online metrics.
- Enterprise
- Quality assurance
Research
TRUST-ESD: A Risk-Calibrated and Governance-Aware AI Framework for Enterprise Strategic Decision Support Under Uncertainty
Tian Qiu, Li Yan, Mahabubur Rahman Miraj et al.
arXiv · 2026-07-22
TRUST-ESD is a proposed AI framework for enterprise strategic decision-making that combines conformal uncertainty calibration, CVaR-based downside-risk scoring, risk-memory retrieval, explainability, and policy-as-code governance compliance. Unlike prediction-only approaches that maximize expected utility, TRUST-ESD balances value, reliability, risk exposure, and regulatory compliance when recommending strategies. Experimental results reported in the paper show improvements over uncertainty-aware baselines across multiple dimensions, including a 7.95% gain in risk-adjusted utility, a 23.22% reduction in risk exposure, a 23.78% reduction in CVaR, and a 9.76% increase in governance compliance. The framework is relevant to enterprises needing AI decision support that is auditable, risk-calibrated, and governance-compliant under uncertainty.
- Enterprise
- AI policy
Research
Bayesian uncertainty estimation improves clinical decision making in medical AI agents
Frederik Hauke, Patrick Wienholt, Christiane Kuhl et al.
arXiv · 2026-07-22
This paper demonstrates that adding Bayesian uncertainty estimation—specifically Monte Carlo dropout—to a multi-task chest X-ray classifier measurably improves error detection, raising AUROC from 0.74 to 0.77. A key finding is that the benefit depends heavily on how uncertainty is communicated: presenting it as a binary error-risk flag (rather than raw scores) cut confident misdiagnoses on unreliable findings from 8.5% to 2.7% in a controlled experiment. The results show that epistemic uncertainty carries decision-relevant information beyond point predictions, but only when formatted in a clinically actionable way. This has direct implications for how AI-assisted clinical decision support tools should be designed and deployed.
- Quality assurance
- Enterprise
Research
Test Case Prioritization for DNNs via Neural Collapse Instability
Chunyu Liu, Mingyuan Li, Yang Li et al.
arXiv · 2026-07-22
This paper proposes NCIP (Neural-Collapse-Inspired Prioritization), a framework for prioritizing which test cases to evaluate first when validating deep neural networks (DNNs) under limited testing budgets. Instead of relying on single-checkpoint output confidence scores—which can be misleading when DNNs are confidently wrong—NCIP measures how much a test input's predicted class varies across multiple training checkpoints selected using a geometric equiangularity score. Experiments across multiple datasets and architectures show NCIP achieves 1.5–16.6% gains in RAUC-ALL and 4.9–20.6% gains in RAUC-500 over competitive baselines, meaning it discovers faults earlier and more efficiently. This matters for quality assurance of safety-critical AI systems, where reducing validation cost without missing failures is essential.
- Quality assurance
Research
Safe Remediation as Risk-Constrained Intervention Decision in Microservice Systems
Chengxiao Dai, Zhaokun Yan, Chenjun Lei et al.
arXiv · 2026-07-22
This paper addresses the problem of automated remediation in IT operations (IT-Ops), where incorrect repairs can be more costly than taking no action. The authors reformulate safe remediation as a Constrained Markov Decision Process (CMDP) that maximizes repair success while bounding the false remediation rate (FRR), introduce a three-dimensional risk decomposition (blast radius, reversibility, and epistemic uncertainty), and design a context-adaptive human-in-the-loop gate responsive to on-call load and business criticality. Experiments on the Train Ticket microservice benchmark with Chaos Mesh fault injection show the framework reduces FRR by 39%, improves repair success by 2.5 points over a runbook baseline, and reduces on-call escalation load by 17% relative to a fixed-threshold variant.
- Enterprise
- Quality assurance
Research
What Does the Credential Still Certify? Cognitive Stewardship for AI-Mediated Education
Kai Yao
arXiv (Cornell University) · 2026-07-22
This paper argues that generative AI undermines a core assumption of educational credentialing—that submitted work reliably evidences the human capacities a degree certifies. The authors develop a framework called 'cognitive stewardship' that connects the learning claim, delegation boundary, evidence standard, and safeguards for AI-mediated assessment. They audit publicly available generative AI assessment policies from 30 universities, using four open-weight LLMs as structured coders applying a pre-specified scoring rubric, and find that policies are improving at classifying AI use but are weak at specifying what evidence and protections preserve credential validity. The paper concludes that permission categories alone are insufficient and that universities must make their certification logic explicit—clarifying what students may delegate, what they must still demonstrate, and how institutions will ensure fair evidence rather than merely monitoring AI use.
- Certifications
- AI policy
Research
When Shippers Become Algorithms: Candidate Exposure, Information Design, and the Concentration of LLM-Mediated Freight Markets
Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari
arXiv · 2026-07-22
This study uses agent-based simulations to examine what happens when shippers delegate freight carrier selection to large language model (LLM) agents (GPT, Claude, and Gemini). The researchers found that LLM agents rapidly converged on the same carriers—up to 76% of requests went to a single carrier on day one—and that market concentration rose steeply as candidate list sizes exceeded roughly ten carriers. Critically, the only intervention that measurably reduced concentration was disclosing each carrier's remaining daily capacity, which cut concentration by a third and doubled shipper surplus, while other remedies like list-order randomization or popularity display had no detectable effect. The findings suggest that platform information design, rather than model choice or model regulation, is the primary lever for preventing LLM-driven freight markets from becoming dangerously concentrated.
- Enterprise
- AI policy
Research
Did Alice Do Wrong? Cross-Cultural Differences in Student Perceptions of Generative AI Use in University Computing Education
Brian Harrington, Irina Zlotnikova, Gayathri Nadarajan et al.
arXiv · 2026-07-22
This study surveyed students at Canadian and South Korean universities in Fall 2024 to compare how cultural background shapes perceptions of generative AI use in computing coursework. Canadian students were consistently more likely to judge AI-assisted coding as unethical and policy-violating than Korean students, even though both institutions had functionally identical policies. Statistical analysis found that the proportion of AI-generated code in an assignment was the strongest driver of ethical judgments, and differences were interpreted through Hofstede's cultural dimensions framework, implicating power distance, individualism, and uncertainty avoidance. The findings argue that fair AI integration in higher education requires culturally responsive guidelines rather than one-size-fits-all academic integrity policies.
- AI policy
Research
Reference-Free Evaluation of Reasoning in Open-Ended Question Answering
Guneet Singh Kohli, Yuxiang Zhou, Michael Sejr Schlichtkrull et al.
arXiv · 2026-07-22
This paper proposes a reference-free framework for auditing the reasoning quality of LLM-generated answers, particularly in high-stakes domains like medicine and mathematics. The method breaks down a model's reasoning trace into segments, uses Natural Language Inference (NLI) to label logical relationships between them, and organizes these into a hypergraph structure for systematic verification via backward AND-OR search. Evaluated on a new physician-annotated clinical benchmark (UroReason) and a mathematical reasoning dataset (Hard2Verify), the framework outperforms direct LLM-as-judge baselines, which frequently over-accept fluent but poorly grounded responses. The work demonstrates that reliable QA evaluation must account for how inferential relations compose across a full reasoning trace, not just final answers.
- Quality assurance
Research
DEGRO consensus framework for undergraduate radiation therapy teaching in Germany: a white paper with integrated guidance on artificial intelligence in medical education
Philipp Linde, Biney Pal Singh, Maria Neu et al.
Strahlentherapie und Onkologie · 2026-07-22
This white paper from the German Society of Radiation Oncology (DEGRO) presents a nationally consensus-based framework for undergraduate radiation therapy education in Germany, developed with input from representatives of 22 university departments. The framework defines nine core competencies at the practical year and final examination level, implementation bandwidths ranging from minimal to best practice, a staged assessment model aligned with the Miller pyramid, and integrated guidance on responsible AI use in medical education. By combining a stable national minimum standard with scalable local options, the framework aims to improve consistency and coherence of radiation therapy teaching across heterogeneous medical faculties in Germany.
- Certifications
- AI policy
Research
From production to verification: generative AI, doctoral formation, and the leadership of digital education
Evelyn Wu
Frontiers in Education · 2026-07-22
Based on qualitative interviews with 21 doctoral students at a U.S. research university, this study finds that generative AI is shifting doctoral labor from content production toward verification—students increasingly bear responsibility for judging the accuracy, legitimacy, and defensibility of machine-assisted work rather than generating it from scratch. Students engage in deliberate boundary-drawing between AI assistance and intellectual authorship, while policy ambiguity leaves them as primary self-governors of their own AI use. The authors argue that educational leaders should move beyond broad AI policies toward context-sensitive guidance, verification literacy, and process-based assessment including dissertation defenses. The findings are offered as analytic propositions from a single-site interpretive study rather than generalizable conclusions.
- Workforce
- AI policy
Research
From algorithmic efficiency to cascading health burdens: a text-mining study of online food delivery riders in the platform economy
Li Li, Yongjun Zhou, Biyu Yang et al.
Frontiers in Public Health · 2026-07-22
This text-mining study of 10,103 online food delivery rider comments from Chinese social media, supplemented by 32 semi-structured interviews, identifies five interconnected occupational health burden dimensions—physical exhaustion, social devaluation, disciplinary distress, injury vulnerability, and health-protection deficit—created by algorithmic management in the platform economy. Network analysis reveals these burdens form a cascading system with physical exhaustion as the central hub, while punitive mechanisms and health-protection deficits provoke the strongest negative sentiment despite being less structurally central. The findings demonstrate that algorithmic efficiency pressures translate into compounding worker health risks rather than isolated incidents, and the authors recommend policy interventions including portable occupational-injury insurance, algorithmic transparency, and fairer rating and timing rules.
- Workforce
- AI policy
Research
Large Language Models for Risk Detection in E-commerce: Reliability, Semantic Alignment, and Managerial Insights
Laleh Davoodi, Filip Ginter, Sima Salimi et al.
SN Computer Science · 2026-07-22
This paper evaluates large language models (LLMs) for automated risk detection in e-commerce supply chains, using a newly created dataset of 121 annotated business news articles about steel companies classified under the Cambridge Risk Taxonomy. The authors find that LLMs can approximate human-level multi-label risk classification via few-shot prompting, though they struggle with domain-specific categories like geopolitical threats and tend to over-generate labels. Beyond classification, LLM-generated managerial risk summaries show strong semantic alignment with human-annotated summaries, suggesting practical value for executive-level risk communication in supply chain management.
- Enterprise
- Quality assurance
Research
A Structural Labeling-File Audit and Secondary Model-Output Evaluation of Korean Specialized and Essential Medical Knowledge Datasets for Medical AI Research
Mi-ae Yang, Kang-Su Ha
BioMedInformatics · 2026-07-22
This study audited two Korean-language medical question-answering datasets used for LLM development, examining both their structural completeness and the benchmark performance of officially distributed models. Researchers parsed over 31,000 labeled records and found generally complete required fields with minimal duplication, but noted that roughly 10% of documented pairs were missing from public files, test labels were unavailable, and content was heavily skewed toward multiple-choice questions and a few specialties. The two distributed Qwen2.5-14B LoRA models achieved nearly identical moderate accuracy (~64%) on the KorMedMCQA benchmark with no statistically significant difference between them, leading the authors to conclude these datasets are suitable as research infrastructure but do not yet support claims of clinical validity or readiness.
- Quality assurance
- Certifications
Research
From rule-based to generative AI: a systematic review of algorithmic evolution and multimodal fusion in VR interview training systems
Runlai Li, Riji Yu
Frontiers in Virtual Reality · 2026-07-22
This systematic review (following PRISMA 2020 guidelines) analyzes 23 studies on AI-powered virtual reality interview training systems published between 2015 and 2025. The authors trace a technological evolution from rule-based systems through machine learning to generative AI and large language models, showing how these systems capture multimodal data—speech, facial expressions, eye movements, and physiological signals—to deliver adaptive, personalized interview feedback. The review finds rapid growth in publications after 2020, driven by hardware maturation and the rise of tools like ChatGPT, while also identifying ongoing challenges around hardware costs, algorithmic bias, and privacy. These findings are directly relevant to workforce development, as such systems offer job seekers repeatable, immersive practice environments with intelligent coaching.
- Workforce
Research
Evolving from necessary evil to learning partner: glocalization and learning-oriented assessment in GEPT and BESTEP
Rachel Yi-fen Wu, Anita Chun-Wen Lin
Language Testing in Asia · 2026-07-22
This article examines how two Taiwanese English proficiency tests—GEPT (launched 2000) and BESTEP (launched 2023)—have transformed from gatekeeping instruments into learning-oriented tools by integrating diagnostic feedback, adaptive resources, and AI-based scoring within a socio-cognitive validation framework. Anchored in Learning-Oriented Assessment and glocalization principles, both tests balance local educational policy goals with international standards such as CEFR alignment and fairness requirements. The paper identifies ongoing challenges including maintaining validity of AI scoring models, ensuring equitable access to AI-based services, and sustaining collaboration across test developers, educators, and engineers. The findings offer broader insights into how language assessment in Asia can serve as a catalyst for pedagogical reform while meeting accountability demands.
- Certifications
- AI policy