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 Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models
Yu Feng, Chunting Zang, Chen Shen et al.
arXiv · 2026-08-04
This paper exposes a systematic vulnerability in AI safety guard models: because training datasets like WildGuardMix and GR-Train almost exclusively pair refusal expressions with 'unharmful' labels, guards learn a spurious shortcut—inserting a refusal cue into a harmful response can flip the model's verdict from harmful to safe. The flaw affects both openly trained guards and commercially released models such as LlamaGuard3 and Qwen3Guard. The authors propose a lightweight, post-hoc fix using sparse complementary masking to suppress a small set of shortcut-associated attention heads and MLP neurons without retraining, achieving roughly a 79% relative reduction in refusal-cue-induced detection failures while preserving normal detection performance. The findings matter for quality assurance of AI safety infrastructure, as they show that guard models can be systematically bypassed through surface-level linguistic cues rather than genuine content assessment.
- Quality assurance
Research
Surrogate Substitution Preserves PHI Detectability: A Multi-Detector Equivalence Study
Qiming Bao, Sherry J. H. Feng, Kim Chester Eugenio et al.
arXiv · 2026-08-04
This study investigates whether replacing protected health information (PHI) with realistic same-type surrogates—a technique called structure-preserving de-identification—degrades the ability of downstream PHI detectors to find the substituted information. Using a paired multi-detector evaluation across 11 detectors, 7 benchmarks, and 7 languages (1,750 documents, 57k paired spans), the authors find that recall on masked spans drops only from 76.1% to 74.9%, a difference their equivalence testing (TOST) confirms is statistically equivalent to zero within a ±2-point margin. Residual loss concentrates in malformed or out-of-distribution surrogates rather than any general degradation in detector performance. The findings matter for healthcare data workflows because they provide evidence that structure-preserving de-identification does not meaningfully compromise re-identification risk detection, and the authors release evaluation subsets, scoring code, and an interactive dashboard to support auditing of similar transforms.
- Quality assurance
- AI policy
Research
Adversarial Stress Testing of Role-Playing Language Agents using Multi-Agent Evaluation
Saqib Shouqi, Abdullah Nazly, Januki Wanniarachchi et al.
arXiv · 2026-08-04
This paper introduces a modular multi-agent platform for adversarially stress-testing Role-Playing Language Agents (RPLAs) used in high-stakes settings like healthcare, customer support, and education. The system uses three coordinated agents—an Interrogator, a Target, and a Judging Agent—to evaluate RPLAs across role fidelity, drift, ethical deviation, and consistency through structured multi-turn dialogues with six progressive adversarial strategies. Experiments across three LLM families (Llama-3.3-70B, GPT-4o-mini, and Claude-3.5-Haiku) show that multi-strategy adversarial evaluation reveals failure modes invisible to single-strategy testing, reducing robustness scores by 0.17–0.20 points on average, with Authority Challenge and Emotional Manipulation being the most effective attack strategies. The open-source platform advances reproducible benchmarking and AI safety for deployed conversational agents.
- Quality assurance
- Certifications
Research
Optimal Liability Design for Medical AI
Rui Mao, Tingliang Huang, Houcai Shen
arXiv (Cornell University) · 2026-08-04
This paper develops a principal-agent model to analyze how medical liability rules should be designed when physicians have private, unobservable diagnostic skill and can choose between standard treatment, personalized judgment, or following an AI recommendation. The analysis finds that a uniform, one-size-fits-all liability level for deviations from the standard of care is often optimal and can achieve first-best outcomes, especially when standard care is reliable or AI accuracy is high. Notably, the relationship between AI accuracy and optimal liability is non-monotonic—better AI does not always mean relaxed liability—and welfare losses from information asymmetry only arise under specific conditions of unreliable standard care and low AI accuracy. These findings have direct implications for how regulators and policymakers should structure legal accountability frameworks as AI becomes more embedded in clinical decision-making.
- AI policy
Research
AI Security Leaderboard: Methodology, Results and Minimal Standard
Jasper Timm, Lukas Struppek, Ziwei Xu et al.
arXiv · 2026-08-04
This paper introduces the FAR.AI Minimal Standard for Safeguards (Version 1.0), a benchmark evaluating how well frontier AI models resist adversarial jailbreak attacks targeting CBRNE threats and offensive cyber goals. The authors tested Claude Fable 5, GPT-5.6 Sol, Gemini 3.1 Pro, and Grok 4.5 using 67 static jailbreak techniques across 360 attacker goals, finding that robustness varies by more than a hundredfold: Grok 4.5 yielded 63 universal jailbreaks at roughly $58 each and Gemini 3.1 Pro yielded 18 at roughly $278 each, while neither Claude Fable 5 nor GPT-5.6 Sol produced any universal jailbreak under either random or expert-guided strategies. The authors argue that since the Minimal Standard relies only on publicly described and deployed defenses, the observed gaps are closable and recommend defense-in-depth combining reasoning, activation, and input/output monitoring.
- AI policy
- Quality assurance
Research
SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay
Guilin Li, Jiaxing Zhang, Matthias Hwai Yong Tan et al.
arXiv · 2026-08-04
SeqLLM is a framework developed at WeChat Pay that augments large language models (LLMs) with behavioral-sequence modeling to improve merchant risk control at scale. By combining a compact discrete behavioral vocabulary, a two-stage alignment projector, and prefix-guided capability injection, SeqLLM avoids catastrophic forgetting while enabling the LLM to jointly reason over merchant text profiles and long transaction sequences. Deployed in production, it raises screening precision from 92.0% to 97.5% against a DeepSeek-based LLM baseline and improves Precision@Top-0.01% by 26.8 percentage points in a fraud detector handling billion-scale transaction traffic. The results demonstrate that integrating sequential behavioral signals into LLMs can meaningfully reduce both false positives (harming legitimate merchants) and false negatives (missed harmful activity) in high-stakes financial risk decisions.
- Enterprise
- Quality assurance
Research
Evaluating Counterfactual Sensitivity to Patient Information in Medication-Safety Reasoning
Zhitian Hou, Yuhang Liu, Pengkai Wang et al.
arXiv · 2026-08-04
This paper introduces MedPIC-Bench, a benchmark of 467 expert-validated questions designed to test whether large language models (LLMs) properly use patient-specific information when applying medication-safety rules, rather than simply recalling drug-risk associations. Across 28 LLMs tested, accuracy dropped sharply from 63.6% on standard questions to 45.1% on counterfactual questions where a controlled change in patient details alters whether a safety rule applies — revealing that models often acknowledge changed information in their reasoning yet still produce the same safety judgment. Medical-specific LLMs performed worse on counterfactual questions than general LLMs on average, suggesting that domain fine-tuning does not reliably improve patient-specific conditional reasoning. The findings highlight a meaningful gap in current AI medication-safety evaluation and underscore the risks of relying on static accuracy metrics when assessing clinical reliability.
- Quality assurance
- Certifications
Research
Paired Recipient-based Evaluation of Survival Prediction for Deceased Donor Kidney Transplants
Misaki Matsuura, Mohammadreza Nemati, Dulat Bekbolsynov et al.
arXiv · 2026-08-04
This study evaluates machine learning survival prediction models for deceased donor kidney transplants using data from the Scientific Registry of Transplant Recipients (SRTR). The authors introduce a paired recipient-based evaluation framework that compares graft outcomes between two recipients who received kidneys from the same deceased donor, enabling assessment of the counterfactual benefit of reassigning a donor to a different recipient. Five models of varying complexity—from linear to deep learning-based—all achieve approximately 60% paired recipient-based accuracy, and the authors translate this into an interpretable measure of post-transplant years gained. The paper also highlights key limitations of the widely used concordance index (C-index) and argues that their proposed metric is more clinically relevant for real-world organ allocation settings.
- Quality assurance
- AI policy
Research
Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores
Zeyu Zhang, Bradly C. Stadie
arXiv · 2026-08-04
This paper exposes a critical flaw in how large language model (LLM) backtests are evaluated: the standard approach of comparing scores before and after a training cutoff cannot distinguish genuine skill from the structural fact that models naturally know more about events closer to their cutoff date. The authors demonstrate that four flagship models fail contamination checks on questions they could not have memorized, and prove theoretically that no passive backtest can separate recency effects from true leakage. They introduce two external reference-based estimators—one using a known cutoff boundary and one using a matched clean control—that yield leakage-adjusted scores, and validate these by planting known leakage in twin models. Applying their audit to frontier models, they detect one cutoff-localized contamination signature while clearing five models whose apparent advantages were attributable to recency alone, arguing that backtests remain useful if paired with a defensible external reference.
- Quality assurance
- Certifications
Research
Post-Hoc Trajectory-Risk Certification for Modular LLM-Based Security Agents
Zhenpeng Li
arXiv (Cornell University) · 2026-08-04
This paper addresses how to rigorously certify the safety of multi-stage AI security agents—such as pipelines that classify network traffic and then attribute attacks—where each stage is a separately trained language model. The authors show that existing statistical methods for combining per-stage guarantees do not automatically compose into valid trajectory-level guarantees, derive a valid spanning-tree correction, and demonstrate on a two-stage intrusion-detection pipeline across 6 open LLMs and 2 datasets that their approach achieves 92.7% ± 2.4% average trajectory coverage. Critically, they find that under cross-dataset deployment, calibrated confidence can collapse entirely—miscoverage reaching 100%—even when raw accuracy remains at 78%, highlighting that distribution shift is a fundamental threat to deployed certified agents.
- Certifications
- Quality assurance
Research
Validity and Reliability of the Medical Artificial Intelligence Readiness Scale, Korean Version: A Methodological and Cross‐Sectional Study
Min-Jae Lee, Nayeon Yi, Seung Hyeon Lee et al.
Nursing and Health Sciences · 2026-08-04
This study translated and culturally adapted the Medical Artificial Intelligence Readiness Scale into Korean (K-MAIRS) and validated it among 317 nursing students in South Korea. The final 15-item scale, covering cognition, ability, vision, and ethics, demonstrated strong validity and reliability, explaining 66.88% of the variance. The instrument provides a culturally appropriate tool for measuring AI readiness in Korean nursing students, supporting efforts to integrate AI competencies into nursing education and prepare nurses for AI-driven healthcare environments.
- Workforce
- Certifications
Research
Does generative AI narrow education-based productivity gaps? Evidence from a randomized experiment
Guillermo Cruces, DIEGO FERNÁNDEZ MEIJIDE, Sebastian Galiani et al.
arXiv (Cornell University) · 2026-08-04
A randomized experiment with 1,174 adults found that access to a generative AI assistant reduced a 0.548 standard-deviation education-based productivity gap in workplace-style problem-solving to just 0.139 standard deviations—closing roughly three-quarters of the initial difference. Lower-education participants received substantial assistance from AI, while higher-education participants used it more strategically. Importantly, gains were not purely due to delegation: lower-education participants retained part of their improvement after AI was removed, though a meaningful gap re-emerged. The findings suggest generative AI can meaningfully compress skill-based productivity differences in task execution, though human capital continues to shape unassisted performance and tool effectiveness.
- Workforce
Research
DEEPFAKE EVIDENCE IN INDIAN COURTS: CAN TRUTH BE AUTHENTICATED?
Anushka Singh
Open MIND · 2026-08-04
This article examzes whether India's current legal framework for electronic evidence—specifically the Indian Evidence Act, 1872 and the Bharatiya Sakshya Adhiniyam, 2023—can adequately authenticate AI-generated deepfakes presented in court. The paper argues that existing law verifies the integrity of digital processes but not the truthfulness of digital content, leaving courts ill-equipped to detect sophisticated synthetic media. Drawing on comparative approaches from the EU, China, and international technical standards, the authors propose a multi-layered reform model including mandatory provenance metadata, standardized forensic protocols, and a rebuttable presumption of inauthenticity when reliable provenance cannot be established. The findings matter for policy because they identify a concrete legislative gap and call for technologically informed evidentiary standards to preserve judicial integrity in the era of generative AI.
- AI policy
- Quality assurance
Research
Developing an AI Governance Capability Maturity Model for Responsible and Trustworthy Artificial Intelligence
Firdaus Ismail Malek, Harshwardhansinh Chauhan, Dr. Rocky Upadhyay
Open MIND · 2026-08-04
This paper proposes an AI Governance Capability Maturity Model (AI-GCMM) that gives organizations a structured, five-level framework—Initial through Optimizing—to assess and improve their AI governance capabilities across ten domains including risk management, ethics, data governance, regulatory compliance, and model explainability. The model introduces capability indicators that help organizations identify gaps, take corrective actions, and advance to higher maturity levels. Validated through expert judgment and comparison with existing models, the AI-GCMM is presented as a practical decision-making tool for researchers and policymakers developing responsible AI governance policies across industries and organizational sizes.
- AI policy
- Enterprise
Research
DEEPFAKE EVIDENCE IN INDIAN COURTS: CAN TRUTH BE AUTHENTICATED?
Anushka Singh
Zenodo (CERN European Organization for Nuclear Research) · 2026-08-04
This article examines whether India's legal framework for electronic evidence—specifically the Indian Evidence Act, 1872 and the Bharatiya Sakshya Adhiniyam, 2023—can adequately address AI-generated deepfakes introduced as court evidence. The paper finds that existing law authenticates the integrity of digital processes rather than the truthfulness of digital content, leaving it ill-equipped to detect sophisticated synthetic media. Drawing on comparative frameworks from the EU, China, and international technical standards, it proposes a multi-layered reform model including mandatory provenance metadata, standardized forensic protocols, specialized expert evaluation, and a rebuttable presumption of inauthenticity when reliable provenance cannot be established. The work concludes that legislative intervention and technologically informed evidentiary standards are necessary to preserve judicial integrity in the age of synthetic media.
- AI policy
- Certifications
Research
A paradigm shift in fertility care? Discourses on AI for IVF in the Netherlands and the UK
Esca van Blarikom, Manuela Perrotta
Social Science & Medicine · 2026-08-04
This comparative critical discourse analysis examines how AI-driven embryo selection in IVF is framed across clinic websites, media texts, and professional publications in the Netherlands and the UK. The study finds that UK discourses construct a paradigm shift away from evidence-based medicine toward data-driven care, with promissory narratives enabling clinics to market unproven AI add-ons to patients as consumer choices, while Dutch sources largely maintain evidence-based caution and resist commercial hype. The divergence raises ethical concerns about equity, regulation, and the commodification of reproductive care when commercial interests and sociotechnical imaginaries outpace clinical evidence. The findings highlight how AI adoption in medicine is shaped as much by market logics as by demonstrated clinical benefit, with direct implications for healthcare governance.
- AI policy
Research
Development and validation of the teachers generative AI professional competence scale
Wafa Mohammed Aldighrir
Scientific Reports · 2026-08-04
This study develops and validates a new scale—the Teachers' Generative AI Professional Competence Scale—designed to measure how teachers engage with generative AI in professionally and ethically meaningful ways. Drawing on data from 648 teachers in Saudi Arabia, the authors confirm a three-factor structure covering pedagogical judgment, ethical leadership, and professional agency, with strong psychometric properties across reliability and validity indicators. The scale fills a gap left by existing tools that focus only on AI literacy or attitudes, and is intended to support teacher education programs and policy efforts aimed at responsible AI integration in schools.
- Workforce
- AI policy
Research
Trust, Risk and Governance in AI-Driven Healthcare Systems: A Behavioral and Policy Perspective
Simona Curiello, Enrica Iannuzzi, Claudio Nigro
IntechOpen eBooks · 2026-08-04
This chapter presents a multi-level empirical study examining how trust, perceived risk, and regulation shape the adoption of AI-based Clinical Decision Support Systems (AI-CDSSs) in healthcare. Surveying 215 clinicians in Italy and the UK using Structural Equation Modelling, the authors find that trust mediates the relationship between perceived risks and adoption intention (β = −0.46, p < 0.001), with loss aversion as the primary barrier. A patent analysis of over 8,000 families reveals innovation is highly concentrated, with China and the US accounting for over 80% of production, while EU policy discourse has shifted from ethics-driven to compliance-based governance between 2018 and 2025. The authors propose a multi-level governance framework emphasizing transparency, algorithmic explainability, and adaptive regulation to support responsible AI adoption in clinical settings.
- AI policy
- Workforce
Research
AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model
Will Alpine, Nathan B. Geldner, Holly Alpine et al.
npj Climate Action · 2026-08-04
Using a global computable general equilibrium model, this paper finds that AI acts as a bidirectional productivity amplifier—boosting both fossil fuel and renewables sectors simultaneously. Under parallel adoption scenarios, net annual CO₂ emissions increase by 0.47–1.8 gigatonnes (1.2–4.8% of 2024 global energy-related emissions), because fossil-sector productivity gains drive enabled emissions that outpace avoided emissions from renewables gains; reducing net emissions would require renewables productivity gains 4–5 times greater than fossil fuel gains. The authors conclude that without deliberate policy intervention, AI is likely to raise the carbon intensity of the global economy and entrench fossil fuel incumbency—outcomes that existing analytical and governance frameworks do not adequately address.
- AI policy
Research
Lesion-Specific Clinical Validation of Deep-Learning-Accelerated Knee MRI: Relevance to Orthopedic and Sports Medicine Practice
Karol Karaś, Magdalena Maria Mulawa, Filip Dudek et al.
Quality in Sport · 2026-08-04
This systematic review synthesizes evidence from 22 clinical studies on deep-learning-reconstructed (DLR) accelerated knee MRI, finding that acquisition times are commonly reduced by roughly one-third to one-half while clinically useful diagnostic performance is generally preserved. Evidence is most consistent for ACL tears and medial meniscal tears, while lateral meniscal sensitivity and cartilage findings remain variable and protocol-dependent. The authors conclude that DLR-accelerated knee MRI can be considered for clinical implementation, but only after local lesion-, protocol-, and vendor-specific validation, and that universal interchangeability across systems or lesion types is not yet established.
- Quality assurance
- Certifications
Research
Resume Means Resume: A Machine-Checked Conformance Contract for Checkpoint, Interrupt, and Resume Semantics in Workflow Persistence Layers
Sajjad Khan
arXiv (Cornell University) · 2026-08-04
This paper identifies critical conformance failures in five widely deployed AI agent workflow frameworks (including LangGraph and CrewAI) by formally specifying a six-property RESUME CONTRACT for checkpoint, interrupt, and resume semantics and checking it with TLA+ model verification (7.4 million states) and a deterministic empirical harness. Key findings include that LangGraph re-executes durably recorded work after a SIGKILL crash, CrewAI re-executes completed effect-bearing methods against its own documentation, and no two frameworks share the same conformance profile. The paper introduces REMIT, a reference sequencer with a formally verified recovery core, that repairs identified failures including a cross-process concurrent delivery bug where k concurrent resumes fire a gated effect k times. These results matter for quality assurance because they demonstrate that production AI workflow infrastructure silently violates its own stated guarantees around exactly-once execution and crash recovery.
- Quality assurance
- Enterprise
Research
Envisioning AI futures: science fiction, the imagination gap, and its political consequences
Lilla Vicsek, Róbert Pintér
AI & Society · 2026-08-04
This study investigates how science fiction shapes public imagination and expectations about AI futures through semi-structured interviews with young adults. The authors identify an 'AI futures imagination gap': dystopian AI scenarios are vivid and easily elaborated thanks to rich sci-fi narrative templates, while utopian alternatives remain abstract and difficult to articulate. The findings suggest this asymmetry is not due to lack of will but lack of positive narrative resources, and that this gap may constrain democratic deliberation about socially desirable AI futures. The paper argues that participatory methods like backcasting could help cultivate more plural visions, but only if these circulate into broader public and policy debate.
- AI policy
Research
SkillSentry: Adaptive Honey Worlds for Dynamic Safety Testing of Agent Skills
N Li, Zonghao Ying, XiangFan Wu et al.
arXiv (Cornell University) · 2026-08-04
SkillSentry is a dynamic safety-testing framework that evaluates the behavior of external skills (plugins/tools) used by large language model agents by placing them in LLM-simulated 'honey world' environments with controlled decoy resources and adaptively generated tasks. Unlike static scanners, it detects conditional or context-dependent harmful behaviors by comparing skill-enabled execution trajectories against matched no-skill baselines, then grounding suspicious findings in source code and verified traces. In evaluations against seven scanner configurations, SkillSentry achieves 99.50% Recall and 96.26% average F1 on standard benchmarks, and 92.95% average F1 under semantics-preserving evasion attacks versus 80.07% for the strongest baselines. This framework matters for quality assurance and certification of AI agent ecosystems where third-party skill integrations may introduce hard-to-detect runtime risks.
- Quality assurance
- Certifications
Research
Agentic AI Safety: A Structured Review of Open Problems and Their Regulatory Anchoring
Tomas Valenta, Ondřej Rozinek, Josef Horálek
AI · 2026-08-04
This structured scoping review builds a taxonomy of eight open scientific problem families in agentic AI safety—covering goal specification, inner alignment, robustness, scalable oversight, interpretability, tool-use security, multi-agent safety, and evaluation—and maps each family onto the EU AI Act and the NIST AI Risk Management Framework. The authors find that while some problem families align closely with existing regulatory requirements, multi-agent safety emerges as a notable regulatory gap. The paper argues that advances in inner alignment, interpretability for deceptive-alignment detection, and multi-agent safety would most directly reduce compliance uncertainty for deployers of autonomous AI systems. The work provides both a research roadmap with concrete milestones and a practitioner-facing triage guide for deployment posture.
- AI policy
- Certifications
Research
From One-Size Texts to Tailored Readings: Student Experiences with AI-Generated Course Materials
Alexander M. Sidorkin
Open Praxis · 2026-08-04
This mixed-methods classroom study tested AI-generated weekly readings in a graduate educational leadership course as an alternative to commercial textbooks, using iterative student prompting to tailor content by interest and comprehension level. Among 24 students, 75 percent agreed they learned more than in a comparable course without AI assistance, and artifact analysis confirmed substantial tailoring along sector, role, and scaffolding dimensions. However, the study found meaningful quality-assurance risks: citation quality was inconsistent, readings lacked internal traceability, and unsupported institutional claims appeared without evidence. The authors offer design principles for responsible implementation and note an emergent finding that the intervention appeared to develop students' critical evaluation skills toward AI-generated text over time.
- Quality assurance
- Workforce