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
Convergence of Digital Transformation in Accounting and Auditing Research and Training: Preliminary Survey and Solutions
V Nguyen
American Journal of Economics and Business Innovation · 2026-09-12
This paper surveys the integration of AI and digital tools into accounting and auditing education in Vietnam, using a Technology Acceptance Model (TAM)-based questionnaire analyzed with SPSS 20.0. Despite participants recognizing AI's potential to improve teaching quality and student employability, actual adoption rates were very low—only 8% for AI and 28% for data analytics—with key barriers being lack of training (80%), insufficient technical support (72%), and inadequate infrastructure (64%). The authors propose solutions including curriculum redesign, faculty development, and the establishment of legal and policy frameworks to accelerate digital transformation in the sector. The findings underscore the need to close digital skill gaps to prepare graduates for labor market demands and support broader economic development.
- Workforce
- AI policy
Research
The New Disease and the Machine Room: Keynes's 1930 Essay, Artificial Intelligence, and the Data-Center Buildout
Doug Doucette
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-12
Doucette rereads Keynes's 1930 essay on technological unemployment as a framework for analyzing AI's early labor-market effects and the U.S. data-center buildout of the mid-2020s. The paper finds that AI-related labor disruption is concentrated in junior cognitive work and entry-level hiring rather than aggregate unemployment, while data centers generate many construction jobs but few permanent ones, creating local conflicts over water, power, and public consent. Drawing on Keynes's distinction between short-run maladjustment and long-run abundance, the paper proposes three policy tests: keeping entry-level career paths open, assigning infrastructure costs to large-load operators, and tracking whether productivity gains reach typical workers' pay. The conclusion argues that failure to meet these tests would be institutional rather than technical.
- Workforce
- AI policy
Research
Testing the Kill Switch: A Conformance-Based Approach to Agentic AI Containment Assurance
Naveen Sundaresan
Zenodo (CERN European Organization for Nuclear Research) · 2026-09-12
This paper proposes a conformance-based audit framework for verifying that 'kill switch' or stop mechanisms in agentic AI systems actually function as required under real operating conditions. Existing regulatory frameworks—including the EU AI Act, NIST AI RMF, ISO/IEC 42001, and Singapore's MAS and IMDA guidelines—mandate human oversight and intervention capability but provide no method for independent auditors to verify these controls work. The authors define a Target of Evaluation, five testable control families (trigger recognition, authority, cessation, latency, and failure resilience), machine-readable control representations, and evidence requirements that distinguish measured system behavior from procedural self-attestation. The approach aims to convert descriptive containment guidance into auditable, independently verifiable criteria using a vendor-neutral assurance harness with pluggable adapters.
- Certifications
- AI policy
- Quality assurance
Research
Reconceptualizing Age Assurance as a Sociotechnical Problem: Connecting Evidence, Evaluation, Claims, and Decisions
Renkai Ma, Prakriti Dumaru, Thomas Synaepa-Addison et al.
arXiv (Cornell University) · 2026-09-11
This paper argues that age assurance—systems designed to verify or estimate whether a user is a child—should be understood as a sociotechnical process rather than a purely technical problem. Reviewing 85 publications from 2020 through early 2026, the authors find that the field uses shared terminology inconsistently, that rights and privacy receive more scholarly attention than accuracy, error, and fairness, and that institutional actors are rarely held accountable when systems fail or users lack remedies. To address these gaps, the authors introduce the Age-Assurance Process Framework, which maps how evidence is evaluated, translated into age-related claims, and used in access or eligibility decisions.
- AI policy
- Quality assurance
Research
Same Patient, Different Order: Action-Level Reliability of Clinical LLM Agents Under Repeated Runs
Rohith Reddy Bellibatlu, Manpreet Singh, Zhoutian Han et al.
arXiv · 2026-09-11
This paper exposes a critical reliability gap in clinical AI agent benchmarks: the same inputs fed to a large-language-model clinical agent can produce materially different medical orders—different tests, medications, or referrals—across repeated runs, even when the benchmark reports the same pass/fail verdict each time. The authors introduce a 'same-input rerun' methodology with six reliability metrics and apply it to 1,000 runs across 50 tasks using two small open-weight models at varying temperatures, finding that under the 8B model at temperature 0.7 all 43 ordering groups produced a different set of orders across five identical runs, and in 22 of those cases the benchmark reported the same failing verdict despite materially different behavior. The study also finds that orders sometimes reach different endpoints, with one rejected by the record server while the agent was told it succeeded. The authors call for repeated-run evaluation, action-level stability reporting, and execution-faithful environment feedback as necessary safeguards before clinical LLM agents can be responsibly assessed or deployed.
- Quality assurance
- Certifications
Research
Harmfulness Propagation Dynamics: Layer-wise Trajectories of Adversarial Intent in Large Language Models
Noor Islam S. Mohammad, Uluğ Bayazıt
arXiv · 2026-09-11
This paper identifies 'Harmfulness Propagation Dynamics' (HPD), a phenomenon where harmful prompts cause a model's internal representation of harm to rise monotonically across transformer layers, while benign prompts stay flat or oscillatory. Building on this, the authors introduce HERALD, a lightweight input moderator that extracts a seven-dimensional feature vector from these cross-layer trajectories and classifies it with a tiny 288-parameter MLP, achieving an average F1 of 89.3 on OLMo2-7B across eight benchmarks and outperforming prior guard models on adversarial jailbreak detection (98.4 vs. 96.9 F1). The method requires only 262 KB of storage and adds negligible compute overhead, while also producing per-instance audit trails that reveal when and how harmfulness emerges within the model. This matters for AI safety and quality assurance by providing an efficient, interpretable, and reproducible mechanism for detecting harmful or adversarial inputs before they are acted upon.
- Quality assurance
Research
Governing at Machine Speed: An Adaptive Intelligence Architecture for Real-Time AI Policy Enforcement
Sandeep Bokkasam, B. Durgalakshmi
arXiv (Cornell University) · 2026-09-11
This paper diagnoses what it calls the 'attestation deficit'—a structural gap in which organizations adopt AI widely but cannot produce auditable, tamper-evident evidence that their governance policies are actually enforced within regulatory timelines. Drawing on the Stanford 2026 AI Index Report (362 documented incidents), the IBM/Ponemon 2026 Cost of a Data Breach study (USD 4.99M average breach cost, 92% of breached organizations lacking access controls), and the EY/AIUC-1 Consortium survey (38% end-to-end monitoring, 17% agent-to-agent coverage), the authors argue the failure is organizational and architectural rather than technical. To address this, they propose AGIL (Adaptive Governance Intelligence Layer), a conceptual five-layer architecture covering shadow AI detection, behavioral risk classification, inline policy enforcement at sub-100ms latency, continuous tamper-evident audit trail generation, and ML-driven cross-jurisdictional policy evolution. The authors explicitly note AGIL is a theoretical framework and that empirical validation through controlled deployment remains future work.
- AI policy
- Enterprise
Research
Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures
Harsh Raj, David Lee, Anas Mahmoud et al.
arXiv · 2026-09-11
This paper addresses the challenge of automated root-cause attribution (RCA) for failures in long-horizon AI agent tasks, where execution logs are too large for human review. The authors identify that existing one-shot LLM-based RCA methods perform poorly on lengthy traces because relevant evidence is sparse and distributed, causing the model to settle on early diagnoses. They propose 'Continual Search,' an iterative framework that repeatedly nudges the LLM judge to keep examining unresolved evidence across successive turns, and introduce MegaRCA-Mix, a new benchmark of 50 human-annotated long-horizon failure trials. Across multiple benchmarks and model families, Continual Search consistently improves attribution accuracy — raising GPT-5.5's F1 score by over 40% on MegaRCA-Mix — and shows that effective search strategy can allow lower-tier models to outperform higher-tier counterparts.
- Quality assurance
- Enterprise
Research
Hindsight Bias in Clinical Temporal Reasoning: How Future Data Exposure Affects Large Language Model Judgment
Misaki Matsuura, Sayantan Kumar, Ojas Kadam et al.
arXiv · 2026-09-11
This paper introduces a paired benchmark to measure 'hindsight bias' in large language models (LLMs) used for clinical reasoning—specifically, whether models give better-seeming answers when they can see future patient outcomes that real clinicians wouldn't have had access to at the time of decision-making. Using 171 PubMed Central case reports (covering sepsis and GLP-1/diabetes cases), the authors test four LLMs (GPT, Gemma, GLM, and Opus variants) under prospective versus full-timeline conditions, finding that exposure to complete timelines consistently shifts model responses toward outcome-consistent 'hindsight traps.' Critically, temporal masking—hiding future data—reduces this bias without hurting accuracy, suggesting that standard retrospective evaluations may systematically overstate model reliability for real-world clinical use.
- Quality assurance
- Certifications
Research
Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models
Utkarsh Soni, Syed Shariyar Murtaza, Yifan Nie et al.
arXiv · 2026-09-11
This paper introduces Tasks over Application Manuals (TAM), a benchmark designed to test whether large language models can follow long, rule-dense procedural documents across two real-world domains: ICD-10-CM clinical coding and U.S. federal sentencing guidelines. Each task requires navigating tens of thousands of interdependent rules across lengthy manuals to produce an exact answer. Even the best-performing approaches—including retrieval-augmented generation, ReAct-style prompting, and an agent harness on GPT-5—achieve only 1% exact-match accuracy on clinical coding and 15.5% on sentencing tasks, revealing that current LLMs are far less capable of reliable procedural reasoning than existing short-horizon benchmarks suggest. These findings have direct implications for high-stakes domains like medical coding and legal sentencing, where rule-following accuracy is essential.
- Quality assurance
- Certifications
Research
Scaling Clinical Judgment to Evaluate Medical AI
Thomas A. Buckley, Zahir Kanjee, Peter G. Brodeur et al.
arXiv · 2026-09-11
This paper introduces PrecepTron, a fine-tuned large language model designed to replicate physician-level evaluation of AI-generated clinical responses at scale. Trained via low-rank adaptation (LoRA) on a 32-billion-parameter model using a small set of physician-labeled examples, PrecepTron is validated against GRAND-ROUNDS, a new benchmark comprising 9,217 physician scores from 11 physicians across seven studies. The authors show that standard 'LLM-as-a-judge' approaches frequently disagree with physicians and with each other, while PrecepTron achieves physician-level consistency and successfully reproduces headline findings from five influential studies published in JAMA, Science, and Nature Medicine without requiring new human grading. This work matters for quality assurance in medical AI by enabling reproducible, large-scale assessment of clinical reasoning in LLMs that was previously infeasible due to the high cost and limited scale of human physician evaluation.
- Quality assurance
- Certifications
Research
When Rubrics Fail: Hallucinations Reveal Blind Spots in Medical AI Evaluation
Griffin Farrow, Lily Sijia Li, Jack Johnson et al.
arXiv · 2026-09-11
This paper investigates whether rubric-based evaluation methods—currently the dominant approach for assessing large language models (LLMs) in medicine—reliably detect clinically relevant hallucinations. Using a controlled setting with MedHallu and a novel clinician-validated error-injection pipeline spanning HealthBench, HealthBench Professional, and LiveMedBench, the authors show that medical hallucinations are frequently missed by rubrics, often leaving scores unchanged. Rubrics perform best when explicitly checking specific facts but fail for unanticipated or additional errors. The findings suggest that rubric scores alone are insufficient to establish clinical reliability and that complementary approaches, such as retrieval-based factuality checks, are needed to close these systematic blind spots.
- Quality assurance
- Certifications
Research
What is the Difference Between Me and You? Benchmarking the Quality Gap Between Human-Written and AI-Generated Code
Cristina Improta, Pietro Liguori, Domenico Cotroneo
arXiv · 2026-09-11
This paper benchmarks code quality differences between human-written and AI-generated code at scale, using 787,562 function pairs across Python, Java, and C produced by GPT models, DeepSeek-Coder, and Qwen2.5-Coder. The study finds that AI-generated code is structurally simpler—roughly half the size and branching complexity of human code—and stylistically more templated, while defect and vulnerability profiles differ by author and language: LLMs produce more and more severe security findings in Python and Java, but fewer high-severity memory-safety issues in C. Naturalness, rather than complexity metrics, is the key differentiator once code size is controlled for. The authors release CQBench, a benchmark of 27,346 issue-prone tasks with evaluation pipelines for quality assurance and security testing, providing a systematic tool for assessing AI coding assistants beyond functional correctness.
- Quality assurance
- Enterprise
Research
Understanding Game Coaching on Gig Platforms
Hwijoon Lee, Saiph Savage
arXiv (Cornell University) · 2026-09-11
This study examines how freelance game coaches operate on gig platforms like Fiverr, drawing on semi-structured interviews with 20 experienced coaches across 17 competitive games. Despite no shared formal training, coaches converged on common practices around rapport-building, individualized diagnosis, and adaptive feedback. The research identifies two structural challenges unique to this workforce: 'dual precarity' from both gig platform instability and the volatile lifecycles of live-service games, and 'earned authority' requiring coaches to continuously legitimize themselves through visible competitive achievement. Coaches welcomed AI for administrative and analytic tasks but resisted it in live coaching interactions where trust and relational judgment were seen as essential.
- Workforce
Research
When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration
CIVIC-AI Collaboration, :, Jiaying Wu et al.
arXiv (Cornell University) · 2026-09-11
This paper proposes a workflow-level framework for evaluating when AI genuinely augments human work, rather than simply automating discrete tasks. The authors define AI augmentation through six conditions covering durable net value, meaningful human control, accountability and recovery, and long-term human development including learning, career pathways, and job purpose. They apply this framework to a case study of AI-mediated social surveys and offer guidance for organizations, researchers, and policymakers on understanding the future of work. The framework matters because it shifts the focus from narrow automation metrics to the broader transformation of work through human-agent collaboration.
- Workforce
- AI policy
Research
Operational Range Bounding in Spectroscopy: A Safety Cage Framework for Machine Learning Models
Nikki Grens, Luís F. Simões, Kai Hou Yip et al.
arXiv (Cornell University) · 2026-09-11
This paper proposes a modular 'safety cage' framework that wraps around black-box machine learning models used in exoplanetary spectroscopy to detect unreliable predictions without modifying the underlying model. The framework combines uncertainty quantification, out-of-domain detection, and influence functions as parallel monitoring indicators, and shows that rejecting only 20% of predictions reduces error by 45–65% across different domains and metrics. The study demonstrates that no single indicator captures all failure modes, making indicator fusion essential. This work advances quality assurance for ML models deployed in safety-critical scientific settings where ground truth is rarely available.
- Quality assurance
Research
Beyond Establishing the Four-Day Workweek: Understanding Adaptation and Long-Term Survival in an Agile Software Organization
Michael Neumann, Darja Šmite
arXiv (Cornell University) · 2026-09-11
This longitudinal case study follows a software organization operating a 32-hour, four-day workweek (4DWW) across interviews conducted in 2022 and 2026, examining how the arrangement was introduced, adapted, and sustained over time. The research finds that the 4DWW is better understood as an evolving arrangement than a one-time intervention, with teams redesigning agile practices, coordination, and meetings to fit reduced hours. When faced with ownership changes, economic downturns, and AI market pressures, employees engaged in voluntary protective adaptations rather than reverting to a five-day week, anticipating that rollback would harm job satisfaction and employer image. The paper contributes a lifecycle model and a survival matrix explaining how external pressures and management rationale interact to determine rollback risk.
- Workforce
- Enterprise
Research
Testing Between the Test Cases: Proving End-to-End Steering in Conditions You Never Drove
Menuka Ghalan, Charles Rodgers, Zachary D. Asher
arXiv (Cornell University) · 2026-09-11
This paper applies formal verification (bound propagation) to end-to-end neural network steering models trained in the CARLA simulator, proving mathematically how far steering outputs can drift under unseen disturbance conditions—such as fog, night, and low sun—without requiring additional simulation runs. By reading trained network weights directly, the method covers combinatorial spaces (e.g., 10^133 pose-disturbance combinations on an arterial road) in minutes on a single GPU, far exceeding what simulation testing alone could achieve. Crucially, formal verification identified failure conditions for a clear-weather-only policy that simulation testing missed, offering preliminary evidence of failures 'between the test cases.' The work demonstrates a scalable approach to safety assurance for autonomous steering systems under distribution shift.
- Quality assurance
- Certifications
Research
PIA-Bench: Towards Automated Privacy Impact Assessment with Large Language Models
Jiamin Zheng, Hao-Ping Lee, Luo Mai et al.
arXiv (Cornell University) · 2026-09-11
This paper introduces PIA-Bench, the first open benchmark for evaluating large language models on privacy impact assessments (PIAs) drawn from 499 expert-authored PIAs published by US federal agencies, with 73 structured PIAs curated for evaluation. The authors find that off-the-shelf LLMs can produce meaningful privacy risk and mitigation assessments, while also identifying gaps that require domain-specific workflows, accountable LLM infrastructure, and new PIA quality standards. The work is directly relevant to automating and scaling a compliance process that currently demands scarce privacy and technical expertise.
- AI policy
- Quality assurance
Research
NovaFabric: Tamper-Evident, Replayable Evidence for Autonomous AI Agent Runs
Mohsen Seyedkazemi Ardebili
arXiv (Cornell University) · 2026-09-11
NovaFabric is a system for producing tamper-evident, replayable audit records of autonomous AI agent runs without modifying the agent itself. It packages an agent's execution into a portable 'Run Capsule' sealed with cryptographic signatures, timestamps, and Merkle logs, then exports an Evidence Bundle that third parties can verify using standard tooling. The paper evaluates eight research questions and finds that mocked replay works reliably (10/10), tampering is rejected across tested classes, and credential redaction covers 14/14 credential types, though tool-using workload replay and stream completeness have notable gaps. This matters because emerging regulations like the EU AI Act and ISO 42001 presume that independent parties can inspect records of AI actions, and NovaFabric directly addresses that audit infrastructure need.
- Certifications
- AI policy
Research
Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews
Brian Jabarian, Luca Henkel
CESifo · 2026-09-11
This natural field experiment with 70,000 real job applicants randomly assigned to human recruiters or AI voice agents finds that AI-interviewed applicants are 12% more likely to receive job offers, with downstream gains in job starts and worker retention and no productivity decline among hires. Analysis of interview transcripts shows AI agents produce more structured, consistent interviews while still adapting to individual applicants, yielding higher-quality, decision-relevant information. The study argues that AI's key advantage emerges in settings where many human workers execute the same task repeatedly, because human variance introduces noise that AI reduces through adaptive standardization.
- Workforce
- Enterprise
Research
ORQA: An Occupation-Realistic Question and Answer Framework for LLM Professional Knowledge
Shreyas Krishnan, Serina Chang, Abhishek Nagaraj
arXiv (Cornell University) · 2026-09-11
ORQA introduces a benchmarking framework that evaluates large language models on occupation-specific professional knowledge by linking O*NET occupations to trusted sources such as regulatory agencies, licensing bodies, and government publications, then converting that content into traceable question-answer pairs. The benchmark spans 116 occupations across all 21 major SOC groups, with 480 questions drawn from 187 websites. Top frontier models (Claude Opus, GPT-5, Claude Sonnet) score only 58–62%, while smaller open-weight models reach 33–41%, with near-zero performance on some occupations like Sheet Metal Workers and Fish and Game Wardens. The findings highlight significant gaps in LLM professional knowledge and offer a scalable, source-grounded method for evaluating AI readiness across occupational domains.
- Workforce
- Certifications
Research
From Review to Reuse: How Post-Task Workflow Can Support Human-AI Agent Interaction
Zekun Wu, Xinru Wang, Rock Yuren Pang et al.
arXiv (Cornell University) · 2026-09-11
This paper investigates 'post-task workflows'—editable, graph-based representations of an AI agent's completed execution—as a way to help users understand, validate, and reuse automated multi-step processes. The researchers analyzed over 10,000 real-world workflow templates and built a tool called Trace2Flow, then tested it with 20 participants reviewing agent executions containing errors. They found that post-task workflows improved user understanding and error detection compared to reviewing only the original prompt, and that adapting a workflow for follow-up tasks was as effective and often preferred over rewriting the prompt.
- Workforce
- Enterprise
Research
How Do Data Collection Strategy and Data Quality Influence the Outcomes of Digital Technology Adoption?
Xuejiao Li, Cheng Yang
arXiv (Cornell University) · 2026-09-11
Drawing on survey data from 86 manufacturing firms, this study uses PLS-SEM to show that data collection strategy and data quality jointly drive successful digital technology adoption. Data quality fully mediates the relationship between data collection strategy and implementation performance, meaning upstream data practices affect digital outcomes only through the quality of data they produce. The findings offer practical guidance for managers designing data governance frameworks, and call for future research integrating quality assurance, standardization, and lifecycle management into data-driven transformation strategies.
- Enterprise
- Quality assurance
Research
Human Touch vs. Tech Talk: Mental Health Clinicians' Perspectives on Artificial Intelligence in Systemic Practice
Merranda Marin, Marcel Montañez, Kourtney Vaillancourt
Journal of Marital and Family Therapy · 2026-09-11
This mixed-methods study surveyed 89 U.S. mental health professionals about their familiarity with and attitudes toward AI tools in clinical practice. About half reported using AI professionally, most commonly for notetaking and transcription, and respondents expressed conditional acceptance—valuing AI for clinician support tasks while raising concerns about confidentiality, professional judgment, and the erosion of human connection. The findings highlight a pressing need for ethically grounded, relationally informed frameworks for AI integration in mental health care.
- Workforce
- AI policy