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
Artificial Intelligence Tool Adoption Among Employees of the Philippine Electronics and Communication Institute of Technology (PECIT): A Qualitative Exploration of Issues and Challenges
Gerardo Jr. S. Carlos, Daryll A. Cabagay, Maria Cecilia Z. Matillano et al.
Proceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...) · 2026-09-01
This qualitative study examines how teaching and non-teaching staff at a Philippine private higher education institution (PECIT) are adopting AI tools such as ChatGPT, Gemini, and Copilot. Using semi-structured interviews and thematic analysis, the research finds that employees use these tools for lesson preparation, research, communication, and administrative tasks, reporting gains in productivity and efficiency. Key challenges include inaccurate AI outputs, difficulties in prompt formulation, and the need to verify AI-generated content. Participants recommended continuous AI training, clear institutional policies, and responsible AI practices to support sustainable integration.
- Workforce
- AI policy
Research
Human-AI Integration in Industry 5.0: Mapping Relational Patterns among Antecedents, Mechanisms, and Outcomes
Mirco Avallone, Gianluca Aquilone, Antonello Cammarano et al.
Journal of Industrial Information Integration · 2026-09-01
This paper develops an Antecedent-Process-Outcome framework for human-AI integration in Industry 5.0 industrial settings, consolidating 70 constructs and coding 162 AI-oriented industrial practices. Using multi-layer association and network analyses, it finds that augmentation-oriented human-AI collaboration aligns more strongly with positive performance and knowledge outcomes than automation-oriented arrangements, which show weaker links to human-centric results. The study identifies structurally central constructs that differ from those most emphasized in existing literature, offering evidence-based guidance for sustainable process innovation and adaptive capacity in industrial workplaces.
- Workforce
- Enterprise
Research
Artificial Intelligence in Academic Research: Adoption, Productivity, and Skill Perceptions
Alessandro Muscio, Sotaro Shibayama
Technology in Society · 2026-09-01
This survey of Italian academics finds that 39% report using AI in their research, with adoption highest among younger scholars and those in applied or interdisciplinary fields. AI use is positively associated with self-reported research productivity and more optimistic expectations about skill development, though trust in AI is uneven and skepticism persists across segments of the academic population. The study provides empirical evidence on how AI is reshaping research workflows and researcher skill perceptions in higher education.
- Workforce
Research
Algorithmic Publicity and the Right to a Fair Trial
Ariss Laghai
Bond Law Review · 2026-09-01
This article analyzes how AI-driven algorithmic systems—including social media feeds, recommender engines, synthetic media, and generative AI—threaten the right to a fair trial in Australian criminal jury proceedings by amplifying and personalizing prejudicial publicity in ways that undermine traditional broadcast-era safeguards. Drawing on empirical research on juror psychology and internet use, it evaluates existing controls such as contempt, stays, jury directions, and judge-alone trials, finding them inadequate against persistent and searchable digital content. The article proposes doctrinal refinements, targeted platform duties, and court-supervised AI monitoring tools, drawing on developments in the UK, EU, US, China, and Estonia. The work is directly relevant to policy debates about how legal systems should regulate AI and platform behavior to preserve fair-trial guarantees.
- AI policy
Research
Governing the Algorithmic Black Box in Talent Acquisition: Towards a Multi-Level Framework of Contestable Accountability
Ramniyata Jairath
Frontiers in Social Sciences Research · 2026-09-01
This conceptual paper argues that existing governance tools for AI-driven hiring systems—such as transparency disclosures, explainability requirements, and algorithmic audits—rest on assumptions that the recruitment context systematically violates. The authors develop a multi-level 'contestable accountability' framework spanning epistemic, procedural, institutional, and contestatory levels, and introduce two diagnostic concepts—vertical displacement and temporal displacement—to explain why governance efforts persistently fail to protect job candidates. The paper reframes the policy problem from making algorithms explainable to making hiring decisions genuinely answerable, and situates its analysis against regulatory landscapes in the EU, US, and India. The work matters because it offers a governance-theoretic vocabulary for addressing AI opacity in talent acquisition and advances nine testable propositions for future research and regulatory design.
- Workforce
- AI policy
Research
Use of artificial intelligence in education and training of radiology
María Belén Morales-Cevallos, Canva Byron Ma Lam, María José López Pino et al.
Frontiers in Radiology · 2026-09-01
This scoping review of 29 studies (2020–2025) examines how artificial intelligence is being integrated into radiology education and training. Skill development was the most investigated outcome (55.2% of studies), and approximately 86% of studies reported positive or improved educational outcomes, with AI-based interventions enhancing learner confidence, AI literacy, diagnostic reasoning, and readiness for clinical implementation. Generative AI tools showed promise for tutoring and assessment but raised concerns around reliability, hallucinations, and bias. The authors conclude that sustainable AI integration requires standardized curricula, faculty development, and ethical oversight to prepare radiology professionals for AI-integrated healthcare.
- Workforce
- Certifications
Research
Stakeholder perceptions of artificial intelligence for climate adaptation in Somalia: Opportunities, readiness barriers, and community resilience
Aweis Ahmed Hussein Jarras, Abdimalik Aden Ibrahim, Abdullahi Ilyas Osman
Progress in Disaster Science · 2026-09-01
This study surveys 353 stakeholders across government, NGOs, academia, and the private sector in Somalia to assess perceptions of AI for climate adaptation. Respondents broadly valued AI for early warning systems, disaster response, agricultural decision-making, and environmental monitoring, yet the strongest consensus was that Somali communities lack the technical capacity to use AI tools, followed by financial and infrastructure barriers. The authors frame this 'high demand–low preparedness' paradox through a proposed 'Readiness-Adjusted Technology Adoption' framework, arguing that perceived usefulness alone cannot drive adoption without complementary investments in capacity, infrastructure, and institutions. The paper offers practical recommendations including mobile-first, voice-based AI tools, community digital literacy programs, and governance reforms relevant to fragile state policymakers and donors.
- AI policy
- Workforce
Research
Civil Liability for Damage Caused Due to the Use of Artificial Intelligence: an Unpopular Study of a Popular Topic
Nataliia Filatova-Bilous
Civìlìstična platforma. · 2026-09-01
This article analyzes civil liability for AI-caused damages, examining regulatory frameworks in the EU and US and their implications for Ukrainian law. The authors argue that Ukraine does not yet need a special AI liability regime, but should instead incorporate EU Directive 2024/2853 on defective products, adapt traditional tort concepts to address the difficulty of establishing fault and causation in AI contexts, and update procedural rules for AI-related evidence. The paper highlights core technical challenges—including AI unpredictability, opacity, and self-learning—that complicate existing legal frameworks.
- AI policy
Research
Digital adoption, AI integration, and labor productivity: empirical insights from the European Union
Mercy Minoo Kavele
Labour & Industry a journal of the social and economic relations of work · 2026-09-01
This study analyzes the relationship between enterprise AI adoption, workforce digital skills, and labor productivity across all 27 EU member states using Eurostat and OECD data, OLS regression, and Pearson correlation. While AI adoption and digital skills show positive bivariate associations with labor productivity, GDP per capita is the only statistically significant predictor once economic development is controlled for. The findings suggest that digital technologies alone do not guarantee productivity gains and that complementary economic capabilities are necessary to translate AI and digital transformation into measurable labor productivity improvements. The study contributes empirical evidence relevant to EU-level debates on workforce upskilling and enterprise digital investment strategies.
- Workforce
- Enterprise
Research
Auditing the Algorithmic Leviathan: A Tiered Accountability and Reporting Standards Framework for Democratic Public Administration
Haris Alibašić
Standards · 2026-09-01
This paper develops a tiered accountability and reporting standards framework—organized around five auditable primitives (provenance tracking, decision logging, role attribution, contestability, and post-deployment audit)—for algorithmic systems used in public administration. Drawing on case studies including the DOGE-Treasury access episode, Australia's Robodebt scheme, and governance arrangements in six countries, the authors argue that current standardization focuses too narrowly on AI system certification and neglects institutional answerability. The framework introduces a Public Sector Algorithmic Accountability Statement (PAAS) with ten disclosure fields crosswalked to GRI standards, escalating obligations across minimum, heightened, and systemic/constitutional tiers based on the severity of public authority consequences. The proposal aims to bridge the gap between technical AI assurance and democratic governance accountability.
- AI policy
- Certifications
Research
Should Businesses Trust AI Advice? A Methodology to Audit the Ethical Integrity of Chatbots
Manuel Chaves-Maza
Computers in Human Behavior Reports · 2026-09-01
This paper introduces the Adaptive Ethical Evaluation Protocol (AEEP), a validated audit methodology for assessing whether AI chatbots maintain consistent ethical stances when subjected to business pressure in small and medium-sized enterprise (SME) advisory contexts. The protocol stages structured five-node adaptive dialogues across ten real-world SME dilemmas and was applied to five frontier LLMs, achieving 93.8% algorithm–expert agreement (Cohen's κ = 0.728). Results revealed meaningful behavioral differences between models—Claude was most consistent under pressure while Grok wavered most—providing enterprise advisors, regulators, and SME managers with a reusable tool to identify where AI advice can be trusted and where human oversight remains necessary.
- Enterprise
- Quality assurance
- AI policy
Research
Development and validation of a human-supervised AI-augmented living oncology evidence platform: a breast cancer pilot study
H. Rugo, A. Forsythe, D. Flora et al.
ESMO Real World Data and Digital Oncology · 2026-09-01
This paper describes and validates a living oncology evidence platform (Living-OEP) for breast cancer that uses an agentic AI system—combining GPT-4.1, o3, and Claude Sonnet-4—to perform daily, human-supervised systematic literature review compliant with Cochrane standards. Trained on over 29,000 annotated clinical trial abstracts, the system achieved review accuracy of 95.1%–97.2% and extraction accuracy up to 99.4% against human annotations, outperforming general AI chatbots on comprehensiveness and accuracy across eight breast cancer treatment scenarios. By integrating structured evidence with guideline-based treatment pathways and FDA labels in real time, the platform aims to help oncologists keep pace with rapidly evolving clinical data. The authors note that future studies are needed to assess impact on physician workflows and clinical decision-making.
- Enterprise
- Quality assurance
Research
Current Landscape of Curriculum Development and Implementation in Medical Artificial Intelligence: A Scoping Review
Yue Wang, He Wang, Ting Wang et al.
Journal of Multidisciplinary Healthcare · 2026-09-01
This scoping review examines 36 implemented medical AI educational programs published between 2021 and April 2026, finding that such programs have grown rapidly but remain largely in pilot stages with small class sizes. Most curricula were developed based on expert experience rather than standardized frameworks, focused on understanding-level learning objectives, and relied on student feedback for evaluation rather than objective assessment tools. The review identifies persistent gaps including lack of standardized curriculum frameworks, limited affective learning objectives, weak integration with existing medical curricula, and insufficient involvement of instructional designers. The authors conclude that medical AI education is still in its early stages and call for accelerated curriculum development, stronger practical and affective competency training, and more rigorous evaluation systems.
- Workforce
- Certifications
Research
Literature Review of Existing Environmental Policies Related to AI Data Centers: Current Regulatory Tools, Applicability, and Shortcomings
Andy Zhang
Future Trends in AI Research · 2026-09-01
This literature review examines whether existing environmental regulations in the US, EU, and subnational jurisdictions can adequately govern the environmental impacts of AI data centers, which concentrate electricity demand, cooling loads, water use, and hardware turnover at unprecedented scale. The review finds that existing law does reach many aspects of AI data center operations—through electricity regulation, water permitting, air quality rules, and waste law—but only partially and unevenly, as most policies regulate inputs or side effects rather than AI operators directly. Key shortcomings include fragmented governance across agencies, incomplete reporting on facility-level water use and embodied emissions, and an inability to address cumulative grid and aquifer impacts. The core policy problem identified is a mismatch between legacy environmental governance frameworks and the scale, speed, opacity, and local concentration of AI infrastructure growth.
- AI policy
Research
Adaptive, ethical and responsible AI governance for smart cities and nations
Z. R. M. Abdullah Kaiser
Discover Cities · 2026-09-01
This paper develops a six-step adaptive, ethical, and responsible AI governance framework for smart cities and nations, integrating legal rules, ethical oversight, organizational capacity, and human–AI decision-making arrangements. It identifies cross-sector governance risks including algorithmic bias, weak oversight, regulatory gaps, vendor dependence, and legitimacy deficits, and distinguishes between 'governance of AI' (regulation and oversight of AI systems) and 'governance by AI' (AI inputs into public decisions). The framework is applied illustratively to New York City and Singapore to demonstrate how its six governance dimensions function across decentralized and centralized contexts. The paper argues that AI governance must remain adaptive rather than fixed, embedding risk management, responsiveness, and ethics-by-design across the full AI lifecycle.
- AI policy
Research
Fallibility, persuadability, and correctability of large language models under sustained conversational misinformation pressure
Jordan Rodriguez, Zachary Hansen, Luis De Anda et al.
Scientific Reports · 2026-09-01
This study systematically evaluated seven major large language models (ChatGPT GPT-3.5/4o/4o-mini, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama-3-70B, and DeepSeek) on three dimensions of susceptibility to conversational misinformation: fallibility, persuadability, and correctability. Across 50-repetition sequences of 100 purposefully false statements, misinformation affirmation rates varied more than 150-fold across models, and a novel 'conversational reverberation' phenomenon was identified where models oscillated between accepting and rejecting the same false statement. Susceptibility was significantly influenced by informational obscurity under repetitive conditions, implicating training data frequency as a key factor, while correctability was highly heterogeneous—four models achieved 100% self-correction but the most accurate model failed to correct any of its rare errors. These findings reveal failure modes invisible to standard evaluation benchmarks, with direct implications for model selection in truth-critical deployment contexts.
- Quality assurance
- Enterprise
Research
Artificial Intelligence in Sustainability Assurance: Accounting Challenges, Audit Risks and a Conceptual Framework for ESG Verification
Radosveta Krasteva-Hristova, Vanya Georgieva
Accounting and Auditing · 2026-09-01
This conceptual paper develops the Responsible AI-Assisted Sustainability Assurance Framework to address how AI should be integrated into ESG verification and sustainability assurance workflows. It identifies five domains where AI adds analytical value—evidence extraction, criteria mapping, anomaly and greenwashing screening, external-data triangulation, and documentation support—while cataloguing risks such as data fidelity, explainability, bias, auditor overreliance, and preparer gaming. The framework sets graded reliance ceilings and prohibits autonomous AI decisions on materiality, evidence sufficiency, and conclusions, offering a testable model grounded in ISSA 5000 and illustrated in a European regulatory context. The work matters because it provides structured governance guidance for accounting and audit professionals navigating the shift from voluntary to regulated, externally assured sustainability reporting.
- Quality assurance
- Certifications
Research
Two Structures of AI Power: National Capacity, Relational Boundary Control, and Systemic Influence in the Emerging AI Order
Shaoyuan Wu
arXiv · 2026-09-01
This working paper proposes a two-structure analytical framework that distinguishes between 'AI National Power' (a country's capacity to generate, mobilize, and sustain AI-relevant resources) and 'AI Boundary Power' (an actor's ability to control the conditions under which another actor can access, transfer, or deploy AI resources). Using export controls on advanced computing and EU AI regulation as illustrative mechanisms, the paper argues these two forms of power are distinct but potentially complementary. The framework is relevant for understanding how policy instruments like export controls and regulatory regimes shape the geopolitics of AI development.
- AI policy
Research
Acceptance Without Choice: Mandated Use and the Limits of Acceptance Theory in AI-Enabled Recruitment
Ramniyata Jairath
Frontiers in Social Sciences Research · 2026-09-01
This conceptual paper critiques how technology acceptance theory (TAT) is applied to AI-enabled recruitment, arguing the approach is fundamentally misspecified when used on populations—both recruiters and candidates—who cannot opt out of the technology. The authors trace the problem to an unresolved assumption in TAT's own lineage: that when use is mandatory, measured 'intention' reflects compliance rather than genuine acceptance, undermining the dependent variable's meaning. The paper proposes relocating the scope condition from the setting to the behavior, limiting acceptance theory to whatever discretionary latitude remains, and distinguishes behavioral conformity from attitudinal endorsement with seven propositions. Notably, it highlights that candidates subjected to AI recruitment tools have no equivalent in the enterprise systems literature from which mandatory-use arguments are typically borrowed, representing a gap with direct implications for fairness and organizational legitimacy.
- Workforce
- AI policy
Research
Conversation Coach: A Voice-enabled AI System that Helps Practice Difficult Workplace Conversations
Fanyou Wu, Suraj Maharjan, Ainur Yessenalina et al.
arXiv · 2026-08-31
Conversation Coach is a voice-enabled AI system designed to help managers rehearse difficult workplace conversations, such as performance reviews and coaching sessions, in a realistic spoken format. The paper compares two architectures—an end-to-end speech-to-speech model and a cascaded approach using automatic speech recognition, a large language model, and text-to-speech—finding the end-to-end approach offers 3× lower median latency and an estimated 8× lower cost, while the cascaded approach provides superior reasoning quality for coaching feedback. The cascaded system was deployed in production, where over 40,000 managers used it across six months, with usage patterns suggesting selective engagement for high-stakes conversations. The work demonstrates that scalable AI-driven voice coaching can reduce the cost of manager training while enabling personalized, policy-aware feedback at scale.
- Workforce
- Enterprise
Research
Do LLMs Know Your Neighborhood? Auditing LLM Priors for Neighborhood-Level Mobility Prediction and Structural Alignment
Saad Mohammad Abrar, Eesha Kurella, Arnav Dadarya et al.
arXiv · 2026-08-31
This paper evaluates whether large language models (LLMs) can predict neighborhood-level human mobility across four U.S. metropolitan areas without task-specific training (zero-shot), comparing them against supervised baselines using anonymized Cuebiq mobility data. Supervised models achieved 0.580 average accuracy versus 0.435 for the best LLM, with LLMs performing especially poorly on spatial extent outcomes. A directional alignment analysis reveals that LLMs rely on coarse, stable predictor-level priors that do not meaningfully vary across cities or outcomes, and that they exhibit asymmetric treatment of protected-group predictors—raising concerns about embedded bias. The findings suggest LLM mobility predictions should not be treated as structurally grounded substitutes for empirical data without auditing for alignment and potential discriminatory patterns.
- AI policy
- Quality assurance
Research
Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries
Phuong Anh Nguyen, Jill Noorily, Matthew Flathers et al.
arXiv · 2026-08-31
This paper audits how three AI consumer platforms—ChatGPT, Perplexity, and Google AI Overview—cite sources when answering mental health questions, recording nearly 16,000 citations across English and six other languages. The study finds citations are heavily concentrated (the top ten domains account for 43.6% of English citations), platforms differ sharply in source consistency and type preferences, and non-English queries receive fewer and less language-appropriate resources. The findings reveal that generative AI systems are now performing source curation on behalf of users, yet do so unevenly across languages and platforms, raising concerns about information equity and reliability. The authors release their typology, classifier, and annotated corpus to enable future audits of AI-generated health information.
- Quality assurance
- AI policy
Research
Workload Identification with Physical Side Channels for AI Governance
Simone Gargiulo, Gabriel Kulp
arXiv · 2026-08-31
This paper demonstrates that an external observer can identify what type of AI workload—training, inference, or non-AI computation—is running on an NVIDIA H200 GPU by analyzing its physical power draw, without relying on operator-reported telemetry that could be spoofed. Using 930 five-second power traces at ~10 MHz across 17 open LLM families and 25 non-AI workloads, the authors achieve 97% accuracy and a macro-averaged F1 of 0.955 in separating these workload classes on unseen model families. The study also tests four adversarial evasion strategies designed to disguise training as inference, finding that a hardened detector catches training at least 99% of the time for three of the four strategies. These findings are directly relevant to AI governance, offering a tamper-resistant technical mechanism for verifying how AI compute is being used by frontier labs or other operators under international agreements.
- AI policy
- Certifications
Research
The Irreversibility Budget: Fleet-Level Risk Accounting and Admission Control for Agent Operating Systems
Bardia Mohammadi, Laurent Bindschaedler
arXiv · 2026-08-31
This paper identifies a critical gap in how fleets of LLM agents are currently controlled: individually authorized agents can collectively exceed a principal's total acceptable risk even when each local approval is correct. The authors propose an 'irreversibility budget'—a runtime-enforced cumulative account of residual value-at-risk shared across all agents, workflows, and tenants for a given principal—that denies any marginal action that would push the aggregate past the limit. A controlled study shows that per-effect gates alone allow fleet-level overdraws of up to 48 times the tenant's risk limit, while the budget mechanism keeps correctly charged runs within that limit. The central open problem is accurate and adversarially robust pricing of heterogeneous, correlated effects across agents.
- Enterprise
- AI policy
Research
Delegation Without Trust: An Empirical Gap Analysis of Identity, Authorization, and Runtime Governance in Multi-Agent LLM Systems
Panduranga Sai Varma Dantuluri, Jyotirmoy Sundi
arXiv · 2026-08-31
This paper investigates the security of multi-agent LLM systems where agents hold credentials, call tools, and spawn sub-agents on a user's behalf. The authors define a threat model with four adversaries—confused deputy, token theft and replay, prompt-injection privilege escalation, and compromised sub-agents—and derive eight security requirements that governed agent systems must meet. They then show that common frameworks (LangGraph, CrewAI, AutoGen, and MCP) fail to provide adequate built-in confinement, and demonstrate that their authorization broker implementation blocks all four threats, accepts 0 of 200,000 forged tokens, and limits a compromised sub-agent to a mean of 1.5 reachable actions versus 8,100 under bearer delegation, enforcing decisions at roughly 2.6 microseconds each. The findings have direct implications for enterprise deployment of agentic AI systems and the policies governing delegation of authority in automated pipelines.
- Enterprise
- AI policy
- Quality assurance