AI Frontier Institute

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

  Five Primitives for Governing Autonomous AI Agents at Runtime

  Jiten Oswal, John Cadeddu

  arXiv (Cornell University) · 2026-08-27

  This paper identifies a fundamental mismatch between existing enterprise control models—designed for human users and long-lived services—and the governance demands of autonomous AI agents, which are ephemeral, have unpredictable action sets, and can be created by anyone with API access. The authors argue that governing such agents is a runtime problem and derive five core primitives required before and after an agent action takes effect: discovery, identity, governance, attestation, and supply chain. They describe a working implementation where agent actions are mediated against policy, authorized against a per-tenant action vocabulary, and recorded in a verifiable hash-linked signed ledger, with four of the five primitives already running in private pilots. The work directly informs enterprise deployments of AI agents by exposing what fails when each primitive is absent and being transparent about real architectural costs such as critical-path enforcement latency and availability trade-offs under fail-closed mediation.

  • Enterprise
  • AI policy

• Research

  Risks and Controls for Multi-Agent Systems: an analytical framework for deployment of AI agents across organisational boundaries

  Alistair Reid, Simon O'Callaghan, Dustin Venini et al.

  arXiv (Cornell University) · 2026-08-27

  This report presents an analytical framework for reasoning about the risks that arise when AI agents interact with each other across organisational boundaries. It defines three deployment tiers—singular governance (one organisation governs all agents), federated governance (multiple organisations share agreed rules), and open environments (no central authority)—and within each tier examines risk factors, failure modes, and available controls. The framework identifies which actors are positioned to apply each control and flags gaps where no single actor can act, characterising the collective action needed to close them. It is directly relevant to organisations, policymakers, and researchers deploying or regulating multi-agent AI systems.

  • AI policy
  • Enterprise

• Research

  Benchmarking Clinical Decision Pathway Adherence in Large Language Models

  Nuo Chen, Xinyang Jiang, Zilong Wang et al.

  arXiv · 2026-08-27

  This paper introduces MEGA-CDP, a benchmark designed to evaluate whether large language models (LLMs) can follow clinical decision pathways (CDPs) as defined by clinical practice guidelines, rather than simply producing correct final answers. Built from 2,274 English and Chinese clinical practice guidelines, it generates 42,353 clinical cases with explicit reference pathways and supports both single-turn and multi-turn evaluation settings. Experiments across 16 LLMs reveal that reliable guideline-adherent clinical decision support remains a significant challenge for current models. The work highlights the importance of evaluating process-level adherence to guidelines, not just outcome accuracy, for safe medical AI deployment.

  • Quality assurance
  • Certifications

• Research

  PCFBench: A Diagnostic Benchmark for Product Carbon Footprint Estimation

  Krishna Rao, Andrew Dumit, Shaena Ulissi et al.

  arXiv (Cornell University) · 2026-08-27

  PCFBench introduces the first benchmark for evaluating AI systems on the task of estimating product carbon footprints (PCFs), decomposing the workflow into six independently-scored tasks covering decomposition, retrieval, ontology matching, and numerical extraction across 614 expert-labelled items. Testing eight frontier LLMs from four providers reveals that while the strongest models estimate total emissions within 2x of declared totals on 77% of products, accuracy drops to 37-58% when PCFs are generated step by step, with only 45-75% of outputs obeying mass conservation. These findings show that AI agents routinely fail at intermediate reasoning steps in a high-stakes decarbonization workflow, with errors that are hidden when only final outputs are evaluated. The benchmark and evaluation harness are released to support targeted improvement in AI reliability for sustainability applications.

  • Quality assurance
  • Enterprise

• Research

  Operationalizing Regulations into Code: A Model to Enhance Governance and Compliance in LLM Selection for Software Engineering

  Jonysberg Quintino, Hermano de Moura, Filipe Calegário

  arXiv (Cornell University) · 2026-08-27

  This paper proposes a three-layer model for selecting Large Language Models (LLMs) in software engineering projects while ensuring compliance with regulations such as the EU AI Act, GDPR, NIST AI RMF, and ISO/IEC 42001. The model uses a multi-criteria decision matrix with knock-out and weighted scoring criteria, and includes a regulatory feedback loop for iterative refinement. A pilot evaluation using 20 adversarial scenarios based on CWE and OWASP Top 10 found distinct risk profiles between commercial cloud-based LLMs and local open-source LLMs, providing preliminary evidence that regulatory disqualification logic can prevent selection of technically capable but compliance-risky models. The work matters because it operationalizes complex regulatory obligations into actionable technical decision criteria for software development teams.

  • AI policy
  • Enterprise

• Research

  Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap

  Jacopo Dardini, Claudio Stanzione, Giordano Colò et al.

  arXiv (Cornell University) · 2026-08-27

  This paper demonstrates that post-training quantization of Large Language Models can activate hidden backdoors that are undetectable at full precision, exposing a structural 'validation–deployment gap' in current auditing practices. The authors formalize this vulnerability through Quantization Behavioral Equivalence Classes and show that certifying a model at source precision does not guarantee safe behavior after INT8 or 4-bit compression. Experiments on tactical machine translation and political content analysis show up to 85% friend-foe corruption inversion and measurable ideological bias shifts after quantization, with attack persistence varying across quantization schemes and architectures. The findings argue that behavioral certification must include the final deployed configuration, not just the source-precision checkpoint.

  • Certifications
  • Quality assurance

• Research

  Thomson: Continual Learning of Frontier Models for SovereignAI

  Shengzhuang Chen, Jerrod Parker, Yejin Bang et al.

  arXiv (Cornell University) · 2026-08-27

  This paper presents Thomson, a frontier AI model built using Continual Learning on open-weight models to demonstrate that high-performing AI can be developed by a broader range of institutions, not just heavily funded players. The authors argue that their mid- and post-training approach—which preserves both plasticity and stability—achieves gains comparable to multiple successive model generations at substantially lower compute and personnel budgets. Thomson performs competitively with recent frontier models on agentic tasks, safety, legal, tax, multilingual, and deep research benchmarks, while largely avoiding the forgetting problems common to narrow domain adaptation. The work directly addresses the concept of SovereignAI, enabling more organizations to independently build, deploy, and govern their own AI systems.

  • Enterprise
  • AI policy

• Research

  When Safety Becomes a Competitive Disadvantage: The Structural Dangers of the Frontier AI Race

  Dean Gary Egan

  Zenodo (CERN European Organization for Nuclear Research) · 2026-08-27

  This paper argues that frontier AI development creates a structural race hazard where safety-conscious behavior—such as delaying deployment, conducting deeper evaluations, or accepting oversight—imposes concentrated costs on cautious actors while spreading benefits to competitors and the public. The authors identify seven conditions that intensify this hazard and review evidence from independent assessments and developers' own safety frameworks, finding that leading actors themselves acknowledge competitor-dependent safeguards and potential races to the bottom. The paper proposes a 'Race Hazard Audit' and a governance design standard requiring that restraint be reciprocal, verifiable, competitively survivable, and enforceable. The core concern is that competitive dynamics may systematically select against the very precautions all participants depend on.

  • AI policy

• Research

  When Safety Becomes a Competitive Disadvantage: The Structural Dangers of the Frontier AI Race

  Dean Gary Egan

  Zenodo (CERN European Organization for Nuclear Research) · 2026-08-27

  This paper analyzes why AI safety measures can become a competitive disadvantage in the frontier AI race, identifying seven structural conditions—such as speed premiums, externalised harm, and weak reciprocity—that create incentives for developers to underinvest in safety precautions. Drawing on independent assessments and developers' own safety frameworks, the authors find that leading actors acknowledge competitor-dependent safeguards and potential races to the bottom, even if internal decision-making remains largely unobservable. The paper proposes a 'Race Hazard Audit' and a governance design standard requiring that safety restraint be made reciprocal, verifiable, competitively survivable, and enforceable. The work matters because it provides a mechanism-based framework for policymakers to prevent competitive dynamics from systematically selecting against the very precautions that protect all participants and the public.

  • AI policy

• Research

  Precision psychiatry in clinical practice: What is clinically actionable, what is promising, and what remains experimental?

  Júlio Torales, Marcelo O’Higgins, Iván Barrios et al.

  International Review of Psychiatry · 2026-08-27

  This structured narrative review evaluates the clinical readiness of precision psychiatry approaches—ranging from pharmacogenetics and digital phenotyping to AI-assisted decision support and multimodal prediction models—using a framework covering validity, replication, actionability, and equity. The review finds that only a few approaches (e.g., measurement-based care, therapeutic drug monitoring, selected pharmacogenetic interactions) are ready for clinical use, while AI-assisted decision support and neuroimaging markers are promising but not yet routine, and automated suicide prediction or fully automated treatment selection remain experimental. Major barriers across all domains include poor external validation, limited evidence of patient benefit, underrepresentation of diverse populations, and unresolved accountability concerns. The findings matter for quality assurance and policy by setting clear benchmarks for when and how AI and precision tools should be adopted in psychiatric care.

  • Quality assurance
  • AI policy

• Research

  Developing a Neuroimaging Training Framework and Certificate Program

  Aneta Kielar, Dianne Patterson, Thomas Hicks

  PsyArXiv (OSF Preprints) · 2026-08-27

  This paper describes the design and evolution of a fully asynchronous neuroimaging education framework at the University of Arizona that trains graduate and undergraduate students, research staff, and faculty in computational skills such as BIDS, Unix, HPC, JupyterLab, and containerization. The program offers both à la carte lessons and a complete neuroimaging certificate path to accommodate learners who prioritize flexibility over formal credentialing. A notable contribution is the framework's approach to AI-resilient assessment, replacing multiple-choice questions with interactive matching questions and structured generative AI workflows while retaining hands-on practicals. The paper positions this case study as a template for institutions seeking to deliver targeted, on-demand computational neuroimaging training to broad and diverse audiences.

  • Certifications
  • Workforce

• Research

  Gender, Power, and Law in European Digital Migration Systems

  Tatiana Maria Cernicova-Dragomir

  Ars æqui · 2026-08-27

  This legal analysis examines whether the EU's current regulatory framework—including the Charter of Fundamental Rights, the GDPR, and the AI Act—adequately protects migrant women and gender-diverse persons from structural inequalities embedded in digital migration systems such as biometric identification and algorithmic risk assessment. Using a doctrinal methodology informed by feminist legal theory, the paper argues that existing safeguards on data protection, non-discrimination, and administrative justice are insufficient because bias arises not from explicit discriminatory rules but from ostensibly neutral digital infrastructures. The article calls for a more substantive interpretation of equality and stronger procedural accountability in regulating automated border governance tools. Its findings are relevant to ongoing EU policy debates about governing AI-driven migration systems fairly.

  • AI policy

• Research

  Containing AI: from responsible to critical AI governance

  Paul Schütze, Benedetta Brevini

  AI & Society · 2026-08-27

  This paper critiques mainstream responsible AI (RAI) governance frameworks, arguing they fail to genuinely address AI's environmental and social costs and instead create an illusion of accountability—sometimes producing a regulatory vacuum rather than meaningful oversight. Using 'sustainable AI' as a case study, the authors propose 'critical AI governance,' which reframes the core question from how to make AI responsible to whether AI is needed at all. The framework rests on three principles: questioning dominant ideologies and economic imperatives, prioritizing social and political solutions over technological fixes, and investing in public digital infrastructure and independent research. The paper has direct implications for how governments and institutions design and enforce AI policy.

  • AI policy

• Research

  Artificial Intelligence and Its Influences on Enterprise Employees

  Zuhao Yu

  Empirical Studies in Econimics and Finance · 2026-08-27

  This systematic review synthesizes 186 peer-reviewed studies (January 2024–May 2026) to map how AI and generative AI affect enterprise employees across productivity and psychological dimensions. The findings show AI both boosts operational efficiency by automating routine tasks and heightens job insecurity, reducing job satisfaction and well-being—with frontline workers bearing disproportionately greater risks than managers. Organizational AI training and transparent algorithmic governance are identified as effective mitigants. The paper develops a theoretical framework and offers practical guidance for human resource management during intelligent transformation.

  • Workforce
  • Enterprise

• Research

  Governing at a distance: The EU AI act and GDPR as twin pillars of digital law and corporate governance

  Maria de L. Haynes, Akram Al Matarneh

  Corporate Law & Governance Review · 2026-08-27

  This article examines how the EU AI Act and GDPR function together as complementary regulatory frameworks that reshape corporate governance well beyond EU borders. Using doctrinal and comparative legal methodology, the study finds that both instruments embed compliance obligations—including risk management systems, internal controls, board oversight, and accountability processes—directly into corporate practice, a principle the authors call 'governance-by-design.' The paper concludes that together the two frameworks enable the EU to govern corporate conduct at a distance globally, offering an analytical framework for both academics and practitioners navigating AI and data regulation.

  • AI policy
  • Enterprise

• Research

  Multisite Implementation of a Digital Wound Model of Care: A Post‐Implementation Multimethods Evaluation of Patient and Clinician Perspectives and Lessons Learned

  Michelle Barakat‐Johnson, Clementine Mak, Michelle Lai et al.

  International Wound Journal · 2026-08-27

  This Australian multisite study evaluated a district-wide digital wound care model combining an AI-enabled application with a virtual command centre across nine facilities. High patient satisfaction was reported—91% rated care as excellent or good—and clinicians valued improved documentation consistency, though barriers like poor connectivity, time pressures, and training burden were identified. The findings demonstrate strong acceptability of AI-assisted wound care delivery while flagging patient involvement in decision-making and clinician workflow integration as areas needing further development.

  • Quality assurance
  • Workforce

• Research

  Not Afraid of AI yet: Indonesian TV Journalists on Negotiating Roles in AI-Assisted Newsrooms

  Yearry Panji Setianto, Rossalyn Asmarantika, Veronika et al.

  Media and Communication · 2026-08-27

  Drawing on interviews with 20 Indonesian television journalists across 12 national newsrooms, this study finds that journalists largely view AI as a supplementary tool for routine production tasks rather than a replacement for core journalistic functions such as editorial judgment, verification, and contextual interpretation. Journalists see video capture and accountability reporting as irreplaceably human, even as they navigate economic pressures and newsroom restructuring. The findings illuminate how journalists in a Global South media context negotiate professional identity and role perceptions amid AI adoption, balancing established journalistic values against changing industry realities.

  • Workforce

• Research

  Readability of U.S. food handler training materials: a natural language processing analysis of worker study guides and the federal Food Code

  Morris Brako, Anirudh R. Naig

  Frontiers in Public Health · 2026-08-27

  This study used a natural language processing pipeline to assess the reading difficulty of 14 food handler training study guides across 11 U.S. jurisdictions, comparing them to the 2022 FDA Food Code. Results showed that 13 of 14 guides exceeded the recommended sixth-grade Flesch–Kincaid readability target, and the FDA Food Code was more demanding than every guide. Washington's Food Worker Manual was the sole exception, demonstrating that accessible materials are achievable in this domain. The findings raise concerns about whether certification training materials are actually usable by the workers required to study them, and highlight audit gaps where content is delegated to private vendors.

  • Certifications
  • Workforce

• Research

  Trusting the AI, Crafting the Show: Algorithmic Dependence and Human–AI Collaboration in Variety-Show Post-Production Across Chinese OTT Platforms

  Fang Wang, Yazhou Huang, Qianxun Jiang et al.

  International Journal of Human-Computer Interaction · 2026-08-27

  This study examines how post-production workers on Chinese OTT variety shows adopt and come to depend on AI tools, using a mixed-methods design combining survey data from 443 respondents (analyzed via PLS-SEM) with 20 qualitative interviews. Findings show that trust in AI—shaped by performance expectancy, effort expectancy, interpretability, controllability, and verifiability—drives behavioral intention and use, which in turn leads to algorithmic dependence as a post-adoption outcome. AI anxiety moderates how workers evaluate AI systems, and the study finds that ease of use can paradoxically trigger resistance when workers feel it threatens their professional value. The research informs how AI tools can be deployed in creative media production in ways that remain compatible with human judgment, occupational sustainability, and professional responsibility.

  • Workforce
  • Enterprise

• Research

  Expert-Guided Visual Correction for Characterizing Diagnostic Performance and Error Patterns of Multimodal Large Language Models Using Periodontal In-Service Examination Images

  Prita Abhay Dhaimade, Robin Henderson

  medRxiv · 2026-08-27

  This study benchmarked six multimodal large language models (MLLMs) on 50 image-based multiple-choice questions from the American Academy of Periodontology In-Service Examination, using a two-phase design where models first answered independently and then re-answered incorrect questions after receiving expert-validated image descriptions. By distinguishing perceptual errors (faulty visual feature extraction) from cognitive errors (flawed reasoning despite accurate perception), the authors found that aggregate accuracy scores mask mechanistically distinct failure modes—some errors resolved with corrected visual input while others persisted regardless. The framework demonstrates that safe clinical deployment and model improvement require separately characterizing perceptual versus reasoning failures, with direct implications for AI-augmented dental education and competency assessment.

  • Quality assurance
  • Certifications

• Research

  Strategic competitiveness between the EU and the United States in artificial intelligence development – human capital, legal frameworks and data

  Robert Lizak

  Kortowski Przegląd Prawniczy · 2026-08-27

  This article compares the strategic competitiveness of the EU and the United States in AI development across three dimensions: human capital availability, regulatory environment, and data access. Grounded in Draghi's Report on European economic competitiveness and key legal instruments including the EU AI Act, GDPR, and US federal laws, the analysis finds that the EU model prioritizes fundamental rights protection and risk management, while the US approach emphasizes innovation, public-private collaboration, and broad data access. The authors conclude that the US's more flexible legal environment accelerates AI commercialization and competitive advantage, and they offer recommendations for the EU to strengthen human capital, simplify regulation, and improve data availability.

  • AI policy

• Research

  Artificial intelligence in social work practice for children and youth: A scoping review

  Zhao Lifen, Ying Feng, Yong‐Xin Ruan et al.

  International Journal of Social Welfare · 2026-08-27

  This scoping review synthesizes evidence on AI applications in social work for children and youth, drawing on 17 studies across eight databases. Only five studies provided empirical evaluations—focused mainly on identifying service users—and their results were mixed, while the remaining studies were theoretical. Across all evidence tiers, AI served a supporting role with social workers retaining final decision-making authority, and key challenges included algorithmic bias, privacy risks, contextual complexity, and unclear accountability. The review highlights the limited and uneven state of evidence, urging responsible AI integration and a clear future research agenda in this domain.

  • Workforce
  • AI policy

• Research

  Non‐inferiority analysis of a fetal heart rate artificial intelligence algorithm to registered nurse assessment

  Rohit Pardasani, Renee Vitullo, Sara Harris et al.

  Pregnancy · 2026-08-27

  This retrospective study evaluated whether an AI algorithm using deep learning and rule-based techniques could assess fetal heart rate and uterine activity from cardiotocography (CTG) tracings at a level comparable to registered nurses (RNs). Across data from five U.S. hospitals (2013–2024), the algorithm met a pre-specified 15% non-inferiority threshold versus RN readers on most primary endpoints—including baseline fetal heart rate, accelerations, decelerations, and contraction frequency—though it fell short for contraction duration under the strict criterion. The findings suggest AI could serve as a consistent, objective decision-support tool in obstetric care, potentially reducing the inter- and intra-observer variability that currently affects clinician interpretation of CTG tracings.

  • Quality assurance
  • Workforce

• Research

  Developing a Neuroimaging Training Framework and Certificate Program

  Dianne Patterson, Thomas Hicks, Aneta Kielar

  arXiv · 2026-08-27

  This paper describes the design and evolution of an asynchronous neuroimaging training program at the University of Arizona that teaches computational skills—including BIDS, Unix, HPC, JupyterLab, MATLAB, Bash, and containerization—to a diverse cohort of students, staff, and faculty. A key feature is its flexible, à la carte structure that allows learners to pursue individual lessons or a full certificate path. The program also addresses AI's disruption to traditional assessment by replacing multiple-choice questions with interactive matching and structured generative AI workflows while retaining hands-on practicals. The case study offers a replicable template for institutions seeking to deliver targeted, on-demand computational training.

  • Certifications
  • Workforce

• Research

  Social support and technological displacement anxiety among ride-hailing drivers: The mediating roles of self-efficacy and qualitative job insecurity

  WU Xin-hui, Yan Dai

  Acta Psychologica · 2026-08-27

  This study of 447 ride-hailing drivers finds that higher social support is significantly associated with lower anxiety about being displaced by autonomous driving technology. Using bootstrap mediation analysis, the researchers identify three psychological pathways through which social support reduces displacement anxiety: via increased self-efficacy, via reduced qualitative job insecurity, and via a sequential chain where social support boosts self-efficacy, which in turn lowers job insecurity, which then reduces anxiety. The authors suggest these findings could inform mental-health and occupational-support interventions for gig workers facing AI-driven disruption.

  • Workforce

← Back to the dashboard