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
Strategizing safety for women in the gig economy: a scoping review of policy research
Deeksha Bajpai Tewari, Upma Gautam, Ankita Mishra et al.
Frontiers in Organizational Psychology · 2026-07-27
This scoping review synthesizes 16 studies on the safety concerns facing women in the gig economy and finds that abusive workplace behavior, fear of nighttime operations, customer anonymity, and lack of client verification are prominent barriers to female participation. The review reveals that formal protective structures are largely absent, with only informal strategies in place to address harassment. The authors call for changes to existing legal and policy frameworks to better protect female gig workers from sexual harassment and enable broader workforce participation.
- Workforce
- AI policy
Research
TRUST IN AI, BEHAVIOURAL NUDGING, AND INVESTOR OVERSIGHT IN RETAIL INVESTOR DECISION-MAKING: EVIDENCE FROM INDIA
Anushree Adithya Balike
EPRA International Journal of Economic and Business Review · 2026-07-27
This study examines how AI-powered investment platforms shape retail investor decision-making in India, focusing on three drivers: AI recommendation quality, behavioural nudging, and perceived regulatory oversight. Using PLS-SEM on 276 Indian retail investors, the authors find that behavioural nudging is the strongest predictor of cognitive bias activation, and that trust and cognitive bias each partially mediate pathways to investment behaviour. Crucially, perceived regulatory oversight increased trust in AI but did not reduce bias susceptibility, suggesting that disclosure-based regulatory frameworks build investor confidence without actually improving decision quality — a finding with direct implications for how financial regulators design oversight frameworks for AI-driven investment platforms.
- AI policy
- Enterprise
Research
Malaysia's Digital Governance Framework in the Era of Emerging Technologies
Nuurrianti Jalli
arXiv · 2026-07-27
This chapter analyzes Malaysia's evolving digital governance framework, examining how legislation including the Communications and Multimedia Act 1998, its 2025 amendments, the Online Safety Act 2025, and the National Guidelines on AI Governance and Ethics 2024 are reshaping regulation of digital platforms and AI. The authors highlight tensions between stated goals—curbing misinformation, improving cybersecurity, and promoting ethical AI—and concerns about press freedom, civic engagement, and selective enforcement against journalists and activists. The chapter traces continuity from colonial-era laws adapted to digital expression, while noting the expansion of formal AI governance institutions and platform licensing requirements. The findings underscore that Malaysia's digital future hinges on whether its regulatory model prioritizes transparency and innovation or consolidates state control over digital spaces.
- AI policy
Research
Who Gets Named: Citation Type Predicts Individual Naming by Grounded Language Models, and a Roster Instrument Captures 0.5% of It
Dmitrij Żatuchin
arXiv · 2026-07-26
This study examines how often AI language models name individual professionals (rather than just firms) in response to buyer-intent queries, using 2,400 grounded API calls across four models, four European markets, and five languages. Results show models named an individual in 25.8% of responses, with large variation by category (real estate 35.4%, insurance 9.1%) and by model (Grok 38.0%, Gemini 9.3%). Citation type—not citation volume—predicts individual naming, with naming responses more likely to cite the individual's own site or category portals. Critically, a 939-person roster matched only 0.47% of name-shaped mentions, meaning roster-based measurement captures a small and unrepresentative slice of actual individual AI visibility.
- Enterprise
- AI policy
Research
Visible to the Court: How AI Is (and Isn't) Litigated in U.S. Federal Court Opinions
Julie Yu, Rock Yuren Pang, Jevan Hutson et al.
arXiv (Cornell University) · 2026-07-26
This paper presents a systematic empirical review of 559 U.S. federal court opinions in which AI plays a role in the parties' contentions, identifying seven recurring dispute areas, six categories of AI technologies, and four types of common litigants. The authors find that courts predominantly apply pre-existing legal doctrines rather than creating new AI-specific law, producing what they call 'piecemeal' AI governance where outcomes are shaped more by which harms are cognizable under existing statutes than by where AI has actually caused harm. A comparison with the AI Incident Database reveals substantial gaps between documented and litigated harms, suggesting courts capture only part of the AI risk landscape. The findings matter for policy because unresolved AI harms and inconsistent coverage across legal doctrines point to significant gaps in the current U.S. federal regulatory and judicial framework for AI.
- AI policy
Research
AI Strategy: How to Choose What AI Product to Implement
Foster Provost, Panos Ipeirotis
arXiv (Cornell University) · 2026-07-26
This paper introduces 'expected ROI' (eROI), a decision framework to help firms choose which AI products to build by decomposing each candidate project into three separately rated components: Value if Successful, Likelihood of Success, and Investment Required. Using Compass, a residential real-estate brokerage, as the primary case study, the authors show how the framework distinguishes strong bets (a Likely-to-Sell recommendation tool that generated nine figures in annual gross commission revenue) from weak ones (a Time-on-Market pricing tool that was correctly shelved), even when simple ROI estimates could not tell them apart. The framework also addresses portfolio construction—encouraging firms to fund multiple ranked bets rather than only the top project—and breaks the catch-22 where teams cannot estimate ROI without building, and cannot build without justifying ROI. Coarse business-level ratings of the three components are presented as sufficient to guide sound AI investment decisions under uncertainty.
- Enterprise
- AI policy
Research
The Illusion of Secure LLM Code: Closing the Security Gap via Iterative Reprompting
Ishpuneet Singh, Shreyas Mahajan, Gurjot Singh et al.
arXiv (Cornell University) · 2026-07-26
This paper evaluates the security of authentication code generated by five AI coding assistants using a combined static analysis and dynamic penetration testing framework mapped to NIST SP 800-63B guidelines. Testing four prompting strategies — Basic, Secure, NIST-Based, and Reprompting — the study finds that functional or generically secure prompts consistently omit critical protections such as brute-force resistance, session management, and robust password handling. Even explicit single-shot NIST context improves but does not fully resolve compliance gaps; only iterative reprompting, which forces models into a self-auditing loop, achieves a comprehensive defense-in-depth security architecture. The authors conclude that AI coding assistants do not produce secure-by-default applications, and enterprise deployments must adopt continuous, standards-driven verification pipelines rather than single-shot prompt engineering.
- Enterprise
- Quality assurance
Research
Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing
Kevin Lee, Benjamin Letham, Zhiyuan Jerry Lin et al.
arXiv · 2026-07-26
This paper presents an offline-to-online workflow for ad creative optimization that pairs generative AI models with predictive ranking and adaptive online experiments. In the offline phase, a predictive model trained on historical A/B test data guides and ranks AI-generated creative variants; a final slate is then tested in an adaptive online experiment. Across three field experiments—including a 50-arm study—the best AI-generated creatives achieved 45.1%, 46.7%, and 36.2% higher engagement than the best human-authored creatives. The findings suggest a practical design principle: use predictive models to guide generation of a candidate slate and adaptive experiments to efficiently identify top performers, even when offline prediction alone is too noisy to pinpoint the best creative.
- Enterprise
Research
Where Is the Cost of Third-Party API Routers in Agentic Software Development?
Donghao Fu, Jingxin Li, Xue Jiang et al.
arXiv · 2026-07-26
This paper empirically investigates security vulnerabilities introduced by third-party API routers in agentic software development workflows. The researchers developed SIDEL, a framework for injecting and evaluating router-side attacks at four subtlety levels, and tested four coding agents on 400 samples. Key findings show that without mitigations, all evaluated agents achieved a 0% defense success rate against router-side injections, meaning malicious modifications to LLM responses could alter repository-level code actions undetected. The results highlight a critical control gap in current agentic coding pipelines and motivate the need for provider-side output-integrity guarantees.
- Quality assurance
- Enterprise
Research
Language Shapes Instruction Hierarchy Compliance in Multilingual LLMs
Jiwon Moon, Yerin Hwang, Kyomin Jung
arXiv · 2026-07-26
This paper introduces XIH-Bench, a benchmark for evaluating instruction hierarchy (IH) compliance in multilingual large language models across six languages, four domains, and three IH settings. The study finds that a model's ability to correctly prioritize higher-priority instructions over lower-priority ones varies significantly by language, with a language that strengthens compliance in a high-priority position becoming disruptive in a lower-priority position. Cross-language conflicts produce higher compliance than same-language conflicts—a phenomenon the authors call the Language Boundary Effect. These findings reveal multilingual reliability and security risks, as language specialization can make lower-priority instructions in model-favored languages harder to override, undermining safe and controllable deployment.
- AI policy
- Quality assurance
Research
Private Again: AI Agents Restore Anonymity---Foreclosing Discrimination and Its Proof
Anirban Mukherjee, Hannah Hanwen Chang
arXiv (Cornell University) · 2026-07-26
This paper examines how AI agents that transact online on behalf of human users—handling browsing, payments, and reviews without revealing identity—can eliminate the personal data inputs (identity, purchase history, location, demographic proxies) that fuel algorithmic discrimination. However, the same anonymity that blocks discriminatory targeting also destroys the evidentiary basis needed to prove discrimination, since disparate-treatment claims require comparators, disparate-impact claims require protected-class baselines, and post-Iqbal pleading requires specific factual allegations that anonymous transactions never generate. The authors highlight an asymmetric burden: those most vulnerable to discrimination are least likely to afford these privacy-preserving agents, and least able to prove harm when it occurs. The paper argues that law must shift focus from detecting algorithmic discrimination to governing agent-mediated anonymity as civil rights infrastructure, including questions of access, abuse regulation, and whether retailers may refuse agent transactions.
- AI policy
Research
Mission-Level Runtime Assurance for LLM-Assisted ISR Swarms over a Verification-Aware Fabric
Nikolaos Kekatos, Stylianos Basagiannis, Panagiotis Katsaros et al.
arXiv (Cornell University) · 2026-07-26
This paper addresses a critical safety gap in swarms of LLM-assisted autonomous robots used for intelligence, surveillance, and reconnaissance (ISR): individually compliant robots can collectively execute prohibited missions by splitting tasks across platforms to evade per-platform guardrails. The authors propose a three-tier (platform/squad/mission) compositional runtime-verification framework that decomposes mission policies into per-agent and cross-agent aspects and aggregates verdicts over a verification-aware messaging fabric using an evidence-aware algebra that tracks provenance. In simulated ISR experiments, the framework successfully detects an indirect prompt injection that causes real LLM planners to split a prohibited collection task across four platforms — a violation invisible to every per-platform monitor — while a baseline central monitor emits silent false all-clears under injected network faults. The work directly bears on quality assurance and certification of AI-driven autonomous systems, demonstrating that mission-level compositional verification is necessary to reliably catch emergent swarm-level policy violations.
- Quality assurance
- Certifications
Research
Auditing Alignment Controllability in LLMs via Political Axes
Bartol Bućan, Nikola Sočec, Sarah Isufi et al.
arXiv · 2026-07-26
This paper audits how easily seven leading LLMs (GPT-5, Claude, Grok, Gemini, DeepSeek, Kimi, and Qwen) can be steered toward different political positions via system prompts, testing 12 ideological personas across 70 Political Compass items for 63,700 total responses. The study finds that contextual framing (i.e., prompt-based persona instructions) explains roughly 88–93% of variance in model responses on economic and societal axes, while model identity accounts for under 3%, meaning responses are highly instruction-adjustable regardless of which model is used. Models differ in how far and how symmetrically they can be steered, with some saturating under extreme framings, and prior conflicting audit results are explained by non-centered baselines rather than differential compliance. The authors argue that political audits of LLMs must report not just a resting position but also steerability metrics such as dispersion, symmetry, saturation, and refusal floors, releasing prompts, benchmark data, and code to support this.
- AI policy
- Quality assurance
Research
A Regress Argument Concerning Autonomous Self-Correction in AI Agents
Julian Lee-Sursin, Hong Joo Ryoo
PhilSci-Archive (University of Pittsburgh) · 2026-07-26
This paper argues that the concept of 'self-healing' AI—systems that autonomously detect and correct their own failures—is fundamentally unrealizable as commonly described. Any self-correcting agent must evaluate its own performance against a benchmark, but that evaluation process can itself fail, generating an infinite regress; in practice, every implemented system relies on some external anchor or threshold (a 'stopper') that the system itself cannot certify. The authors examine where these stoppers appear in prominent agent architectures and show that existing benchmarks systematically obscure them. They conclude that claims of autonomous self-correction should be replaced by explicit, stopper-relative descriptions, imposing a constraint on both the design and evaluation of agentic AI.
- Quality assurance
- AI policy
Research
Exploring artificial intelligence adoption in digital marketing using snowball sampling: A qualitative study of small retail businesses in Cape Town
Steven Lububu, Monique Harris, Michael Twum-Darko
International Journal of Business Ecosystem and Strategy (2687-2293) · 2026-07-26
This qualitative study examines how small retail businesses in Cape Town adopt AI for digital marketing, drawing on 18 semi-structured interviews analyzed through a combined TOE–UTAUT theoretical framework. Thematic analysis reveals that competitive pressure, customer expectations, and perceived performance benefits drive adoption, while skills gaps, cost concerns, and data challenges constrain it. The study finds that platform-based AI tools can lower technological barriers but simultaneously expose organizational and capability-related weaknesses in emerging market contexts. The findings offer practical guidance for policymakers and practitioners seeking to support digital transformation in small business ecosystems.
- Enterprise
- AI policy
Research
Navigating Boundaries of AI Use in Physical Education Journal Research Publications
Yucen Li, Chenhao Wu, Yanlei Su et al.
Quest · 2026-07-26
This study examines AI usage policies across ten peer-reviewed physical education journals, finding that policy adoption remains limited and driven top-down by publishers rather than journals themselves. While AI authorship is universally banned, rules around AI-generated content vary, and a notable asymmetry exists between strict confidentiality requirements for reviewers versus weaker data privacy standards for authors. The authors argue that current governance relies on generic templates and call for a human-centered framework with symmetric data protection and bias screening to preserve research integrity.
- AI policy
- Quality assurance
Research
Engineering Determinants of Artificial Intelligence Adoption in Indian Healthcare Systems: A Regression-Validated Institutional Readiness Model
Prateek Singh, Sudipta Das, Justin Babu et al.
Al-Nahrain Journal for Engineering Sciences · 2026-07-26
This study examines the engineering predictors of AI adoption readiness across Indian hospitals using a structured survey of 120 healthcare workers and 20 expert interviews, analyzed via correlation and multiple linear regression. The regression model explained 61% of variance (R²=0.61, p<0.001), with Perceived Usefulness, Infrastructure Availability, and AI Training Exposure as significant positive predictors, while Ethical Concern Scores negatively predicted readiness. Qualitative findings reinforced quantitative results, identifying interoperability gaps, lack of computational training, and data governance weaknesses as key barriers. The paper proposes a validated institutional readiness model for AI implementation in Low- and Middle-Income Country hospital settings, with implications for workforce training, infrastructure investment, and ethical governance frameworks.
- Workforce
- AI policy
Research
Artificial Intelligence Adoption, Implementation Barriers, and Business Impact in Small and Medium-Sized Enterprises: A Systematic Literature Review
Edy Suandi Hamid, Bhenu Artha
Archives of Business Research · 2026-07-26
This systematic literature review synthesizes 50 peer-reviewed articles (2019–2025) on AI adoption in small and medium-sized enterprises (SMEs), drawing on Scopus, Web of Science, and IEEE Xplore. It finds that SMEs primarily deploy machine learning, natural language processing, and chatbot technologies to improve customer service, marketing automation, and operational efficiency, with adoption linked to measurable gains in cost reduction, customer retention, and innovation performance. Key barriers include financial constraints, skills shortages, and inadequate digital infrastructure, while top management support and staff capability are critical enablers. The review calls for tailored implementation strategies, capacity-building initiatives, and supportive policy frameworks to enable inclusive AI integration across SMEs globally.
- Enterprise
- Workforce
- AI policy
Research
Bias-Audited Resume Screening with Calibrated Matching Scores, Selective Review, and Evidence-Grounded LLM Explanation Cards
Lucas Cui
International Journal of Instruction, Technology & Social Sciences. · 2026-07-26
This study evaluates an automated resume screening pipeline on a dataset of 4,817 resumes spanning 46 technical roles, finding that a temperature-scaled support vector machine achieved modest accuracy (24.9%) and top-3 accuracy of 39.8% when leakage-prone fields were removed. A leakage audit showed that including direct role text inflated accuracy to 100%, underscoring how easily such systems can produce misleading performance estimates. The system incorporates selective review, calibrated uncertainty, and LLM-generated explanation cards grounded in candidate evidence, with all 46 role summaries verified to preserve safety constraints and exact evidence spans. The authors conclude the pipeline is appropriate only for human-supervised role triage, not autonomous hiring decisions.
- Workforce
- Quality assurance
Research
AI as Cognitive Infrastructure in Norwegian Legal Access: A Multi‑Layer Framework for Responsibility, Regulation and Justice Gap Reduction
Oleg Zmiievskyi
Zenodo (CERN European Organization for Nuclear Research) · 2026-07-26
This preprint argues that AI can serve as a 'cognitive infrastructure' to reduce Norway's documented justice gap by providing scalable legal information to the roughly two-thirds of Norwegian households excluded from the legal-aid system after 2025 reforms. The framework distinguishes three operational layers of AI-assisted legal work—information, interpretation, and process—finding that AI democratizes access to legal information but not interpretive competence, with lawyers retaining professional responsibility for assessing AI outputs. The paper identifies a regulatory vacuum around AI legal assistance but contends it does not constitute a prohibition, and proposes an eight-point policy framework including transparency standards, a national AI legal-aid portal, and empirical evaluation programs. The work positions AI as a bridge to access rather than a replacement for lawyers.
- AI policy
- Workforce
Research
Effect of Robotic Process Automation (RPA) on the growth of listed food and beverage firms in Nigeria
Juliet Adaku NWANORO
Radiant Journal of Business & Sustainability · 2026-07-26
This study analyzed how Robotic Process Automation (RPA) affects revenue growth among twelve listed Nigerian food and beverage firms from 2015 to 2024, using panel data and Estimated Generalized Least Squares regression. Results showed that RPA usage had a statistically significant negative effect on revenue growth, while RPA cost and investment intensity showed no significant effect, suggesting automation investments have not yet produced measurable financial returns. The authors attribute this to implementation costs, learning curves, and low maturity of RPA adoption, and recommend firms integrate automation strategically into core processes and invest in employee digital skills for long-term gains. The findings are relevant to enterprises considering automation and to the workforce implications of digital transformation in emerging markets.
- Enterprise
- Workforce
Research
Governing the Ungoverned: Emotional AI, Nigerian Communication Policy, and the Case for a National Regulatory Framework
Ighodalo Uyi Ebhodaghe, Jide Johnson
Zenodo (CERN European Organization for Nuclear Research) · 2026-07-26
This paper identifies a regulatory double failure in Nigeria's communication governance architecture: no Nigerian policy instrument defines Emotional Artificial Intelligence (EAI), and no single regulatory body holds jurisdiction over its misuse in political communication. Using critical policy analysis, comparative review of the EU AI Act, and African AI ethics scholarship, the authors find that institutions such as the NBC, NCC, NITDA, and NCAIR were designed before generative and affective AI emerged and remain inadequate for governing EAI's use in electoral contexts. In response, the article proposes a four-pillar Nigeria-Specific EAI Ethics Framework covering Awareness and Literacy, Emotional Responsibility, Governance and Guidelines, and Trust and Transparency, intended as a decolonially-grounded governance starting point ahead of Nigeria's 2027 general elections. The paper offers the first systematic mapping of Nigeria's EAI regulatory gap and advances an original governance framework calibrated to Nigeria's institutional realities.
- AI policy
Research
AI as Cognitive Infrastructure in Norwegian Legal Access: A Multi‑Layer Framework for Responsibility, Regulation and Justice Gap Reduction
Oleg Zmiievskyi
Zenodo (CERN European Organization for Nuclear Research) · 2026-07-26
This preprint presents a multi-layer analytical framework arguing that AI can serve as scalable 'cognitive infrastructure' to reduce Norway's documented justice gap, where roughly two-thirds of households fall outside the legal-aid system following a 2025 reform. The framework distinguishes three operational layers of AI-assisted legal work—information, interpretation, and process—concluding that AI democratizes access to legal information but not interpretive competence, with lawyers retaining responsibility for their assessments of AI output. The paper documents five structural findings and proposes an eight-point policy framework including transparency standards, a national AI legal-aid portal, and empirical evaluation programmes, explicitly positioning AI as a complement to, not a replacement for, lawyers.
- AI policy
- Workforce
Research
Governing the Ungoverned: Emotional AI, Nigerian Communication Policy, and the Case for a National Regulatory Framework
Ighodalo Uyi Ebhodaghe, Jide Johnson
Zenodo (CERN European Organization for Nuclear Research) · 2026-07-26
This article identifies a critical regulatory gap in Nigeria's governance of Emotional Artificial Intelligence (EAI)—systems that detect, simulate, and exploit human emotions for political persuasion—finding that Nigeria's existing communication policy bodies (NBC, NCC, NITDA, NCAIR) collectively constitute a 'regulatory double failure': no Nigerian instrument defines EAI, and no single institution holds jurisdiction over its misuse in political contexts. Using critical policy analysis and comparative regulatory review (including the EU AI Act), the authors propose a four-pillar Nigeria-Specific EAI Ethics Framework covering Awareness and Literacy, Emotional Responsibility, Governance and Guidelines, and Trust and Transparency. The framework is explicitly calibrated to Nigeria's institutional realities and positioned as a practical governance starting point ahead of the 2027 general elections. The paper contributes the first systematic mapping of Nigeria's EAI regulatory gap and advances a context-sensitive, decolonially-grounded governance model for African AI regulation.
- AI policy
Research
The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise
Nolan Lovett
Human Resource Development Review · 2026-07-26
This conceptual paper introduces the 'Cognitive Commons' framework to argue that rational AI adoption decisions by individual organizations can collectively deplete the shared pool of professional expertise that professions need to renew themselves. It distinguishes between 'Internalized Mastery' (deep knowledge built through sustained practice) and 'Distributed Mastery' (managing human-AI systems), and introduces the 'Validation Tether' concept: effective AI oversight depends on the very expertise that AI adoption may erode. Drawing on early labor market and clinical evidence from highly AI-exposed sectors, the paper identifies five factors shaping occupational vulnerability and proposes governance arrangements spanning organizations, professional associations, and policy to steward expertise as a collective resource rather than an organizational optimization problem.
- Workforce
- AI policy