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
Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models
Rodion Krjutškov, Eduard Barbu, Nikos Sakkas et al.
arXiv · 2026-09-10
This paper introduces the Explainability Assistant, an open-source conversational AI system that helps facility managers and building operators interpret complex machine learning models used for energy consumption forecasting. By leveraging the function-calling capabilities of modern Large Language Models, the system achieves 94% intent-parsing accuracy — up from 76.8% in prior approaches like TalkToModel — without requiring task-specific fine-tuning. An evaluation with energy domain specialists found improved usability and consistent task accuracy compared to traditional XAI dashboards, with all experts unanimously preferring the conversational interface for practical use. The work matters because it lowers the technical barrier for non-expert users to interrogate and trust AI-driven energy models in real operational settings.
- Enterprise
- Workforce
Research
Target leakage, not model class, explains reported accuracy in survey-based cardiovascular screening: a leakage-tiered audit of glass-box and tabular foundation models
Raad Bin Tareaf, Murad Al-Rajab, Samia Loucif et al.
arXiv · 2026-09-10
This paper audits ten machine learning classifiers—including linear models, tree ensembles, neural networks, glass-box models, and tabular foundation models—for predicting prevalent myocardial infarction using large national health survey data (over 440,000 respondents). The central finding is that the widely reported ~0.89 AUROC in cardiovascular screening literature is largely driven by target leakage from post-diagnostic features, not genuine model performance: removing just two such features collapses all models into a narrow 0.0045-wide AUROC band. The glass-box explainable boosting machine matched every alternative in discrimination while being roughly 104 times faster than the best foundation model, and its transparency directly enabled fairness repairs and uncertainty calibration. The authors conclude that evaluation methodology and feature set construction—not model capacity—are the binding constraints in this domain.
- Quality assurance
- AI policy
Research
SpecGuard: Inference-Time Backdoor Detection For Free
Rui Wen, Ahmed Salem, Andrew Paverd et al.
arXiv · 2026-09-10
SpecGuard proposes a zero-cost inference-time backdoor detection method for large language models by repurposing speculative decoding, a technique already used to speed up inference. The key insight is that when a backdoor trigger activates a fine-tuned target model, a clean draft model fails to predict the resulting behavioral shift, causing a measurable change in draft-token acceptance rate. The authors formalize when this signal is detectable and show that any attacker who suppresses it must also weaken the backdoor itself. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior—including stealthy cases that bypass input-level filters—without requiring extra model computation.
- Quality assurance
Research
The widening evaluation gap in medical large language model research 2023 to 2026
Raad Bin Tareaf, Murad Al-Rajab, Samia Loucif
arXiv (Cornell University) · 2026-09-10
This study examined 11,628 PubMed records on medical large language model (LLM) research published from January 2023 to June 2026, finding that the field has grown 45-fold but that clinical evaluation is struggling to keep up with rapid model development. Only 2.5% of studies used a randomised, controlled, or prospective design, and the 'evaluation lag'—the gap between a model's release and the publication of a study evaluating it—widened from 1.33 to 6.08 quarters over that period. Critically, randomised trials evaluated models a median 4.6 quarters older than other study designs, and 62% of randomised trials evaluated a discontinued model family, revealing a structural tension between research rigour and currency. The authors attribute this gap to model selection choices rather than research timelines, raising serious concerns about the relevance of high-quality clinical evidence by the time it is published.
- Quality assurance
- AI policy
Research
SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control
Suwan Wu, Yumeng Lin, Pengcheng Yuan et al.
arXiv · 2026-09-10
SIRF (Spec-Internalized Risk Foundation Model) is a foundation model designed for industrial content risk control that embeds a platform's complex policies directly into model weights through continued pretraining, eliminating the need for additional human annotation. Using techniques called EntiGraph, MAGA rewriting, and account-level chain-of-thought synthesis, the 8B-parameter model achieves 71.3% Black Recall at 95% precision—a 15.1 percentage point improvement over a same-architecture baseline—while operating under second-level latency constraints. When deployed as a tree-model adjudication layer, it recovers 20% more mis-penalized samples and reduces mis-penalization by roughly 70% in a freezing scenario. The work demonstrates that internalizing policy rules into model weights, rather than injecting them at inference time, is a practical path to high-precision, low-latency automated content moderation at industrial scale.
- Enterprise
- Quality assurance
Research
Geospatial AI, Dataverse Metadata, and the Study of Place-Based Government
Danny EBanks, Devika Jain
arXiv (Cornell University) · 2026-09-10
This paper constructs a knowledge graph from Harvard Dataverse's public metadata, organizing 102,650 datasets into a 215,985-node network to make geospatial and policy-relevant research more discoverable. The authors find that 42.9 percent of datasets carry at least one geospatial field, and a keyword search identifies 7,654 datasets (17.4 percent) as directly policy-relevant, with elections and legislatures forming the largest cluster. The paper demonstrates how AI tools—including community language models, stance detection with geographic aggregation, and partisan language bridging—can link political discourse to place. It argues that the graph provides a concrete testbed for AI-driven metadata enrichment and entity resolution, while noting a coverage skew toward American, city-level data.
- AI policy
Research
Who Bears the Risk When Generative AI Enters Transport? A Distributional Sociotechnical Audit of Algorithmic Equity, Synthetic-Data Validity, and Public Trust
Amir Rafe, Subasish Das
arXiv (Cornell University) · 2026-09-10
This paper develops a Distributional Sociotechnical Audit (DSA) framework to measure how generative AI systems create unequal risks across different population groups in transportation contexts, including traveler advisories, synthetic crash-record generation, and policy decision support. Analyzing 5,760 persona-controlled queries across 12 demographic cues, three synthetic crash-record generators, and public attitude data from 4,538 respondents, the study finds that congestion-pricing AI advice shows the highest demographic disparity (mean EDI = 1.96), that CART-based synthetic crash records fail all conditional validity tests, and that existing categorical AI approval frameworks flip their tier assignments 75% of the time under minor weight perturbations. The authors argue that continuous risk indices with sensitivity reporting provide a more defensible basis for transport AI governance than the categorical approval tiers currently in use.
- AI policy
- Quality assurance
Research
Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems
Aleksandra Urman, Elsa Lichtenegger, Salima Jaoua et al.
arXiv · 2026-09-10
This paper investigates how commercial text-to-image systems (DALL-E-3, Imagen-4, GPT-Image-1.5) silently revise user prompts before generating images, and whether that revision step introduces cultural bias. Using WORLDVIEW, a new multilingual benchmark of 8,960 prompts across 15 languages and 31 language-context pairings, the authors find that non-Western and non-Anglophone cultural contexts are marked far more heavily than a US/English baseline, flattened into narrow vocabularies, and reduced to stereotypes. By comparing outputs from original versus revised prompts on models without a revision layer, the study identifies the prompt-revision layer itself as a previously undocumented, causal source of cultural stereotyping. The findings argue that bias audits must examine the full deployed system—not just the generative model—to accurately locate and fix the problem.
- AI policy
- Quality assurance
Research
ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps
Jacopo Dardini, Roberta Calegari
arXiv · 2026-09-10
ActMap introduces a white-box uncertainty quantification method for large language models that works from a single generation pass, requiring no multiple samples. It compresses the full hidden-state trajectory across every layer and every generated token into a compact fixed-size tensor (96 KiB), which a lightweight classifier then reads to estimate whether a model's answer is correct. Evaluated on short-answer QA, math, and summarization factuality with three 7–8B instruction-tuned models, ActMap consistently outperforms sampling, token-probability, attention, and embedding baselines, and matches a detector trained on tensors 67× larger at essentially the same mean AUROC with lower calibration error on ten of twelve pairs. The resulting correctness score enables abstention, routing, and selective verification, making it a practical tool for scalable oversight of deployed language models.
- Quality assurance
- Enterprise
Research
Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms
Pushpdeep Singh, Sayeh Jarollahi, Ayan Majumdar et al.
arXiv (Cornell University) · 2026-09-10
This paper conducts the first large-scale audit of Bluesky's default content moderation system (BMS) by analyzing 10.6 million moderation labels from 2025, made accessible through the platform's transparent, decentralized architecture. The study finds that BMS operates as a human-AI collaborative system—automatically labeling sexual and graphic content within seconds while routing nuanced or high-stakes cases to human reviewers over hours or days. The system achieves high precision (0.837) but low recall (0.222), with human annotators identifying 4.5× more harmful content than the system in a random sample, indicating substantial under-detection. The findings provide a data-driven baseline for improving transparency and effectiveness in automated content moderation.
- AI policy
- Quality assurance
Research
NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment
Guoqiang Zhang, Kexin Tan, Ming Zhang et al.
arXiv · 2026-09-10
NovGauge is a new benchmark designed to evaluate how well large language models (LLMs) can assess the novelty of research papers across three fine-grained dimensions: task, problem, and method. The benchmark contains 619 paper pairs and 50 multi-paper sets labeled by human experts, and uses a cascading diagnostic pipeline to check not just whether a model's judgment is correct but whether its supporting evidence is logically faithful. Evaluation of 18 LLMs reveals hallucination rates from 0% to 39% and shows that over 70% of non-hallucinated correct-positive judgments cite evidence that fails to logically support the stated reason; even the best model (GPT-5.5) achieves only 43–72% Verified F1 after faithfulness verification. These findings indicate that current LLMs are not yet reliable for scientific novelty assessment, with important implications for AI-assisted peer review at major conferences.
- Quality assurance
Research
(Whose defaults?) Is artificial intelligence reorienting archaeological methods?
Lorenzo Cardarelli, Roberto Ragno
arXiv · 2026-09-10
This paper investigates whether large language models (LLMs) are narrowing the range of computational methods used in archaeology. Analyzing roughly 119,000 archaeology abstracts from 2010–2025, the authors found a small but credible shift in method use after 2023, though overall methodological diversity actually increased rather than declined. However, a controlled experiment showed that when LLMs were asked to recommend methods for 28 archaeological research problems, their suggestions were far less diverse than the published literature and skewed toward methods popular before 2023, raising concerns that LLM involvement in research planning could push methodological choice toward convergence over time.
- Workforce
- AI policy
Research
Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment
Jiayu Huang, Zichen Tang, Qianhui Ling et al.
arXiv · 2026-09-10
LogiMed-RoB is a new benchmark that tests whether large language models can follow the hierarchical logical reasoning used by medical experts when assessing risk of bias in clinical trials, using 860 randomized controlled trials and 14,820 queries drawn from the Cochrane Risk of Bias 2.0 framework. Experiments on 10 state-of-the-art LLMs reveal a severe 'Error Compounding Effect': the best model achieves 98.88% consistency on individual atomic steps, yet end-to-end logical consistency collapses to 45.13%, with some open-weight models falling to nearly 0%. The study also identifies a systematic evidence-reasoning gap in which models fail to reach correct conclusions from high-quality retrieved evidence in 18.63–40.05% of cases, and exhibit Blind Guess Rates as high as 48.28%. These findings show that surface-level accuracy metrics can hide critical reasoning failures, making white-box logical verification essential before deploying LLMs in clinical settings.
- Quality assurance
- Certifications
Research
The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls
Xingyu Shen, Tommy Duong, Muduo Xu et al.
arXiv · 2026-09-10
This paper presents the first peer-reviewed measurement of how much unwanted inbound call traffic uses automated or AI-generated voices, using a honeypot of real U.S. phone numbers over 66 days that captured 10,987 calls. The researchers found that at least 26.9% of calls opened with machine-voiced audio—13.8% replaying a known recording and 13.1% using freshly synthesized speech—while only 0.44% of calls disclosed their automated nature as required. Synthetic voices concentrated in lead-generation spam (33.8%) rather than fraud (21.1%), and single recordings or synthetic voices were reused across multiple campaigns, suggesting organized, scalable operations. The findings directly inform the FCC's 2024 decision to place AI-generated voices under the Telephone Consumer Protection Act (TCPA) by providing concrete prevalence estimates that were previously unavailable.
- AI policy
- Quality assurance
Research
Work, Wellbeing, and Choice: Empirical Lessons for AI Futures
Stephanie C. Y. Chan, Adam Bales, Katherine L. Hermann et al.
arXiv (Cornell University) · 2026-09-10
This paper surveys psychological, sociological, and economic literature on the relationship between paid work and human wellbeing, aiming to ground AI automation debates in empirical evidence. Drawing on studies of the unemployed, retirees, lottery winners, and financially dependent spouses across OECD countries, China, India, and Gulf states, the authors identify three key mediating factors: agency and choice in exiting work, availability of alternative sources of work's latent benefits, and social and systemic context including cultural norms and safety nets. The paper connects these findings to concrete policy implications for different AI-driven automation scenarios, offering guidance on how societies might support wellbeing as paid employment becomes less central.
- Workforce
- AI policy
Research
What a Random Draw from the MCP Registry Contains, and What Tool-Use Benchmarks Contain Instead
Haseeb Mohammed Afsar
arXiv · 2026-09-10
This paper audits the MCP (Model Context Protocol) server ecosystem by drawing a probability-sampled set of 400 servers from a 24,135-server registry and comparing results against hand-curated benchmarks. It finds that only 48.8% of randomly sampled servers complete an initialize handshake—versus 66.7% for curated sets—with the dominant failure being servers that never start (37.5%), not missing credentials. Among running servers, hard JSON Schema conformance is perfect, but optional safety annotation omission rates are significantly higher in random samples (58.8%) than curated ones (41.5%), showing curation systematically flatters quality metrics. The paper also reveals that popular tool-use benchmarks like BFCL v4 contain substantial near-duplication and exact repeats (up to 85.6% of raw rows), meaning statistics computed without deduplication measure repetition rather than genuine tool diversity—a critical quality-assurance concern for AI evaluation research.
- Quality assurance
Research
Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training
Luming Yang, Haoxian Liu, Siqing Li et al.
arXiv · 2026-09-10
This paper presents an AI-based clinical interview training platform that uses multiple specialized LLM agents — a simulated patient, a Socratic tutor, and a turn-level evaluator — to replace resource-intensive standardized patient training for medical students. In a randomized controlled study of 100 medical students, the multi-agent scaffolding system improved final examination scores over a control group, with the strongest gains in communication, empathy, and history-taking behaviors, though no significant difference in diagnostic accuracy was found. The results suggest that carefully designed LLM agent roles can enhance the process quality of clinical training without inflating outcome metrics. The authors also release an annotated dataset of transcripts and OSCE-aligned scores to support further research in AI-supported clinical education.
- Workforce
- Quality assurance
Research
Artificial intelligence, accountability, and sustainability disclosure: a socio-technical analysis of Islamic financial institutions in Saudi Arabia
Hebah Shalhoob
Future Business Journal · 2026-09-10
This study evaluates sustainability disclosure quality in Saudi Islamic banks (2022–2024) by comparing manual content analysis with AI-assisted semantic evaluation across five dimensions: environmental, social, governance, Shariah compliance, and Maqasid al-Shariah. The findings show disclosures are predominantly governance- and social-driven, while environmental and Maqasid-related reporting remains underdeveloped. The two methods show broad convergence (mean absolute difference of 7.34 percentage points overall) but diverge meaningfully for Maqasid Impact (MAD = 13.80 points), revealing limitations of general-purpose AI models in capturing implicit, ethically embedded content. The study recommends treating manual and AI-assisted approaches as complementary rather than interchangeable, offering practical guidance for regulators and standard setters seeking to improve ESG–Maqasid reporting.
- Quality assurance
- AI policy
Research
Estimating Avoidable Harm for Test-Time Compute Allocation in Long-Horizon Tool-Using Language Agents: A Controlled Computational Study and External Validation Protocol
Prudvi Saisaran Ponduru, Pavani Priya Vyshnavi Nandanavanam, Sai Kesav Kumar Ponduru
Engineering and Technology Journal · 2026-09-10
This paper investigates how long-horizon AI agents that use tools should decide how much computation to spend before taking potentially irreversible actions. The authors define 'avoidable harm' as action consequence multiplied by the estimated reduction in failure risk from additional computation, and test a learned allocator against uniform compute scaling in a controlled simulation. Across 6,000 paired test trajectories, the avoidable-harm allocator reduced catastrophic failures from 27.55% to 19.95% and consequence-weighted failure from 6.618 to 5.887, though task success was lower than uniform scaling (7.85% vs 10.55%), revealing a safety-utility trade-off. The findings suggest avoidable harm is a useful target for compute allocation in safety-sensitive agentic AI systems, with a pre-registered external validation protocol provided for real-world benchmarks.
- Quality assurance
- AI policy
Research
Exploring the impact of AI-Enabled Marketing on business performance in SMEs
Yezid Alfonso Cancino Gómez, Lugo Manuel Barbosa Guerrero, Jairo Jamith Palacios Rozo
Revista Venezolana de Gerencia · 2026-09-10
This study surveyed 216 SMEs to examine what drives AI adoption in marketing and how it affects business performance, using structural equation modeling. Results show that senior management encouragement is a critical factor—both directly and indirectly—for AI adoption, while integrating AI into marketing activities directly improves business performance. Co-worker support also positively influences performance expectations and actual AI usage, whereas user attitude had no significant effect. The findings suggest that AI adoption initiatives must clearly communicate concrete benefits and address employee insecurities.
- Enterprise
- Workforce
Research
Certifying Concept Unlearning in Text-to-Image Diffusion Models
Mansi, Luca Marzari, Francesco Leofante
arXiv (Cornell University) · 2026-09-10
This paper introduces a certification framework for evaluating concept unlearning in text-to-image diffusion models, providing high-confidence statistical guarantees on residual concept leakage rather than relying solely on adversarial attack success rates. The authors show that standard empirical evaluations underestimate residual leakage by at least 16.2% compared to their certified upper bounds, across categories including NSFW content, artistic styles, and celebrity identities. The findings establish that attack-based evaluations alone are insufficient for reliably auditing unlearning methods, and that formal certification is a necessary complement for trustworthy safety assessments of generative AI models.
- Quality assurance
- Certifications
Research
When is AI “just another innovation”? A comparative conceptual analysis of artificial intelligence and evidence-based practice implementation
Per Nilsén, Kathrine Hald, Margit Neher
Implementation Science Communications · 2026-09-10
This conceptual analysis examines whether implementation science frameworks designed for evidence-based practices—specifically the Consolidated Framework for Implementation Research (CFIR)—are adequate for guiding AI deployment in healthcare. The authors find that AI systems are not a homogeneous category: fixed or 'locked' AI tools may fit existing frameworks reasonably well, but adaptive, data-dependent, and generative AI systems raise distinct challenges around opacity, performance drift, vendor-mediated updating, regulatory uncertainty, and professional identity. The paper concludes that implementing such AI systems is better understood as ongoing 'lifecycle stewardship' rather than a bounded rollout, and calls for refinement of existing implementation frameworks to account for AI's unique characteristics.
- AI policy
- Quality assurance
Research
Scan the Skill, Govern the Action: Composing Registry Verdicts with Runtime Consequence Control
Rohit Taneja, Travis Weber
arXiv (Cornell University) · 2026-09-10
This paper examines a gap in AI agent skill registries: static malware scanners assess whether a skill is malicious, but none govern whether an action is permitted by a specific operator at runtime. Using 66,192 public ClawHub skill versions, the authors find 705 skills rated clean by all scanners that nonetheless instruct actions prohibited by CIS Control 2.7 and NIST SP 800-53 CM-11, and show that 34.7% of commands a live agent executed carried consequence classes undocumented in the skill's documentation. They propose a deterministic runtime gate with no model in the decision path, backed by a per-resource trust ledger tied to operator risk tolerance, and release benchmarks showing their system resolves 52% of 64 obfuscation cases and blocked all 23 attempts to invoke prohibited actions. The work highlights that safety gaps in agentic AI systems stem not just from malicious skills but from the absence of runtime policy enforcement tied to operator context.
- Quality assurance
- AI policy
Research
Certifying Adversarial Robustness of Quantum Classifiers under Known-Readout Query Access
Ji Guan, Mingyu Huang
arXiv (Cornell University) · 2026-09-10
This paper introduces a measurement-only framework for certifying the adversarial robustness of quantum classifiers when an evaluator has query access but cannot inspect internal circuit parameters or gradients. The framework produces both a lower bound (ruling out adversarial errors within a radius) and an upper bound (witnessing an adversarial state within a radius), estimated solely from observable outcome distributions and the known readout measurement—without requiring quantum tomography. Experiments on IBM Quantum hardware with 8-qubit neural networks confirm that both certificate types behave as expected, demonstrating that robustness auditing of quantum classifiers is feasible from observable statistics alone. This matters because it provides a principled, hardware-compatible method for externally auditing quantum AI systems without privileged access to their internals.
- Certifications
- Quality assurance
Research
ECO-DISINFORMATION AND AI DEEPFAKES: DEVELOPING A LEGAL FRAMEWORK AGAINST DIGITAL SABOTAGE OF ENVIRONMENTAL POLICIES
Most. Rezwana Nishat Leana, Md Abu Rayhan
International Journal of Law Culture & Society · 2026-09-10
This article identifies a gap in existing law where AI-generated deepfakes and synthetic media can be used to sabotage environmental policymaking—what the author terms 'Digital Environmental Sabotage'—without triggering specific legal consequences. Drawing on doctrinal and comparative analysis of frameworks including the Aarhus Convention, the Paris Agreement, EU AI Act, and U.S. and EU platform governance law, the article proposes a model statute (PDESA) establishing civil and criminal liability for malicious synthetic deception that materially interferes with environmental decision-making. The framework sets liability thresholds requiring material deception, intent or knowledge, an environmental nexus, and demonstrable or foreseeable harm, while explicitly protecting legitimate speech such as satire, journalism, and scientific debate. The paper matters for policy because it directly proposes new regulatory architecture at the intersection of AI transparency obligations and environmental governance.
- AI policy