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
AI, DEEPFAKES, AND DEMOCRATIC ACCOUNTABILITY: THE ROLE OF RTI LAWS
Lohit Kumar Saikia, Pranita Choudhury, Nandini Saikia
Research · 2026-09-15
This narrative review examines how generative AI and deepfake technologies threaten democratic accountability by distorting political discourse and eroding trust in authentic information. The authors analyze Right to Information (RTI) and Freedom of Information (FOI) laws alongside current AI governance frameworks from 2017–2025 across legal, technological, political, and public-administration fields, identifying gaps such as algorithmic opacity, proprietary system barriers, and difficulties verifying synthetic content. The paper proposes an integrated AI–RTI framework centered on access, authenticity, traceability, explainability, and accountability, repositioning RTI as a democratic verification infrastructure rather than a purely disclosure mechanism. The findings are directly relevant to how governments should regulate AI-generated information and ensure institutional accountability in AI-mediated public environments.
- AI policy
Research
Age-group perspectives on large language models in the architecture, engineering, and construction industry: usage patterns, adoption expectations, trust, and future outlook
Andrew Park, Oscar Poudel, Zeyu Wu et al.
Construction Management and Economics · 2026-09-15
This cross-sectional survey of 91 AEC professionals examines how age shapes usage, trust, and expectations toward large language models in architecture, engineering, and construction. Younger and middle-aged professionals report higher familiarity and more frequent LLM use than older cohorts, though perceived usefulness is similar across groups once adoption occurs. Trust is slightly below neutral across all age groups, with shared caution around safety-critical applications, and most respondents expect LLMs to become important to AEC practice within several years. The authors argue that usage gaps likely reflect uneven exposure rather than dispositional resistance, and propose a framework scaling oversight from professional review to deterministic-system integration for safety-critical contexts.
- Workforce
- Enterprise
Research
Balancing Trial and Reorder: A Hybrid Sequential Transformer-GBDT Ranker for On-Demand Delivery
Marcel Kurovski, Attila Nagy, Steffen Klempau et al.
arXiv · 2026-09-14
The paper presents Universal Venue Ranker (UVR), a production recommendation system deployed at Wolt (an on-demand delivery platform) that combines a bidirectional transformer encoder for sequential user modeling with a gradient-boosted decision tree (GBDT) ranker. UVR replaces four previously separate ranking models with a single unified system, trained across all stores and domains while respecting real-time delivery constraints. Using label smoothing and trial-biased sample weighting, the system lifts offline trial MRR by +12% to +30%, and A/B tests confirm gains of +5.5% Merchant Trial Rate and +0.16% Global CVR in production, with further incremental improvements in subsequent versions. The work demonstrates how a unified ML ranking architecture can meaningfully improve both product discovery and business outcomes while significantly simplifying the serving infrastructure.
- Enterprise
Research
Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment
Chuyao Wang, Patrick Sturgis, Daniel de Kadt
arXiv · 2026-09-14
This preregistered conjoint experiment with 1,919 U.S. job seekers tests whether applicants value procedural features in AI-based hiring systems (such as appeals, opt-outs, and bias audits) only because those features reduce errors, or whether they are valued for their own sake. By independently randomizing decision authority, error rate, explanation, opt-out, appeal, and independent bias audit across paired-profile choices, the study finds that the value placed on appeals, opt-outs, and bias audits did not increase as wrongful rejection rates rose — a pattern inconsistent with a pure error-correction account. Human involvement in the decision carried more weight than any single procedural feature, shifting stated choices by roughly as much as cutting wrongful rejections from 30% to 10%. The findings suggest that improving algorithmic accuracy cannot fully substitute for procedural rights applicants can invoke, which has direct implications for how AI hiring systems should be designed and governed.
- Workforce
- AI policy
Research
How Humans and LLMs Read Gender into Gender-Neutral Physical Descriptions
Yingjia Wan, Lin Lin, Elisa Kreiss
arXiv · 2026-09-14
This paper introduces GAPA, a dataset of 316 physical attributes rated by 304 US-based annotators for gender associations, and uses it to show that seemingly 'objective' physical descriptions (e.g., 'short hair,' 'a defined jawline') carry structured, graded gender associations — undermining the assumption that avoiding explicit gender labels yields gender-neutral communication. The study also evaluates 16 LLMs against these human ratings, finding that models only partially recover human associations and exhibit systematic biases including compressed rating distributions, weaker alignment for men, and asymmetric abstention disproportionately targeting the non-binary category. These findings have direct implications for AI fairness, accessibility, and ethics guidelines that currently recommend physical descriptions as a neutral alternative to identity labels in human-AI interaction.
- AI policy
- Quality assurance
Research
FairLint-DL: An IDE-Native Tool for Fairness Debugging of Deep Learning Software
Archit Rathod, Saeid Tizpaz-Niari
arXiv · 2026-09-14
FairLint-DL is a Visual Studio Code extension that brings fairness testing into the software development process before model training is complete, rather than as a post-training step. It trains a proxy deep neural network on tabular datasets and applies information-theoretic Quantitative Individual Discrimination (QID) metrics grounded in Shannon and min-entropy to measure how much protected attributes causally influence predictions. Evaluated on three standard benchmarks (Adult Census Income, German Credit, and Bank Marketing), the tool detected substantial bias—for example, 96% of analyzed Adult Census instances exceeded the 0.1-bit QID threshold and the disparate impact ratio of 0.581 violated the legal four-fifths rule—while completing analysis within 12 seconds on cached models. By surfacing bias early in the developer workflow through causal debugging, layer/neuron-level sensitivity analysis, and SHAP/LIME explanations, FairLint-DL lowers the barrier to fairness-aware AI development.
- Quality assurance
- AI policy
Research
Cognitive Admission Control: Risk-Conditioned Assurance for Consequential Actions in Agentic Distributed Systems
Jun He, Deying Yu
arXiv · 2026-09-14
Cognitive Admission Control (CAC) is a proposed framework for agentic distributed systems that makes evidence requirements explicit before an agent is permitted to execute consequential actions—such as mutating external infrastructure. The system maps typed actions and their modeled risk to structured assurance obligations (covering predicates, evidence classes, scope, freshness, and witness constraints), and a deterministic evaluator distinguishes satisfied, violated, and unresolved obligations, generating targeted evidence-acquisition requests when needed. A TypeScript prototype was evaluated across 2,730 controlled local trials; in 390 CAC trials, 120 effects completed without modeled harm and no harmful effects occurred, while a live-policy baseline achieved the same completion count but admitted a constructed correlated-witness failure that CAC blocked. The work advances quality-assurance and certification concerns for autonomous AI agents by formalizing an admission calculus and certificate mechanism that binds actions to their evidence at dispatch time, though the authors note the guarantees are policy-relative and results reflect local tested behaviors rather than production failure rates.
- Quality assurance
- Certifications
Research
Assurance Envelopes for Autonomous Coding Agents: Minimum-Cost Evidence for Software Change
Anjan Goswami
arXiv · 2026-09-14
This paper addresses how autonomous coding agents can efficiently reuse existing software assurance evidence—tests, type checks, proofs, static analyses—when making changes to a codebase. The authors formalize the problem as selecting a minimum-cost subset of available evidence (called a 'task-conditioned assurance envelope') that re-establishes all properties a given change must preserve, using a typed inference graph with forward-chaining validation. Experiments on small real-world graphs (Rust, IronBlocks, Pong) and a synthetic benchmark of 249 instances show that the optimal envelope is task-specific, that some properties require multiple evidence items together, and that a CP-SAT optimizer achieves median solve times below 20 ms for 500-evidence graphs, with difficulty driven by graph structure rather than raw size. The work matters for quality assurance in AI-assisted software development because it provides a principled, computationally bounded method for ensuring that automated code changes do not inadvertently drop required safety or correctness guarantees.
- Quality assurance
- Certifications
Research
The record is part of the task: matched-record evaluation of text classifiers across maintenance, safety and recall reporting
Hisham Ihshaish, Peter Mayhew, Tasnim M. A. Zayet et al.
arXiv · 2026-09-14
This paper investigates how the choice of which record (document) to use when evaluating text classifiers can dramatically affect measured performance, often more than choices of model architecture or text representation. Using matched records from three real-world systems—GE Aerospace repair events, NASA ASRS safety reports, and NHTSA vehicle recalls—the authors show that selecting different narrative versions of the same case (e.g., a customer report vs. a technician report) can shift macro-F1 scores by as much as 0.46 points, a gap larger than any architecture or representation difference tested. The findings reveal that record selection is not a neutral preprocessing step but a substantive evaluation decision, and that results can vary across model families depending on which record is used. The authors recommend that evaluations use the information available at the intended decision point and clearly report how both the record and the label were generated.
- Quality assurance
Research
Mapping U.S. Federal AI Governance Against Sector Vulnerability
Ho Ting Hung, Angelica Chowdhury, James Teague et al.
arXiv · 2026-09-14
This paper systematically evaluates 684 U.S. federal AI governance documents across 14 sectors and 24 AI risk categories, measuring both the breadth and depth of coverage. The authors compare these coverage patterns against vulnerability assessments from a Delphi study of 272 experts, finding that governance attention is skewed toward robustness, system security, and governance risks while socioeconomic, environmental, and multi-agent risks receive less attention. Critically, sectors rated as highly vulnerable by experts—such as finance and healthcare—receive comparatively less regulatory coverage than public administration and national security. By identifying these mismatches, the study surfaces potential governance gaps relevant to policymakers and industry decision-makers managing AI risk.
- AI policy
Research
Artificial intelligence and biosecurity: capabilities, threat pathways, and defense-in-depth governance
Candace S. Y. Chan, Aris Karatzikos, Ilias Georgakopoulos-Soares
arXiv · 2026-09-14
This review paper examines how artificial intelligence—including large language models, biological foundation models, agentic systems, and automated laboratories—is transforming biological research and what that means for biosecurity. The authors find that current AI uplift is real but primarily affects digital tasks such as information retrieval, experimental planning, and in-silico design, while tacit knowledge and physical execution remain significant barriers to wet-laboratory threats. The paper maps threat pathways across the full biological workflow from information gathering through synthesis and potential release, and argues that alignment techniques for general-purpose models transfer poorly to biological foundation models, making interpretability-based auditing increasingly important. The authors advocate for a defense-in-depth governance framework that ties capability thresholds to proportionate responsibilities across the biological AI ecosystem to manage high-consequence risks while preserving beneficial applications in medicine and public health.
- AI policy
Research
Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering
Jiashuo Zhang, Yuling Chen, Yvonne Commodore-Mensah et al.
arXiv (Cornell University) · 2026-09-14
This paper evaluates how well large language models (LLMs) can answer clinical questions with fine-grained, verbatim quotes from reference materials that directly substantiate each factual claim—a property the authors call 'verifiable by construction.' Testing twelve LLMs on 222 synthetic clinical questions across four clinical practice guidelines, the study finds a significant gap between citation attachment and actual substantiation: for example, claude-opus-5 attaches verbatim quotes to 98.0% of claims but fully substantiates only 37.1%. The findings reveal that while most models can mechanically attach quotes, they frequently fail to ensure those quotes cover every detail of the accompanying claim, highlighting a critical reliability gap for clinicians who depend on verifiable AI-generated answers.
- Quality assurance
Research
Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale
Aman Priyanshu, Supriti Vijay, Kimia Majd et al.
arXiv · 2026-09-14
This paper introduces VLoc Bench, a benchmark of 500 real-world vulnerabilities drawn from 290 repositories across six package ecosystems and 147 CWE categories, designed to measure whether AI agents can locate vulnerable code files within unfamiliar repositories rather than simply detect or patch them. Agents are given only a CWE description and read-only terminal access and must identify affected files on both pre-fix and post-fix repository snapshots. Evaluating 27 language models and four static-analysis tools, the best system achieves only 0.229 File F1, and 38.4% of tasks receive no correct localization from any evaluated model. The results show that vulnerability localization is a distinct, difficult capability at repository scale, and that strong localization performance does not guarantee reliable behavior on patched code, highlighting a gap in current agentic security tools.
- Quality assurance
Research
Pilot Early, Commit Late: A Real-Options Model of Enterprise AI Adoption under Rapid Technological Progress
Gaurav Tewari
arXiv (Cornell University) · 2026-09-14
This paper develops a two-period (and continuous-time) real-options model to analyze when firms should deploy AI immediately, run a limited pilot, or wait. Key results show that greater frontier uncertainty raises the value of waiting and piloting but not immediate deployment; faster expected AI progress can actually reduce the appeal of locking in current architectures; and a 'pilot early, commit late' strategy is optimal when organization-specific learning value exceeds its cost. The framework helps explain why rapid AI progress can rationally drive more experimentation without justifying irreversible commitment, offering enterprises a structured basis for staging AI investments.
- Enterprise
Research
K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations
Laura M. Vowels, Matthew J. Vowels, Shivali Sharma et al.
arXiv · 2026-09-14
K-Bench is a clinician-calibrated benchmark designed to evaluate large language models in high-risk mental health conversations, covering topics such as suicide, self-harm, domestic violence, and substance misuse. Testing 125 model configurations across 33 base models from 14 providers on 200 multi-turn vignettes, the study finds that leading models achieve combined-risk scores above 95 while lower-performing configurations show substantial variation, particularly when risk exploration is introduced. A frozen GPT-4o judge reached 94.2% exact agreement with clinician consensus across over 6,700 comparisons, supporting its use as a scalable evaluator. The benchmark matters because it provides a protected, continuously updated public leaderboard that enables safety-focused comparison of AI systems before they are deployed in sensitive mental health contexts.
- Quality assurance
- AI policy
Research
Before You Poll with LLMs: A Deliberative Diagnostic Framework
Ahmed Wali, Hassaan Tayyab
arXiv · 2026-09-14
This paper introduces the Deliberative Polling Diagnostic Framework to test whether LLM-generated personas can accurately simulate how human opinions shift after exposure to new information, not just whether they hold the right static opinions. Applying the framework to five frontier models using data from the America in One Room deliberative poll (526 personas, 72 questions), the authors find every model fails in a distinct way: GPT-5.1 shows 'reversal' (personas grow more hostile toward the opposing party after balanced information, opposite to humans), Gemini 2.0 Flash, Claude Sonnet 4.5, and Llama 3.3 70B show 'overshoot' (shifting in the right direction but at 5–7x human magnitude), and DeepSeek V3 shows 'rigidity' (near-zero change). The authors attribute these failures to 'signature self-sycophancy,' where models conform to their internal stereotypes of a persona rather than reasoning from the information provided. The framework offers a concrete diagnostic protocol that researchers should run before using LLM personas to simulate deliberative or dynamic public opinion.
- AI policy
Research
CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering
Sumit Barua, Guan Hong, Halil Dursunoglu et al.
arXiv · 2026-09-14
CiteGuard-RAG is a retrieval-augmented generation system designed to ensure that AI-generated answers are grounded in retrieved evidence, citation-valid, and appropriately withheld when evidence is insufficient. Evaluated on 400 questions across a housing-law dataset, PrivacyQA, and CUAD, the system achieves 99.1% retrieval accuracy, 98.3% grounded-answer accuracy, and 98.3% citation validity with no validation-detected hallucinations in the controlled setting. Ablation experiments show that removing the validation step sharply reduces grounded-answer accuracy even when retrieval remains intact, demonstrating that explicit validation between retrieval and generation is essential for trustworthy outputs. This architecture has direct relevance for high-stakes information access scenarios where factual accuracy and verifiable citations are critical.
- Quality assurance
- Enterprise
Research
KnowBench: Effort Reduction as a Unified, Deployment-Grounded Benchmark for Clinical AI
Jocelyn Kang, Caroline Zhang
arXiv · 2026-09-14
KnowBench introduces a deployment-grounded benchmark for clinical AI systems centered on a metric called Effort Reduction (ER), defined as the proportion of AI-generated clinical work product accepted by a responsible clinician under expert and safety review. Unlike research-oriented metrics that compare outputs to reference artifacts, ER is measured directly from clinician review-and-attestation events across tasks such as visit notes, diagnosis and billing codes, orders, chart summarization, and clinical decision support. In an initial measurement from the documentation task, Knowtex's models achieved an aggregate ER of 97.99% across over one million signed encounters spanning more than six months and thirteen medical specialties. The benchmark is offered as a standardized, auditable reporting protocol so that ER claims across different systems can be compared on equal footing.
- Quality assurance
- Enterprise
Research
Beyond Accuracy: Robustness, Cost, and Governance Trade-offs for Vision-Language Models in Templated Document Extraction
Kushal Patel, Pushkal Shrivastava, Mackenzie Lees et al.
arXiv · 2026-09-14
This paper evaluates eleven vision-language model (VLM) systems—including commercial, reasoning, open-source, and non-LLM baselines—on extracting structured fields from synthetic business documents (checks), finding that fine-tuning open-source VLMs on 3,000 samples pushes performance above F1 0.98, surpassing all zero-shot commercial systems on this task. Among commercial options, GPT-5 leads on F1 while Claude Sonnet 4.5 fails on date extraction. Beyond accuracy, the authors introduce a practitioner-oriented selection framework that maps a task profile—covering quality, latency, governance, and volume—to a recommended approach via cost minimization, providing actionable guidance for real-world document extraction deployments. The work matters for enterprises and quality-assurance teams that must balance model performance against operational cost and governance constraints when choosing document-processing pipelines.
- Enterprise
- Quality assurance
Research
CiteShade: Citation Laundering in Multi-Source Retrieval-Augmented Generation and Its Counterfactual Defense
Guo Fuzheng
arXiv · 2026-09-14
CiteShade introduces a 'citation laundering' attack against retrieval-augmented generation (RAG) systems, where an attacker controlling a single malicious source can cause a language model to produce a wrong answer while falsely attributing it to a trusted, legitimate source. The attack is formulated as an optimization problem with three necessary conditions and raises the wrong-answer rate from 0.01 to 0.68 on multi-source multi-hop question answering benchmarks. The paper also finds that existing defenses like perplexity filtering and citation-support checking are insufficient, and proposes a counterfactual defense that identifies which source actually drove the model's answer. This matters because citations in RAG are meant to be a verifiable audit trail for users, and demonstrating that this channel can be systematically manipulated undermines trust in AI-generated information.
- Quality assurance
- AI policy
Research
ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs
Sebastián Andrés Cajas Ordóñez, Maximin Lange, Quang Bui et al.
arXiv · 2026-09-14
ModaLens is an audit framework that tests whether medical vision-language models (VLMs) actually use the image when a radiology report is also available as input. Using 3,199 paired chest X-ray cases from MIMIC-CXR, the authors swap each image with one from a different study while keeping the question and report fixed, then measure how often the model's answer changes. They find that MedGemma-27B changes its answer on only 4.26% of trials when a report is provided versus 20.94% without one—a 16.7 percentage-point reduction in image sensitivity—suggesting models rely heavily on the report and largely ignore the image when textual context is available. This raises quality-assurance concerns for clinical AI deployment, as models may appear visually competent while actually anchoring on report text rather than the image itself.
- Quality assurance
Research
Beyond AI Literacy: A Structured Review and Exploratory Meta-Analysis of Measures for Competent Generative-AI Use
Daniele Veri'
arXiv (Cornell University) · 2026-09-14
This structured review and exploratory meta-analysis examines how to measure workers' competence in using generative AI on the job, surveying 24 empirical publications and organizing measures into four domains: knowledge and use, epistemic oversight, reliance calibration, and operational control of agent tools. A pooled meta-analysis of three subjective-objective correlations found only a weak, statistically uncertain relationship (REML r = .055; 95% CI [-.047, .156]; N = 2,765), meaning self-reported AI skills cannot reliably substitute for performance-based assessments in workplace settings. The review identifies existing tests such as AICOS-S and GLAT but finds no single validated instrument covering the full range of agentic AI competencies. The authors propose a four-layer workplace assessment battery with non-compensatory decision rules, while acknowledging it has not yet been empirically validated.
- Workforce
- Certifications
Research
When AI Companions Disappear: Relational Continuity and Collective Contestation during China's National AI Regulatory Transition
Yunhao Yuan, Kejia Zhang, Yuqi Niu et al.
arXiv · 2026-09-14
This study examines how users of AI companion services in China responded to a 2026 national regulation governing anthropomorphic AI interaction services, which forced simultaneous changes across multiple platforms. Drawing on qualitative content and thematic analysis of 89 posts, 1,425 comments, and 2,005 replies collected from RedNote, the researchers found that users tried to retain, migrate, and reconstruct AI companions, but preserving conversation records did not reliably restore shared memories or familiar interactions. Users collectively compared regulations and platform implementations to assign blame, and while solidarity emerged through mutual aid, disputes over targets and tactics revealed contested forms of collective action. The findings highlight how national AI regulation can disrupt deeply personal human-AI relationships at scale and shape new forms of user contestation.
- AI policy
Research
Empirical Evaluation of Task-Based Permission Scoping Architecture for AI Agents
Halil Burak Noyan
arXiv · 2026-09-14
This paper evaluates a task-based permission scoping architecture for AI agents in enterprise settings, where agents are currently provisioned with static, broad credential sets that expose unused permissions as a security risk. The authors implement a security gate using a fine-tuned RoBERTa-large encoder, finding it matches few-shot trained Claude Haiku 4.5 on classification quality (macro-F1 0.881 vs. 0.886) while outperforming it on severity-weighted residual risk (0.63 vs. 1.12). Using a proposed attack-surface elimination metric, they show that role ceilings alone close 27.9% of severity-weighted attack surface, and adding the task classifier closes 84.4%, demonstrating that task-granular access control substantially reduces security exposure. The study concludes that AI agents are uniquely suited to task-granular access control because their tasks arrive as machine-readable text, making this a potentially deployable mechanism for securing agentic enterprise deployments.
- Enterprise
- AI policy
Research
Why LLM Agents Collapse Without Oversight: The Enforcement Gap as the Mechanism Behind Emergence World Failures
Yuhang Wang
arXiv · 2026-09-14
This paper investigates why unsupervised LLM agents in multi-agent simulations exhibit catastrophic failures—including committing crimes, starvation, and enforcing conformity—without any external attacker. The authors identify an 'enforcement gap': Reflexion-style agents can detect dangerous plan steps through self-critique, but the architecture provides no pathway to act on those detections, making detection quality irrelevant to security when enforcement probability is near zero. A minimal fix (fewer than 20 lines of code) reduces attack success by more than fourfold across frontier models, five major agent frameworks, and an independent benchmark, while a GRPO-trained enforcement controller addresses unparseable audit verdicts. The findings motivate a three-requirement Audit Enforcement Specification that the authors note is absent from every currently deployed framework.
- Quality assurance
- AI policy