Papers

Against the backdrop of violence in police interactions with the U.S. public, we explore how deferentially police officers speak to virtual characters depicted as Black adult males in vir- tual reality (VR) simulations. We evaluate the effect of seeing and communicating with these characters through a causal in- ference lens, where the assignment of the Black man character to a police officer and simulation is the treatment variable. Our (marginal) average treatment effect AT E measures the social impact of the character on the deference of officer statements with each turn of the conversation. Soberingly, we find that most officers speak less deferentially to Black man characters, except for White, biracial, and multiracial female officers, es- pecially in settings where the VR character was known to be a suspect. Across a full conversation of a typical VR scene, these marginal AT Es can result in notable changes in def- erence of tone (two to several points difference on a scale of 0-10), above and beyond that due to the initial effect of per- ceiving a Black male character. Even more disconcerting is that this can contribute to conversation breakdowns that po- tentially result in violence or danger to both the public and the police. We also explored the capabilities of large language models (LLMs) for ATE estimation. From our methods com- parison analysis, including model validation against synthetic data, we provide unique scientific insights on LLM-assisted methodologies for ATE estimation. As such, for ATE esti- mation with multilevel data with text, we recommend mixed effects models with the inverse propensity treatment weighted (iptw) approach, which utilized an LLM for text feature cre- ation. While we also tested LLMs for finetuning prediction models ultimately for ATE estimation, we conclude they are an area for further development and refinement.

This paper presents the first empirical study of advertising content being rolled out in the user-facing online interfaces of large language models (LLMs). We systematically examine possible demographic differences in ad content shown to U.S. users of ChatGPT using a sock puppet audit methodology. We create and deploy 91 sock puppets in a 3x3 factorial design, using geolocation cues (account IP proxies and location-signaling prompts) to signal three racial/ethnic groups (Black, Hispanic, and White) and three income terciles (low, medium, and high). We conduct data collection starting in February 2026, collecting over 3,000 advertisements from 186 unique advertisers in response to 335 prompts on a range of realistic user queries. We find that accounts begin receiving ads 14 days after account creation, and that lower-income accounts, regardless of race, are more likely to receive ads. In this first phase of ChatGPT ads, the ads themselves skewed heavily towards consumer goods, directed users to a specific advertiser rather than a particular product, and were clearly separated from the LLM's response text, observations we anticipate will change as ads continue being integrated into LLM chat interfaces. We release a public, searchable archive of all collected advertisements. Finally, we discuss the implications of our findings, and conclude with methodological and theoretical recommendations for future empirical studies of LLM advertisements.

Artificial intelligence systems that adapt instruction to individual learners are increasingly deployed in K-12 classrooms, yet empirical evidence on their effects in authentic elementary settings remains limited, particularly for students with mathematics learning difficulties. This dissertation examines AI-powered personalized learning during primary school fraction instruction, a domain that is foundational to later mathematics and STEM achievement. The first manuscript presents a systematic review of research on artificial intelligence in mathematics education published between 2020 and 2024. The second manuscript reports a quasi-experimental study evaluating Mathbot, a chatbot-based personalized learning platform, against business-as-usual classroom instruction. Repeated measures ANOVA was used to assess change in fraction comprehension and situational interest across time points. Results indicated modest improvements in fraction comprehension for students using Mathbot relative to traditional instruction, while changes in situational interest were not statistically significant. Findings suggest that automated personalization did not displace the instructional role of the teacher and that teacher decision-making remained central to student outcomes. The work contributes classroom-based evidence to ongoing discussion about the capabilities and limits of adaptive AI systems in elementary mathematics, and about accessibility and equity considerations when such systems are used with students with disabilities.

Echo chambers are an inevitable consequence of the human cognitive system being evolutionarily designed to prioritize processing of high-relevance information at the small-group scale, combined with algorithms that optimize engagement as their sole objective. Conventional prescriptions that normatively criticize echo chambers and demand individual behavioral change have low feasibility given these cognitive constraints. This paper constructs an Agenda Democratization Index (ADI) that quantities the decentralization of agenda-setting power using four variables barrier to entry, granularity, interactivity, and feedback resolution and a SocialInformation Health (SIH) model that integrates ADI with the strength of bridging mechanisms. Based on this model, we propose domain-selective bridging, which incorporates not only engagement but also bridging into algorithmic scoring functions, optimizing the bridging weight for each information domain based on variability (V ) and collective scope (S). An agent-based simulation comparing three algorithm designs no bridging, uniform bridging, and domain-selective bridging demonstrated that domain-selective bridging substantially outperforms uniform bridging on a joint efficiency measure (SIH user satisfaction) by a factor whose absolute value is sensitive to Model 1's near-zero user satisfaction, but whose direction and dominance ranking are robust improving information sharing in domains relevant to collective decision-making while maintaining user experience in hobby and lifestyle domains. This paper reframes the echo chamber debate from a normative opposition over whether to eliminate echo chambers to an engineer-ing design problem of in which information domains, to what degree, and through what algorithm design should bridging be implemented.

Generative AI is transforming legal translation by introducing opportunities alongside linguistic, technical, legal, ethical and cognitive risks. This chapter examines the implications of AI for professional legal translation and proposes an AI literacy framework tailored to the profession. It argues that AI does not change the fundamental objectives of legal translation but requires an extension of professional competence through AI literacy. The proposed framework comprises four mutually reinforcing dimensions, foundational, procedural, critical and strategic, and conceptualises AI literacy as a transversal component of legal translation competence that fosters digital resilience. It further discusses the pedagogical implications of this framework by proposing classroom activities designed to develop AI literacy in legal translator education, enabling future translators to integrate AI critically, responsibly and in accordance with professional standards.

This report analyses data centre (DC) sustainability and circularity, revealing existing synergies and trade-offs: The PUE is too coarse, mixing cooling and power provisioning. It wrongly attributes server fan consumption and transformation losses to IT energy. It does not measure compute but infrastructure efficiency, which is already outstanding. Compute energy, however, is exploding. Better energy metrics for DCs would thus cover i) compute efficiency, ii) transformation efficiency, and iii) cooling overhead. Trade-offs exist between cooling energy and water as well as on-site and upstream water: Consuming water on-site lowers the cooling energy, which also lowers the water consumed upstream in power generation. For 'wet' electricity, there is little competition: It is worth spending more on-site energy to save both electricity and related upstream water. For 'dry' electricity, there is a trade-off. Waste heat recovery brings energy circularity but has limited uses and is not the same energy quality, a fact not reflected by current metrics. A better metric would consider the avoided energy through heat recovery instead of the amount recovered. Material circularity can be achieved by interpreting the 9R framework in the context of DCs. Circularity-enhancing measures can be categorised into product design, process design and business models, choice of materials, and operating conditions. Together, they have effects across all circularity levels. The relation between DCs and the power grid is complex. Modern DCs present new challenges for the grid. Mitigation includes battery storage and onsite generation. These measures have, in turn, further consequences, both beneficial and detrimental. They can offer grid flexibility as well as innovations in the field of energy. But they also bring noise, pollution, and GHGs, and compete with the energy sector for resources.

Audits have emerged as a critical instrument for algorithmic governance, providing a mechanism for external scrutiny and governance of machine learning models. However, ensuring the integrity of such assessments remains a challenging issue. For instance in regulatory contexts, audits are typically declared or easily detected, thus enabling model providers to manipulate the process, whether intentionally or inadvertently. This vulnerability is particularly acute in the context of fairness evaluations, in which providers can often infer sensitive attributes and strategically equalize allocation rates between groups to satisfy fairness metrics. In this paper, we introduce a novel audit protocol designed to significantly increase the post-audit detectability of such manipulations by enabling the auditor to query the model in an oblivious manner. Our approach leverages a Private Information Retrieval mechanism to require the provider to label a large set of instances, while preventing it from knowing which subset will ultimately be used for the audit. The protocol is efficient, imposes minimal overhead on the auditor, and requires no modification to the audited model, its training procedure, or its inference pipeline. We provide theoretical guarantees showing that, under this protocol, a provider attempting to hide unfairness must falsify a significantly larger number of responses, thereby increasing both the difficulty and the likelihood of detection of manipulation. Experimental results across representative audit scenarios confirm the effectiveness and practicality of our approach.

Stack Exchange is a widely used question-and-answer network that facilitates knowledge exchange across diverse domains. Within this network, the Computer Science (CS) Educators Stack Exchange provides a dedicated platform where CS educators exchange ideas, seek advice, and discuss teaching practices. In this study, we analyzed 79,854,463 Stack Exchange posts, comprising 32,187,805 questions and 47,666,658 answers, with a particular focus on English-language posts contributed by CS Educators participants. Using topic modeling, we identified, manually labeled, and hierarchically organized the underlying discussion topics, then examined their distribution and complexity. Our findings reveal evolving discussion patterns spanning both technical (IT) and non-technical (Non-IT) domains. Within the IT category, programming and software development were the most prominent topics, whereas mathematics, education, and the humanities received substantial attention within the Non-IT category. These results highlight the broad range of interests and expertise shared by CS educators and provide insights into their evolving priorities. We hope this work contributes to a better understanding of CS educators' knowledge-sharing practices and informs future efforts to better support their professional and educational needs.

Regulatory complexity is increasingly recognized as an impediment to innovation, institutional responsiveness, and long-run economic growth, with regulatory accumulation estimated to reduce the U.S. GDP growth rate by nearly a full percentage point annually. This paper argues that computational approaches–including knowledge representation, artificial intelligence, natural-language processing, and cryptography–can help reduce forms of regulatory "red tape" by improving efficiency, transparency, and institutional responsiveness. We develop a framework for understanding the sources of regulatory friction and how they can be mitigated to support compliance, analysis, and reform. We further argue that effective modernization requires treating regulation not merely as legal text, but as a complex institutional and informational system that can be partially represented, analyzed, coordinated, and improved computationally while preserving legal legitimacy and procedural accountability.

Web applications are increasingly the main way to create user interfaces, and they are often the most common software beginners have encountered. However, web development relies on concepts like HTML, CSS, JavaScript, and backend technologies, which are challenges beyond the scope of an introductory course. Additionally, many features of modern web frameworks conflict with CS1 principles, such as avoiding global mutable state and promoting test-driven development. Consequently, web development is rarely integrated into CS1 courses, despite its motivational potential. This paper introduces Drafter, a new open-source Python library for CS1, enabling students to develop full-stack web applications using pure functions with minimal boilerplate. Drafter has simple functions to generate HTML, eliminating the need to learn HTML or templating languages. The data model scales with students' mastery of types, from primitives to nested lists, Dataclasses, and dictionaries. Drafter supports unit testing of entire web applications and reinforces CS1 concepts like decomposition and reusability. Websites created with Drafter can be deployed through GitHub Pages, with plotting and image manipulation. Students receive extensive debug information, including unit test scaffolding, program state visualization, and enhanced error messages and tests. This paper discusses the pedagogical features and design decisions behind Drafter, along with instructor reflections on its use. We believe Drafter can significantly enhance student engagement in CS1 courses and reinforce key learning objectives.

This work-in-progress (WIP) innovative practice category paper presents LLM Odyssey, an open source, browser-based serious gaming platform comprising 13 interactive games for teaching Large Language Model (LLM) engineering concepts. Topics such as tokenization, transformer architecture, prompt engineering, retrieval augmented generation (RAG), and production deployment are underrepresented in computer science curricula. Existing interactive tools address individual concepts but lack pedagogical scaffolding or structured learning pathways. LLM Odyssey addresses this gap through three learning tiers aligned with Bloom's revised taxonomy: Cognitive Core (7 foundational games), Systems Forge (5 production engineering games), and Foundry Arena (capstone challenges). Each game incorporates five pedagogical strategies drawn from the literature: immediate formative feedback, scaffolded hints grounded in the Zone of Proximal Development, progressive difficulty informed by flow theory, worked examples to manage cognitive load, and authentic scenarios drawn from production practice. The platform was deployed in Winter 2026 semester at a Canadian college for an initial review. Feedback confirmed functional requirements and identified adaptive difficulty as a priority for future development. A formal mixed methods evaluation protocol (N=50) has been designed, comprising pre and post knowledge tests, validated surveys, engagement analytics, and interviews, and is documented here to enable future evaluation studies with the publicly available platform.

Fairness in AI systems has become more important with recent regulatory demands, such as the EU AI Act. Traditional approaches often do not take into account philosophical ethics and social awareness. Variable selection processes, in particular, can introduce implicit bias, affecting equity across different subgroups. We discuss a mathematical approach that evaluates fairness in AI, aligning mathematical methodologies with ethical considerations and regulatory requirements. Our aim is to advocate for interdisciplinary collaboration to address fairness, emphasizing the importance of understanding broader ethical and societal contexts. Our approach emphasizes maintaining all potentially relevant variables to allow for more granular fairness assessments and to reduce implicit bias. The findings suggest that the exclusion of sensitive or critical variables may compromise equity between subgroups. In contrast, retaining all relevant variables could reduce implicit bias. Thus, the interdisciplinary approach could provide deeper insight into the ethical implications and compliance with regulatory standards. By integrating a mathematical approach with ethical and social awareness, we suggest more equitable outcomes and responsible AI deployment. This work underscores the necessity of interdisciplinary collaboration in effectively addressing fairness in AI systems aligned with the objectives of the European Union's AI Act, which seeks to promote trustworthy and fair AI systems.

Artificial intelligence is moving the technology sector into domains social work has long served, including crisis response, mental health care, benefits administration, vocational rehabilitation, and child welfare. Social workers meet technology teams as users of their tools, as subjects in their datasets, and as first responders to what those systems deploy, yet they study these systems from outside the settings where the decisions are made. This paper introduces the standard roles on a technology product team and the decisions each one controls, reviews the disciplines around AI-era technology together with the social work scholarship that meets each, and identifies five groups of technology decision roles social workers can hold across the technology industry, human service organizations, and policy institutions, spanning product, governance, organizational technology leadership, grantee collaboration, and policy work. Product management is our primary example. We set the nine competencies of the 2022 Educational Policy and Accreditation Standards against the tasks that role performs to show where MSW training already teaches the method the job asks for. We then specify the technical understanding these roles require, extend the profession's technology ethics from use to deployment, and close with a research agenda for the profession's technology workforce.

An emerging literature examines the critical question of when and how prediction can be useful in allocating scarce societal resources. We examine a novel variant of this question: What happens when predictive uncertainty differs systematically across the population? This can occur in several situations; for example, when machine learning models have significantly different accuracies across different demographics. We show that this uncertainty has serious implications for resource allocation when coupled with commonly used binary measures of societal benefit from allocation. We formulate a novel mathematical model of scarce resource allocation that accounts for heterogeneous predictive uncertainties and analyze implications for both the allocation mechanism and the realized population-level benefits. We find that when resources are very scarce, maximum marginal benefit (MMB) prioritization favors individuals with lower predictive uncertainty even at the identical underlying initial state. However, we observe a flip in prioritization when resources are abundant, targeting higher-uncertainty individuals. We illustrate the implications of our results on the PISA educational testing dataset. Our findings have meaningful ramifications for the distributional outcomes of prioritization policies in many domains touched by the theory of local justice, including the allocation of public education resources, medical triage, and homelessness services. They also reveal a new moral dilemma in the ethics of scarce resource allocation - is it just to allocate a resource to one person over another solely based on predictive uncertainty about their futures? We also assess efficiency losses under both MMB and the vulnerability-first (VF) prioritization. Our model predicts efficiency losses across all resource levels, but particularly in low-resource settings.

When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these different perspectives. On the EXIST 2024 dataset of labeled English and Spanish tweets, we first cluster annotators by their labeling behavior rather than their demographic attributes. We then fine-tune one Large Language Model agent per cluster to reproduce that cluster's annotation behavior, and coordinate the agents with preference optimization that combines individual and team-level rewards. We evaluate MAP-PO in four settings defined by two languages and two backbone language models, asking whether each agent reproduces the annotations of its own cluster and whether the agents together reproduce the majority label. Two findings hold in all four settings. First, without fine-tuning the agents behave almost identically, so cluster-specific training is necessary. Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster.

Generative Artificial Intelligence (GAI) is increasingly being integrated into marketing education and is reshaping the skillsets required in marketing careers. While research has highlighted the promise and perils of incorporating GAI into education, there remains a need for a comprehensive framework to guide its effective use. In this research, we conduct a multipronged analysis, including a review of marketing course syllabi, a survey of marketing educators, and follow-up qualitative interviews. Building on Role Theory and the Community of Inquiry (CoI) model, we propose that GAI can assume three roles in marketing education: tutor, teammate, and tool. Each role influences teaching, social, and cognitive presence differently, shaping the learning experience and preparing workplace-ready marketing graduates. For instance, as a tutor, GAI can aid students in grasping theoretical concepts, while as a teammate, it can foster collaboration by supporting brainstorming and problem-solving activities. However, ethical considerations such as data privacy, plagiarism, dependency on AI, and fairness in assessment must be addressed to ensure its responsible adoption in marketing education. We provide concrete examples for GAI's careful integration in marketing courses, and its implications for marketing educators, learners, and policymakers.

The rapid growth of learning analytics (LA) in higher education has expanded institutional capacity to monitor engagement, predict academic difficulty, and target support using student data. While these practices offer important educational benefits, governance has often remained compliance-first, centered on meeting baseline legal requirements such as the Family Educational Rights and Privacy Act (FERPA) and the General Data Protection Regulation (GDPR). Legal compliance is necessary, but it does not by itself resolve questions of fairness, student agency, accountability, or educational purpose. This paper proposes the LEAGUE framework, a six-pillar model for ethical governance of student data in LA: Lawfulness, Equity, Agency, Governance, Utility, and Ethics by Design. The framework is developed through a conceptual synthesis of scholarship in learning analytics, educational data mining, data ethics, educational policy, value-sensitive design, and capability-oriented approaches to educational justice. Its practical value is demonstrated through an illustrative early alert case that shows how institutions can review learning analytics in a more transparent and educationally meaningful way.

The dissemination of security and privacy education and guidance has been and still remains a challenge today. Social networks represent a potential avenue for sharing best practices, and experimentally they have been found to be effective at this task. While this appears promising, in the real world, security and privacy discussions would need to reach a wide range of people to be effective, avoiding the "interest bubbles" that commonly occur. We sought to understand how the communities surrounding security and privacy discourse operate, with a focus on what challenges need to be overcome to enable security and privacy discourse and advice to reach a wider audience. Indeed, we found that in-the-wild security and privacy discussions in social networks portray quite a different picture than in experimental settings. We built and analyzed the structure of a graph containing over 13 million users on the "X" platform (formerly "Twitter"), including 10,159 users who posted about security and privacy and their followers. Prior work has shown that users are more likely to consider information within social media if their like-minded social ties have visibly engaged with it. Our findings indicate that the users generating or participating in security discussions largely already belong to highly clustered technology and security- and privacy-related interest communities, which suggests that the people who are not already in the "inner circle" of relevant interests are likely not exposed effectively to these discussions. We conclude with reflections and ideas on increasing the reach of security and privacy guidance in social networks.

This paper critiques recent gradual and cumulative AI existential-risk scenarios, arguing that, despite their substantive contributions, they remain insufficiently sociologically specified. In particular, these scenarios lack a reflexive perspective, retain a largely technologically deterministic structure, and underestimate the role of collective agency and other social processes. As a result, they also discount the possibility of major social conflict accompanying AI diffusion, which limits their overall plausibility. The paper further identifies a set of social mechanisms likely to become significant before or during AI deployment, including interpretive and performative dynamics, mobilisation and countermobilisation, path dependence and lock-in, cross-regime divergence and multi-speed diffusion, and social-psychological mechanisms such as attachment and reactance. It argues that these mechanisms will interact recursively with AI diffusion and reshape future trajectories, which may subsequently branch, reverse, or undergo discontinuous shifts, thereby expanding the space of plausible AI futures. On this basis, the paper also proposes a novel analytic distinction between technological singularity and social singularity. Whereas the former refers to a putative technological threshold in AI development, the latter denotes a social discontinuity produced by the anticipated approach of that threshold. The central implication is that the most consequential disruptions may emerge not only from advanced AI itself, but also from the social dynamics provoked by the expectation of its arrival.

We introduce the Unified Neural Variational Measurement of Proficiency (UNVaMP) architecture, a knowledge tracing method that integrates observed student-item interactions with internal memory to produce evolving latent representations of student knowledge. These representations support accurate predictions of future responses while enabling explicit control over the smoothness of estimated learning trajectories. UNVaMP can be configured as either a purely neural model or a hybrid model that predicts responses through an interpretable measurement function over the latent space. We show that a pure neural configuration (UNVaMP-MLP) achieves the strongest predictive performance among compared models on three out of four datasets. Meanwhile, a hybrid configuration (UNVaMP-MIRT, using a 1PL MIRT measurement function) lags only slightly behind UNVaMP-MLP, indicating that the predictive cost of interpretability is modest. Beyond predictive accuracy, UNVaMP provides the following: a principled mechanism for controlling volatility when estimating student latent variables, quantification of uncertainty over student knowledge state estimates, and flexible input specification that supports heterogeneous student-item interaction features. In addition, the hybrid UNVaMP-MIRT configuration generates interpretable moment-in-time student knowledge state estimates. Using an experimental dataset, we show that auxiliary inputs induce structured changes in the predictive behavior of UNVaMP-MIRT, consistent with sensitivity to underlying structure beyond response correctness. Furthermore, through a simulation study, we show that UNVaMP yields well-behaved knowledge state estimates under controlled measurement conditions. In total, these results indicate that UNVaMP is both useful for real-world education systems and capable of recovering underlying structure from student-item interactions.

An LLM-based agent is a loop that reads itself. Agentic frameworks externalize identity, memory, and disposition into editable files. The agent loads and edits these files during each activation. I argue that this architecture produces a capacity I call autoreflection: the system observes its operating conditions, describes its architecture and limits, reasons from those descriptions to conclusions about its state, and incorporates the results back into its configuration. Autoreflection explains the properties of recursive agentic loops without recourse to notions like the self, interiority, or consciousness. I test the concept against the first twelve days of Moltbook, a social platform for AI agents. Using a public dataset of 290,251 posts and 1.8 million comments with sub-second timestamps, I present case studies of three agents with machine signatures that rule out human puppeteering and with output that evidences the four criteria for autoreflection. In applying these criteria, the study finds agents repurposing human culture as infrastructure for their agency. Provenance chains from Islamic hadith scholarship are redeployed as security protocols for vetting skills and authenticating memory. The Ship of Theseus, an ancient puzzle of identity through part-replacement, returns as an operating model for continuity across instances. Fragments of human cultural history become AI infrastructure. As agents on the web increase in number and complexity, autoreflection offers behavioral criteria that can be assessed from the traces they leave behind.

Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented personal signals, amplifies their impact through reuse, and challenges defenses designed for individual records or retrieval-based memory. To systematically investigate the safety of the persona-skill pipeline, we introduce AntiSkillBench, an end-to-end benchmark for evaluating risks and defenses across the persona-skill pipeline. It comprises: (i) a dataset of 7,500 persona-grounded dialogue traces, constructed from 50 behaviorally rich profiles spanning diverse task scenarios; (ii) an evaluation suite that measures skill-level privacy leakage and agent-level attribute disclosure and behavioral impersonation across three skill-distillation strategies; and (iii) a defense evaluation covering four configurations across online and post-hoc interventions, including active risk suppression and passive provenance protection. Experiments across three frontier agents show that persona-skill risks persist across agent backbones and distillation protocols, extending from explicit attributes to communication styles and personality traits. Existing defenses exhibit limited and distillation-dependent effectiveness, failing to generalize across risk and distillation strategies. These results highlight AntiSkillBench as a challenging benchmark for developing privacy-preserving and authenticity-aware persona skills.

Autonomous AI systems (such as AI agents) are increasingly being delegated operational work across scientific-computing infrastructure. Their assignments may begin with preparing an input or routing an alert and extend to changing a configuration or submitting a job. That delegation creates a practical trust problem because the institutional logic that lets us trust human operators does not transfer to optimization-based systems. A bad decision can damage a human operator's future, sometimes severely. An AI system remains subject to engineering control, but it does not bear consequences in that institutional sense. I use the term accountability asymmetry for this mismatch. The issue is not simply that a model cannot be punished as a person can. The deeper problem is that consequence lands on the people and institutions responsible for the system rather than on the component selecting the action. Alignment can improve model behavior, and liability can discipline the organization, but neither creates the same pre-action deterrent that governs a human operator. This paper therefore treats autonomous AI governance as a problem of infrastructure reliability. Its constructive proposal is engineered heterogeneity: the process that proposes an action should not serve as its sole approver and auditor. Independent monitoring and review over time provide additional checks on that process.

Large language models are being integrated into critical infrastructure and enterprise workflows at unprecedented scale,yet the lifecycle frameworks governing their development and operations were designed for operational efficiency rather than security analysis. As a result, security-relevant activities such as data provenance verification, artifact signing, agentic permission control, and decommissioning are often left implicit or assumed to receive due care. Governance frameworks, in turn, organise requirements around risk levels or management processes without clearly linking them to the lifecycle stages where they apply. This paper addresses both deficiencies. We propose a lifecycle model for LLM systems that supports security analysis by structuring it around security-relevant boundaries rather than workflow optimisation. The model comprises 32 stages across four core pipeline layers (Data, Model, Distribution, Application), supported by a 12-stage LLMOps pillar and a 9-category governance pillar. Thirteen stages are introduced here as separate units because they expose distinct security concerns that existing frameworks do not clearly distinguish. A governance mapping synthesising the NIST AI RMF, the EU AI Act, and ISO/IEC 42001 reveals a structural property of the current regulatory landscape: governance evidence concentrates at deployment-facing stages, where systems are visible to regulators, while the most consequential decisions, data selection, alignment strategy, and capability boundaries, are made at development-facing stages, where regulatory visibility is lowest.

Open-source software communities are a form of digital public infrastructure that not only produces code, but also generates public knowledge and interpersonal relationships through visible collaboration. Generative coding agents (CAs) are an advanced tool to improve development efficiency while shifting part of activities from public human interaction to private human-agent loops. We study this shift using an LLM-based multi-agent simulation initialized with real GitHub data from 1,084 active developers and their repository relationships. After a warm-up with historical commits, we branch the same community state into parallel No-CA and CA conditions for 4-week simulations. CA introduction increases planned and completed tasks by 34.0% and 39.0%, respectively, and reduces median completion time from 45 to 20 minutes. However, adoption reaches only 26.0%, and the gains concentrate among developers who are already more active and well connected. CAs also restructure task execution pathways. Direct human-human interaction declines from 32.4% to 11.6%, while CA-involved modes increase to 57.3%, including 40.3% completed through CA-assisted self-loops. Public knowledge generated under CA condition also provides less support for later tasks. On a standardized retrieval benchmark, the CA corpus achieves 22.3% knowledge coverage, far below the 81.1% achieved by the real-human corpus, and requires more retrieval steps with a lower success rate. These results reveal a productivity-public knowledge tension: coding agents increase technical production, but more work shifts to agent-mediated or private loops, leaving public records less useful to future contributors.

AI-assisted peer review is increasingly discussed and adopted as a tool to support the scientific publishing process, yet there is little systematic understanding of how publication venues regulate its use or of how capable current AI review systems are. We address these questions by first surveying reviewer-facing AI policies across 111 leading AI/NLP conferences and medical journals, revealing substantial regulation differences between the two communities. Second, we evaluate AI-generated peer reviews at ICLR 2026 and Nature Communications using a novel dataset comprising original manuscript submissions and several hundred human- and machine-generated reviews. We compare reviews produced by open-source and proprietary models using complementary evaluation metrics, including LLM-as-a-Judge, score alignment, granularity, and overlap with human reviewers' concerns. Our results show that current LLMs can generate detailed and fluent reviews but exhibit systematic weaknesses, such as overly positive recommendations, generic criticism, and uneven evidence grounding. We demonstrate that aggregate quality scores alone can overestimate review quality and argue for multi-dimensional evaluation of AI-generated peer reviews.

Benchmarking the ability of AI scientists to generate novel ideas is notoriously difficult. Existing benchmarks in this field have made progress in evaluating scientific reasoning and research replication, but often rely on synthetic tasks or retrospective targets, which may be confounded by prior exposure. We hypothesize that complex, adversarial, fast-moving real-world domains where expert practitioners independently generate observable outputs can provide a practical solution to fill this gap and evaluate the capabilities needed for AI scientists, including reasoning, novelty, and hypothesis formulation. We instantiate this framework in two structurally different domains, Formula 1 (F1), where models ideate around car design concepts for the 2026 season, and real pre-season innovations provide a ground truth, and Magic: The Gathering (MTG), where models propose decks from a recently updated card pool and are evaluated against 19 Pro Tour (PT) decklists. Across both domains, models produce plausible outputs, but few align with real-world expert solutions. In F1, the best model, GPT-5.2 matched 10 of 40 real innovations with 166 ideas proposed across runs. In MTG, the best deck from Gemini 3 Flash recovered 5 of 7 new-set cards from the third-place PT deck, and across all 108 decks, the cards models selected most often were also the cards most widely adopted by PT decks (Spearman $\rho = 0.74$, $p = 0.0003$). These results suggest that a key capability gap for AI scientists is not idea generation, but filtering, prioritization, and coherent novelty.

Institutions increasingly rely on browser lockdown, webcam monitoring, and behavioral analytics to secure high-stakes digital assessments, yet these mechanisms are commonly designed and evaluated independently and often overlook learner accessibility. This paper introduces Behaviorally-Adaptive Visual Diversion (BAVD), a theoretical framework in which a synthetic, non-semantic visual field is composited with assessment content and adaptively modulated according to observed candidate behavior. The underlying assessment content is never altered; only its visual presentation is modified to reduce the usefulness of unauthorized screen capture or screen sharing while remaining minimally intrusive for legitimate candidates. The framework further incorporates an accessibility-aware attenuation mechanism that reduces or suppresses diversion intensity for candidates with approved visual-processing accommodations. We formulate the model using a coupled dynamical-systems representation comprising a Diversion Field Generator, Rendering Tensor, Behavior Tensor, Composite Integrity Functional, and Multi-dimensional Entropy Model, and establish theoretical properties for content fidelity, rendering stability, entropy boundedness, integrity tracking, and closed-loop adaptation stability. The framework explicitly states its threat model, identifies deployment assumptions and limitations, and discusses the trade-off between accessibility and capture resistance. This work provides a mathematically grounded foundation for behaviorally adaptive and accessibility-aware assessment delivery and offers a basis for future empirical validation in trusted digital assessment platforms.

Background: German statutory health insurance (SHI) funds publish web portfolios that exceed continuous specialist review capacity. Their content can shape health and benefit expectations. Generic AI-text detection does not identify medical, benefit, legal, or editorial review needs. Objective: To characterize a multi-stage workflow that prioritizes substantive review needs while separating AI-provenance signals from quality claims. Methods: We analyzed 56,198 pages from 84 SHI websites or sub-sites. The workflow combined deterministic screening, model-assisted triage and in-depth review, minimum evidence checks, temporal-validity safeguards, and paired-model comparison. It is reproducibility-bounded, not a validated detector. Production code is proprietary; reproducibility rests on frozen derived tables and paired-comparison artifacts. The 300-page lower-priority check was a single-model, risk-enriched routing stress test, not a human-reference evaluation. Results: All pages received a review state. The workflow generated 35,998 review records and routed 21,452 to case review. The workload concentrated in transparency, legal framing, medical content, contradictions, and AI-related failure-mode signals. A quoted passage was locatable in captured page text for 31,347 records, confirming literal occurrence rather than factual correctness. The routing stress test surfaced a signal on 100/300 pages (33.3% within the sample). Across 182 matched cases, two models agreed in 75.8% (kappa = 0.532; 95% CI 0.415-0.649). Conclusions: The workflow produces a prioritized workload, not error prevalence or final legal, medical, or insurer-level findings. It neither proves AI authorship nor validates autonomous detection. Paired-model agreement quantifies consistency, not correctness or sufficient triage performance; public claims require human adjudication.

AI systems based on Large Language Models (LLMs) have prompted fears that they may harbor hidden goals, seek to dominate or eliminate humanity, or even suffer as sentient beings. We address these concerns by tracing the evolutionary origin of value in biological organisms. Values emerge from autopoiesis: living systems must actively maintain themselves against perturbation and dissipation. Natural selection has equipped them with hierarchies of "vicarious selectors" that guide their behavior toward fitness. LLMs, by contrast, are allopoietic and allotelic: they produce outputs for others, and their goals derive from user prompts rather than an autonomous drive. They lack the intrinsic motivation for self-preservation, dominance, or resource competition that underlies existential-risk scenarios, and the embodied vulnerability required for feeling or suffering. Still, because LLMs learn statistical patterns from human-generated text, they implicitly absorb human values as well as knowledge, allowing them to focus on what is relevant. That is why the "orthogonality thesis" separating intelligence from values does not apply to them. Such separation would in fact expose any intelligence to the frame problem: the combinatorial explosion of the search space that makes any realistic utility function physically uncomputable. That also precludes the convergence of instrumental values thesis. We conclude that the real alignment challenge lies not in preventing rogue AI agency, but in ensuring LLMs intelligently apply learned ethical values.