ethics review

To address inconsistencies in ethical standards, protracted review processes, and variable assessment quality in Institutional Review Board (IRB) oversight, this paper introduces IRB-LLM—the first domain-specific large language model designed explicitly for IRBs. IRB-LLM integrates domain-adaptive fine-tuning, retrieval-augmented generation (RAG), and multi-task prompt engineering to establish a dynamic human-AI collaborative decision-making framework capable of semantically modeling ethical texts. It delivers three core functionalities: pre-review screening, consistency verification, and decision support—collectively enhancing both efficiency and standardization of ethical review. Experimental evaluation demonstrates that IRB-LLM reduces average processing time by 32% and improves inter-reviewer consistency in feedback by 27%. The model provides a reproducible, empirically validated paradigm for AI-augmented governance of research ethics.

This study addresses the absence of unified and transparent research ethics guidelines in top-tier security and privacy conferences, which has led to ambiguous review criteria and inconsistent enforcement, thereby hindering the community’s ethical awareness. Through a systematic analysis of ethics policies across four leading conferences over multiple years and semi-structured interviews with 20 researchers, this work presents the first comprehensive account of the evolution of ethical practices in the field, identifying a critical gap in ethics education as the primary bottleneck. Drawing on qualitative findings and principles of community-based participatory design, the paper proposes an innovative framework featuring an inter-conference coordination mechanism and an open Ethics Wiki. It delineates current progress and key barriers to consensus-building and has already launched the Ethics Wiki as an initial step toward collaborative governance.

Small- and medium-sized enterprises (SMEs) struggle to implement conventional medical AI ethics frameworks due to resource constraints and fast-paced development environments. Method: This paper proposes the Scalable Agile Framework for Ethics in AI (SAFE-AI), which deeply integrates ethical governance into agile software development. It introduces a novel scenario-based probabilistic analogy mapping mechanism for responsibility quantification, establishes testable metrics for fairness, transparency, and uncertainty management, and incorporates a lightweight ethics review model to support iterative development. Contribution/Results: Unlike static, compliance-centric approaches, SAFE-AI operationalizes and scales ethical practice through test-driven acceptance criteria, full-lifecycle monitoring, and business-aligned design. Empirical evaluation demonstrates its applicability in organizations lacking dedicated ethics teams, significantly enhancing model trustworthiness and stakeholder confidence.

Current AI ethics assessments are fragmented, focusing predominantly on fairness, transparency, privacy, and trust at the model or output level while neglecting inter-component system interactions, real-world harm contexts, and causal harm propagation pathways—resulting in evaluations disconnected from actual risk scenarios and lacking actionable thresholds. Method: Through a scoping review synthesizing nearly 800 ethics metrics, this study constructs the first four-dimensional relational framework—“System Components–Attributes–Risks–Harms”—to systematically map ethical assessment dimensions. Contribution/Results: The framework uncovers three critical gaps: insufficient system integration, weak contextual embedding, and poor actionability. It advances AI ethics evaluation from isolated metric measurement toward a systemic, traceable, and intervention-oriented paradigm—enhancing regulatory alignment and practical deployment efficacy in industry settings.

This work addresses the current lack of a systematic framework for evaluating ethical risks in data collection practices for large language models (LLMs). It proposes the first quantifiable assessment framework that integrates multiple prominent ethical theories, structuring evaluation around core ethical principles through a set of targeted questions and establishing a scoring system to measure ethical risk. This approach enables systematic, quantitative ethical auditing of LLM data curation processes. By offering a practical tool for assessing ethical compliance in AI development, the framework fills a critical gap in existing research—particularly in the integration of diverse ethical theories and the empirical evaluation of real-world data practices—thereby advancing the responsible development of artificial intelligence.

Current ethical review systems struggle to address structural ethical risks in large-scale, interdisciplinary research due to insufficient capacity, inconsistent standards, and privacy constraints. This work proposes Mirror, a multi-agent framework that integrates normative understanding, an executable rule repository, and a multi-role collaborative deliberation mechanism to support both rapid compliance checks and in-depth committee-like evaluations. Leveraging a newly constructed domain-specific ethical QA dataset, EthicsQA, the authors fine-tune a specialized language model, EthicsLLM, and integrate it with a rule engine and a structured ethical dimension assessment framework. Experimental results demonstrate that the proposed approach significantly outperforms general-purpose large language models in evaluation quality, consistency, and domain expertise, making it suitable for projects ranging from minimal-risk studies to complex scientific endeavors.

While Layer 2 (L2) rollups improve scalability and reduce costs, operator discretion and information asymmetry introduce novel ethical risks. Method: This paper introduces the first systematic ethical risk analysis framework for L2 scaling, proposing a role-based decision-power–risk-exposure classification model. It integrates role modeling, cross-sectional analysis of 129 L2 projects, a manually curated dataset of on-chain events (2022–2025) covering sequencer liveness and transaction inclusion failures, and mechanism mapping. Contribution/Results: We establish an empirically testable ethical risk metric system, revealing critical concerns—including near-universal absence of withdrawal grace periods for urgent upgrades (86%), single-point proposer control over withdrawal freezing (50%), and ethical vulnerabilities in data availability and forced transaction inclusion. We further propose co-designed technical–governance mitigation strategies to enhance accountability, transparency, and user sovereignty.

This study addresses the frequent neglect of environmental impacts in computationally intensive research—such as artificial intelligence—due to ambiguous ethical review policies. It presents the first systematic framework integrating environmental sustainability into the ethical oversight of computational research. By delineating clear review boundaries, establishing evidentiary standards, and developing researcher self-assessment tools, the framework enables institutional ethics committees to effectively evaluate the environmental costs of proposed projects. This approach provides actionable guidance for ethical review processes and encourages researchers to proactively consider the ecological footprint of their work during early design stages, thereby addressing a critical gap in current research ethics frameworks concerning sustainability.

This study addresses the absence of ethical governance mechanisms in clinical multi-agent systems by proposing ETHOS, a modular ethical framework. Designed as a plug-and-play meta-agent, ETHOS integrates seamlessly into existing architectures and introduces a novel three-tier runtime supervision mechanism that combines deterministic rules, contextual review, and ethical critique. This approach translates abstract principles into executable checks within a dynamic feedback loop. Validation in a liver disease decision-support system demonstrates that ETHOS effectively identifies evidentiary deficiencies, intercepts unsafe responses, and appropriately refuses to answer when necessary. Consequently, this work successfully operationalizes high-level ethical principles into deployable and auditable safety safeguards for clinical AI, significantly enhancing decision-making reliability.

研究通过同行评审跟踪AI论文伦理标记的长期影响，发现作者更倾向于修改呈现方式而非改变研究方向。