AI and ethics最新文献

This study aims to explore the integration of artificial intelligence (AI) in human resources (HR) in depth, focusing on its impact on organizational productivity, ethical considerations, and the emerging challenges that come with widespread adoption through an extensive literature review. AI-driven HR solutions significantly increase efficiency in hiring, employee engagement, and workforce analytics by automating repetitive tasks and enabling data-driven decision-making. However, the study revealed persistent ethical concerns, particularly regarding algorithmic bias, transparency, and data privacy. Additionally, challenges such as inadequate skills among HR personnel, resistance to change, and ambiguous rules regarding AI deployment were recognized as major barriers. The findings highlight the necessity for organizations to connect efficiency enhancements with solid ethical standards and transparent AI governance. HR leaders are encouraged to emphasize skill development, foster a culture of responsible AI utilization, and engage with evolving legal standards. Addressing these factors is crucial for realizing the full potential of AI in HR while mitigating risks to equity, trust, and employee well-being. Employing AI tools in HR procedures without transparent communication can reduce employee trust. Company leaders should prioritize transparency in their AI applications, particularly when it is utilized to monitor employees or impact decisions that directly concern them. Consequently, to emphasize the significance of artificial intelligence in the current business landscape, we aim to explore the challenges and opportunities of implementing AI in human resource management, along with its ethical implications.

This paper asks the question of whether there are performance reasons for leveraging Human Values in the pre-training of Modern Reinforcement Learning (RL) models? As part of exploring this question, this paper looks at Modern RL algorithms generally, and specifically Model Based and Model Free RL. It then examines the treatment of Rewards in such model types as the Solution Concepts of: Common Reward Games, Zero Sum Games, and General Sum Games (incl. Nash Equilibriums). The Value Alignment Problem is then described as a result of a Zero Sum Game Solution Concepts and actual examples of this problem are provided. The paper then goes on to explore how Ruchard Sutton and Stewart Russell propose to address the Value Alignment Problem. Finally, the paper examines the possible use of Supervised Learning to effectively pre-train Modern RL algorithms to address the Value Alignment Problem, and cites the success of the AlphaStar algorithm as an example how pre-training with Human Values may have technical benefits.

Artificial Intelligence (AI) is transforming recruitment by allowing organisations to employ data-driven hiring decisions. However, AI-powered tools are prone to reinforcing biases instead of eliminating them, posing ethical and fairness concerns. Since AI-powered tools are critical in hiring, this study introduces HITHIRE, an AI-driven recruitment model designed to enhance transparency, fairness, and inclusivity within the diverse workforce of Saudi Arabia, which aligns with Vision 2030. A baseline model was first evaluated using Llama 3.1, BERT, and regular NPL techniques. However, the baseline model revealed significant biases in gender and nationality-based hiring, making it inappropriate for the Saudi Arabian diverse hiring environment. The Llama 3.1 model was enhanced through data augmentation, sentence transformers, standard scoring, and transparency mechanisms, resulting in the HITHIRE model. Fairness analysis demonstrated improvements across gender and nationality dimensions, with reduced Statistical Parity Difference (SPD) and Disparate Impact (DI) scores. The findings highlight the potential of ethical AI integration in recruitment, ensuring unbiased, accountable, and transparent hiring practices. HITHIRE sets a precedent for AI-driven fairness in recruitment, contributing to HR policies and ethical AI discourse globally.

This study investigates whether AI-powered Low-Code/No-Code (LCNC) solutions may unintentionally generate gender-biased responses. We developed four AI-powered LCNC implementations (i.e., Spreadsheet-based, Workflow-based, Web-Application-based and Mobile-Application-based), using different generative AI models, including those from OpenAI, DeepSeek, Claude, and Google DeepMind, and evaluated their outputs in response to prompts designed to highlight potential gendered associations in roles, traits, and personal preferences. Our analysis consists of two parts. First, we applied a mixed-methods structured content analysis to systematically identify potential stereotypical patterns in the responses of the AI models. Second, we compared the outputs across the different AI models for each prompt to explore variations in gender bias-related behavior. Our findings raise an ethical concern: without appropriate policies and guidelines in place, AI-powered LCNC solutions may replicate or even amplify existing societal biases. This work contributes to ongoing discussions on responsible AI integration and bias-aware design, especially within the evolving LCNC ecosystem.

This study evaluates the reliability, completeness, and thematic consistency of responses generated by legacy versions of ChatGPT—3.5 (2023) and 4.0 (2024)—across five autism-related domains: Diagnosis, Prognosis, Prevalence, Evaluation, and Treatment. The study provides a historical snapshot of model performance while highlighting broader patterns of accuracy, limitations, and implications for ongoing use of large language models in autism-related contexts. Sixty-nine questions across five domains were presented to both ChatGPT-3.5 and ChatGPT-4.0. The responses were evaluated for accuracy, length of explanation, completeness, and thematic consistency. Comparative analyses incorporated both descriptive and inferential statistics, and thematic patterns were examined using cosine similarity. Both ChatGPT-3.5 and ChatGPT-4.0 exhibited high accuracy, particularly in structured domains such as Diagnosis and Treatment. ChatGPT-4.0 provided slightly richer descriptive detail in more complex areas, though this occasionally introduced thematic variability. Completeness scores were moderate across domains (ranging from ~ 0.20 to 0.50), reflecting that responses often captured some, but not all expected key points. A limited consistency checks with ChatGPT-5.0 (September 2025) demonstrated broad stability of conclusions, with no decreases in accuracy and only minor updates observed in prevalence-related answers. ChatGPT-3.5 and ChatGPT-4.0 show substantial potential as supplementary resources for autism-related information. Although these models consistently provide accurate content, their moderate completeness underscores the need for human oversight to ensure quality and comprehensiveness. Understanding their reliability, limitations, and evolving nature is essential as large language models continue to be explored for healthcare education and patient-facing information, though not as replacements for professional training or clinical care.

Explainability is a leading solution offered to address the challenge of AI’s black boxing. However, a lot can go wrong when trying to apply explainability, and its success is far from certain. Moreover, there is insufficient empirical data regarding the effectiveness of concrete explainability efforts. We examined an explainability scenario for an AI decision support tool under development for the early detection of cancer-related cachexia, a potentially fatal metabolic syndrome. We conducted 13 interviews with clinicians who deal with cachexia, and asked about their prior experience with AI tools, their views on explainability, and presented an explainability scenario based on the Shapley Additive Explanations (SHAP) method. Most clinicians we interviewed had limited prior experience with AI tools, and a majority of them believed that the explainability of such an AI system for the early detection of cachexia is essential. When presented with the SHAP explainability scheme, they had limited familiarity with the features that contributed to the tool’s ruling, and only a minority of the clinicians (nuclear medicine experts) stated that they could utilize these features in a meaningful manner. Paradoxically, it is the clinicians who come in contact with patients who cannot make use of this specific SHAP explanation. This study highlights the challenges of offering a hyper-selective explainability tool in clinical settings. It also shows the challenge of developing explainable-by-design AI systems.

The rapid expansion of artificial intelligence (AI) across sectors brings significant benefits but also substantial risks, such as bias, discrimination, and lack of transparency. Mitigating these risks requires AI governance frameworks that ensure ethical and responsible use. While existing studies highlight strategies and ethical guidelines, comparative analyses of emerging responsible AI (RAI) frameworks, standards, and regulations remain limited. This study aims to fill this gap by employing a rapid review methodology to examine 17 responsible AI frameworks, standards, and regulations which we named as AI policies throughout this research, from diverse regions, including Singapore, the US, the UK, Canada, Hiroshima, and Australia, and global organizations including the Organization for Economic Co-operation and Development (OECD), and International Organization for Standardization (ISO). This research aimed to address four primary questions on identifying global and local AI policies, identifying and analyzing their key features, assessing implementation challenges, and determining the essential components for designing an integrated AI governance framework. There are eleven key features identified, including RAI Principles, Stakeholders, Stages (AI software development life cycle), Targeted audiences, Scalability, Enforce-ability, Resource Intensive, Region, Technology, AI governance practices (Prerequisites, outcomes, implementation tools or guides), and AI governance area. The comparative analysis highlighted that while the AI policies offer detailed implementation guidelines, they differ in their approaches, mandatory nature, scalability, and resource demands. These differences are critical for organizations seeking to implement these policies effectively. Challenges related to resource intensity, scalability, governance practices, and ambiguous targeted audiences were noted as significant barriers to successful adoption. Based on the analysis, key components for an RAI framework were proposed, and categorized into qualities (scalable, extensible, adaptive, efficient), dimensions (scope, context, implementation practices), and governance practices (prerequisites/outcomes, resources, governance steps). These components aim to guide organizations in developing AI governance frameworks.

As artificial intelligence becomes increasingly embedded in academic life, it is beginning to reshape the cognitive, ethical, and institutional foundations of scholarly practice. This paper argues that large language models and automated research tools do more than enhance productivity, fundamentally altering how ideas are generated, valued, and legitimized. We conceptualize this transformation as a shift from the “natural selection of skills” to a “natural selection of minds,” where academic success increasingly depends on the ability to frame generative questions, synthesize across disciplines, and ethically navigate human–machine collaboration. Through interdisciplinary case studies and critical analysis, we highlight both the promises and the risks of this epistemic transition, emphasizing the need for reflection and potential reevaluation of academic standards to align with the realities of hybrid intelligence and cognitive inequality in the AI era. While our analysis captures dynamics most visible in fields shaped by publish-or-perish pressures and formalist conventions, it may not generalize equally across all disciplines.

AI systems increasingly mediate high-stakes decisions and expose value conflicts that fixed principles or design-time value lists cannot settle. This article proposes the Meta-Layered Framework (MLF): a diagnostic, governance-first architecture that renders pluralism and persistent contradiction procedurally governable The MLF distinguishes six non-substitutable layers—structural, epistemic, relational, political, ontological, and reflexive—and coordinates action through a simple grammar: type → route → review. The accompanying Ethical Tension Map operationalises this grammar by linking each live dispute to layer-proper evidence, a competent forum with standing and remedy (including pause/rollback), and a fixed/triggered review cadence. A comparative analysis clarifies how this approach diverges from Value-Sensitive Design, Explainable AI, and process ethics, which surface concerns and improve legibility but under-specify standing, authority, and cadence. An illustrative healthcare case (MedIntra) shows how the framework converts bedside reasoning, disparity control, and sacred-role boundaries into auditable routines. The method is intended for AI designers, governance bodies, and policy-makers, and presupposes minimal assurance conditions: separation of build and adjudication, named authorities with halt powers, and proportional documentation. Its value is procedural: decisions are made provisionally, monitored against thresholds, and altered or unwound when evidence demands.

While interest towards AI ethics grows, it remains challenging for companies to implement in practice. Some existing studies have sought to understand the current state of practice in AI ethics and the challenges companies face while trying to take ethics into account. However, there is still work to be done in this regard. Specifically, we consider the viewpoint of roles in the practical implementation to be underexplored. In other words, "who" should be doing AI ethics? We conduct an exploratory multiple case study to understand how three companies tackle AI ethics, especially in terms of roles. We conduct semi-structured interviews with respondents working in both management and developer roles. The resulting interview data is analyzed using thematic analysis. Our results showcase different ways in which AI ethics can be approached as far as roles are considered. Our case companies employ different approaches to AI ethics and, in addition, also discuss likely future scenarios for their organizations, as well as how they perceive things should be ideally done. In addition, we present a framework that highlights four general ways to approach AI ethics in terms of roles and responsibilities. This study furthers our understanding of how AI ethics is approached in practice in terms of organizational roles and processes, serving as an initial exploration on the topic. The framework we propose provides a tangible way to start considering roles related to AI ethics in both research and practice.