AI and ethics最新文献

In this paper, we challenge the overreliance on conventional ethical frameworks commonly observed in current AI ethics literature. We begin by surveying the ethical concerns and frameworks that dominate this field. Following this, we categorize and critically review the existing objections to these traditional approaches in terms of conceptual challenges, professional and regulatory challenges, and challenges from practical implementation. Finally, we present three key arguments against conventional ethical frameworks, drawing upon the fact that they fail to preclude ethics washing, stifle innovation in ethical research, and the drastic changes brought about by AI technology in the ethical landscape.

Why do users follow moral advice from chatbots? Arguably, a chatbot is not an authoritative moral advisor, but it can generate plausible arguments. We conducted a large pre-registered vignette experiment (N = 1269) that controlled for the effect that ethical justification has on subjects’ propensity to accept the chatbot’s advice. Furthermore, to study the influence of the source of the advice, we manipulated subjects’ belief of being advised by a chatbot or another human. In our experiment, we find that users did not accept reasoned more readily than unreasoned advice. However, this was also true if we attributed advice to a moral advisor, not a chatbot. Hence, we suggest that advice might offer users an easy way to escape from a moral dilemma. This is a concern that chatbots do not raise, but they may exacerbate it as they make advice on versatile issues easily accessible. We conclude that it may take ethical in addition to digital literacy to protect users against moral advice from chatbots.

Academic philosophers sometimes quip that in the future the only job safe from automation will be that of philosophy professors. However, the current AI revolution has inspired some AI scholars to propose the future establishment of closed-loop AI systems as a type of superintelligent robot that would essentially outperform and replace human scientists Zenil in the future of fundamental science led by generative closed-loop artificial intelligence, arXiv:2307.07522v3, 1-40, 2023; Schmidt in Mach Learn Sci Technol 5(3): 035045, (2024); Kitano in Npj Syst Bio Appl 7: 1–12, (2021). In this paper, I investigate whether, analogously, academic philosophy could be automated by putative, sufficiently advanced future AI, potentially featuring artificial embodiment as robots. To this end, I distinguish two mutually exclusive metaphilosophical conceptions of the nature of philosophy circumscribed as philosophy as a set of propositions (PP) versus philosophy as an activity (PA). Granting AI proponents that – iff artificial general intelligence (AGI), potentially embodied as superintelligent robots is achieved – natural sciences may be fully automated in the future, I argue for the conditional that if PP is true (but not if PA is true), then it is possible that AI can automate philosophy. Additionally, I consider what it would mean to automate philosophy given the current state of LLMs (e.g., the GPT-5 era). Finally, I briefly consider whether it would be preferable for us to have philosophy automated and argue that there are two prima facie reasons why automating philosophy, if possible, might be undesirable: the reason from obsolescence and the reason from ultimate answers.

Artificial intelligence is widely expected to intensify and accelerate creative destruction, resulting in displaced workers, disrupted business models, and obsolete products. Does this violate individuals’ legitimate expectations? If so, it would provide a powerful argument for blocking innovation or compensating its losers. We argue it does not. This conclusion does not depend on rejecting the doctrine of legitimate expectations. Rather, we show that the distinctive features of the market process mean that expectations about markets are rarely legitimate in the first place. As a result, the doctrine of legitimate expectations cannot justify compensating the losers of creative destruction—an implication of particular importance for current debates about AI.

The integration of AI into education is often framed as a neutral or beneficial response to pressures of efficiency, scalability, and personalisation. In this paper, I challenge that framing by examining how educational AI reshapes teaching as a form of labour. Drawing on Karl Marx’s theory of alienation, I offer a conceptual analysis of how AI mediated systems reorganise pedagogical work in ways that risk estranging teachers from the products of their labour, the labour process itself, their species being, and their relationships with students and colleagues. Rather than treating AI in education as a monolithic phenomenon, I differentiate between generative tools, automated assessment, learning analytics, and administrative systems, showing how each participates differently in processes of standardisation, surveillance, and managerial control. I situate educational AI within wider dynamics of platform capitalism, datafication, and audit culture, arguing that alienation is not an inevitable outcome of technology but a contingent effect of ownership structures, governance arrangements, and institutional imperatives. I conclude by outlining policy, design, and philosophical interventions aimed at reducing alienation, while acknowledging the limits of reform within marketised education systems. In doing so, I reframe AI in education as a political and ethical question about labour, authority, and the purpose of teaching, rather than a purely technical innovation.

Identifying hidden biases in AI documentation metadata (model, data, and dataspace cards) is essential for responsible AI; yet this domain remains largely unexplored. The proposed work evaluates four Transformer models (XLNet, DistilBERT, RoBERTa, and ELECTRA) for bias detection across publicly available, synthetic, and custom datasets. On the BABE news corpus, all models achieved 77–80% accuracy, with only ELECTRA exceeding 80% on every metric. To address the absence of publicly available AI-card datasets, we generated synthetic metadata for two use cases (Customer Interaction and Customer Data Uploaded by Organisations) using ChatGPT. Models trained on this synthetic corpus displayed near-perfect scores, reflecting shared stylistic cues embedded in the generated text. To test real-world robustness, we curated a Hugging Face dataset by scraping documentation comments, filtering for bias-related keywords, and obtaining annotations from four independent labellers in a single-blind setting. Partial fine-tuning (zero-shot) evaluations of models trained only on BABE or synthetic data revealed substantial performance drops on this real-world set. To mitigate this cross-domain loss, we introduce a cascaded, full fine-tuning (few-shot) pipeline in which Transformer models are sequentially fine-tuned on BABE, synthetic text, and a subset of the Hugging Face corpus. Evaluation on the remaining portion achieved over 85% across all performance metrics, enhancing precision and generalisation. This study demonstrates the challenges of bias detection beyond controlled or synthetic data and highlights cascaded fine-tuning as a practical, low-resource strategy. Future directions include leveraging evidence fusion methods, integrating cross-attention with bias taxonomies, and adopting dual-encoder architectures to advance bias detection toward more in-depth, knowledge-guided reasoning.

To examine how clinical AI systems establish de facto standards of care through design and administrative mechanisms; analyze power redistribution across stakeholders in EU, US, and UK governance regimes; and propose comprehensive policy tools to mitigate unintended normative effects while enhancing equity, accountability, and public trust in healthcare AI systems. A comparative qualitative document analysis was conducted on regulatory texts (e.g., the EU AI Act, FDA guidance), hospital protocols, and vendor materials (2015–2025) from the EU, the US, and the UK. The NORM5 typology (Nudges, Override friction, Responsibility choreography, Metric coupling, Scripted workflows) was developed through systematic inductive coding, informed by political science theories of institutional drift and bureaucratic power. Data collection involved systematic sampling of 127 documents across three jurisdictions, with thematic analysis conducted using established qualitative research protocols. NORM5 mechanisms subtly shift clinical norms through five distinct pathways, redistributing power toward vendors, payers, and healthcare organizations while systematically eroding clinician autonomy. The EU’s risk-based regulatory framework formalizes compliance burdens but enables systematic oversight; the US’s fragmented incentive structures promote defensive AI adoption patterns; and the UK’s polycentric governance approach supports coordinated policy responses. Analysis reveals significant societal risks including algorithmic bias amplification, professional autonomy erosion, and reduced public trust in healthcare systems. Algorithmic normativity governs healthcare by design, necessitating ethical governance to balance innovation with human-centered values. The NORM5 typology and policy toolkit offer actionable pathways for responsible AI governance.

Along with the rapid development of AI systems, the literature addressing the moral concerns raised by AI—stemming from different directions like computer science, medicine, and philosophy—has substantially grown. In solely focusing on AI ethics principles and guidelines, most overviews of the field adopt a principle-based understanding of these moral concerns. However, as our review illuminates, there is more richness and diversity in the current body of literature than this dominant principle-based approach suggests. Within this vast literature, we identify three approaches by which authors tend to formulate the moral concerns raised by AI: principles, lived realities, and power structures. These approaches can be viewed as lenses through which authors investigate the field, and which each entail specific theoretical sensitivities, disciplinary traditions, and methodologies, and hence, specific strengths and weaknesses. The first “principle-based” approach takes moral concerns to be universal, stable, and fixed principles; which are globally shared, may travel between contexts and are often predetermined. What we call the “lived realities” approach foregrounds the interaction between people and AI systems, focusing on local practices and everyday experiences, generally on a micro-level. Thirdly, what we call the “power structures” approach argues for the need to account for the cultural, social, political, and economic context of AI development, hence human-AI interactions at a macro-level. In bringing together different moral frameworks, traditions, and questions, our structure may serve as a bridge for comparing AI ethics with other areas of applied ethics—considering AI systems are quickly integrated into different spheres of society.

AI is increasingly used in learning environments to monitor, test, and educate students and allow them to take more individualized learning paths. The success of AI in education will, however, require the acceptance of this technology by university management, faculty, and students. This acceptance will depend on the added value that stakeholders ascribe to this technology. In two empirical studies, we investigate the hitherto neglected question of which impact educational ideals have on the acceptance of AI in learning environments. We find clear evidence for our study participants’ conviction that humanistic educational ideals are considered less suitable for implementing AI in education than compentence-based ideals. This implies that research on the influence of teaching and learning philosophies could be an enlightening component of a comprehensive research program on human-AI interaction in educational contexts.

The human propensity for possession and ownership has been a defining characteristic throughout history, influencing the structuring of societies. The advent of digital technology, however, challenges traditional notions of possession and necessitates a redefinition of the concept. This paper aims to explore the resignification of possession in the digital age through a philosophical reflection based on an anthropological analysis. It contrasts two paradigmatic forms of possession: the material collection in the ‘thing’ civilization and the data-based access in the digital civilization. Ultimately, the research shows, analyzing the OnlyFans platform model, how the digital age has fundamentally changed the human mode of possession, demanding a new paradigm in both virtual and augmented reality, while emphasizing the irreplaceable material reference when it comes to the intimate aspects of Homo Sapiens. The transition from material collection to digital access signifies a profound change in human self-perception and social structures and reflects the general cultural and existential changes triggered by technological progress.