AI & Society最新文献

Despite significant advancements in AI and automated driving, a robust ethical framework for AV decision-making remains undeveloped. Such a framework requires clearly defined moral attributes to guide AVs in evaluating complex and ethically sensitive scenarios. Existing frameworks often rely on a single normative ethical theory, limiting their ability to address the nuanced nature of human decision-making and leading to conflicting outcomes. Augmented Utilitarianism (AU) offers a promising alternative by integrating elements of virtue ethics, deontology, and consequentialism into a non-normative framework. Grounded in moral psychology and neuroscience, AU employs mathematical ethical goal functions to capture societally aligned attributes. In this study, we propose and evaluate a method to elicit these attributes for AV decision-making. One hundred participants were presented with traffic scenarios, including critical and non-critical situations, and tasked with evaluating the relevance of an initial set of 11 attributes (e.g., physical harm, psychological harm, and moral responsibility) while suggesting additional relevant attributes. Results identified two new attributes—environmental harm and energy efficiency—and revealed that four attributes (physical harm, psychological harm, legality of the AV, and self-preservation) varied significantly between critical and non-critical scenarios. These findings suggest that the weight of attributes in ethical goal functions may need to adapt to situational criticality. The method was validated based on key evaluation criteria: it demonstrated sensitivity by producing varying relevance scores for attributes, was deemed relevant by participants for eliciting AV decision-making attributes, and allowed for the identification of additional attributes, enhancing the robustness of the framework. This work contributes to the development of a dynamic and context-sensitive ethical framework for AV decision-making.

This article investigates the pervasive issue of bias within AI-driven cultural heritage collections, emphasizing how digital technologies both inherit and amplify existing societal and historical prejudices embedded in analogue records. It outlines the multifaceted nature of bias—ranging from data selection and annotation to algorithmic design and user interaction—demonstrating how each stage of the AI pipeline can introduce or perpetuate distortions in representation. Through a critical review of current scholarship and practical case studies, particularly in image classification, the article evaluates technical strategies such as data augmentation, adversarial debiasing, and monitoring plans for bias mitigation. The findings reveal that while methods like noise injection and colour jittering can balance datasets and improve model fairness, effective bias mitigation ultimately depends on interdisciplinary collaboration between heritage professionals, subject experts, and data scientists. The article concludes that addressing bias requires an ongoing, holistic approach, integrating both technical and humanistic perspectives from data collection to model deployment. This ensures more inclusive, accurate, and ethically sound representations of cultural heritage, supporting the sector’s goals of diversity and accessibility for future audiences.

AI algorithms require human input to achieve technological aims. This fact is often overlooked in discussions of autonomous systems and AI safety, to the detriment of both philosophical discourse and practical progress. One potential remedy is to ground our theorizing more fundamentally in the idea that AI technologies are sociotechnological systems with human and artifactual components. In this article, I pursue this strategy, aiming to shift the focus in AI ethics from artifacts and their intrinsic properties—what I refer to as the robotic conception of AI—to the relationships among elements embedded in AI-involving sociotechnological systems. First, I defend the claim that the sociotechnological-system perspective provides an accurate description of some of our most advanced AI. Second, I argue that the dominance of the robotic conception has steered AI safety research down unproductive paths, while the sociotechnological perspective has the capacity to set us right. Specifically, the robotic conception encourages the development of artificial moral agents—whose creation we should avoid if possible—and distracts researchers with hypothetical trolley cases. In contrast, the sociotechnological approach coheres with actual progress being made on AI safety (e.g., networking, shared user-artifact control, and value alignment) and makes vivid solutions to the safety problem that do not require the creation of humnanlike moral decision-makers.

The primary objective of this study is to determine whether the level of digital literacy among pre-service teachers reliably correlates with their anxiety levels concerning artificial intelligence. The study was conducted as a correlational study, with a sample size of 221 pre-service teachers. The study’s population comprised 3922 pre-service teachers enrolled at Turkish state and private universities. To collect study data, the researchers used the “Personal Information Form,” “Digital Literacy Scale,” and “Artificial Intelligence Anxiety Scale.” The data were analyzed using stepwise regression analysis and descriptive statistics. The study’s results indicated that pre-service teachers exhibited high levels of digital literacy and moderate degrees of anxiety regarding artificial intelligence. Regression analysis revealed that 10.3% of pre-service teachers’ anxiety concerning artificial intelligence could be predicted by the technical sub-dimension of digital literacy. Consequently, it was demonstrated that pre-service teachers’ apprehensions regarding artificial intelligence decreased as their perception of technical digital fluency increased. Other sub-dimensions of digital literacy were deemed insignificant in predicting the anxiety levels of pre-service teachers regarding artificial intelligence. Based on these findings, suggestions for future study directions were proposed.

Artificial testimony generated by large language models (LLMs) can be a source of knowledge. However, the requirement that artificial testifiers must satisfy for successful knowledge acquisition is different from the requirement that human testifiers must satisfy. Correspondingly, the epistemic ground of artificial testimonial knowledge is not the well-known and accepted ones suggested by renowned epistemological theories of (human) testimony. Based on Thomas Reid’s old teaching, we suggest a novel epistemological theory of artificial testimony that for receivers to justifiably believe artificially generated statements, testifiers of the statement should robustly perform the propensities of veracity and cautiousness. The theory transforms the weakness of Reid’s view to an advantage of its own. It sets an achievable standard for LLMs and clarifies the improvement that current LLMs should make for meeting the standard. Moreover, it indicates a pluralistic nature of testimonial justification pertaining to the pluralistic nature of possible testifiers for knowledge transmission.

This article situates current deep learning (DL) artificial intelligence (AI) within Leroi-Gourhan’s deep history of the human species’ relation to technology. According to Leroi-Gourhan, technology is both a key element of anthropogenesis and a source of later tensions (or disentanglement) between the human species and its external and increasingly autonomous technics. Human organic (life-oriented) intelligence at first extends itself through technical (machine-oriented) intelligence, only to be later left behind by it. We propose a concept of machine intelligence that goes beyond technical intelligence, the latter a (still) hybrid human–machine intelligence. This new, emerging machine intelligence is DL AI. DL AI developed out of the failure of symbolic AI to instantiate a key generic component of intelligence: creativity. While symbolic AI was rigid and pre-programmed, DL is flexible and unpredictable, presenting an embryonic form of actual machine intelligence. Its creativity can be likened to the ancient Greek concept of metis, a cunning and polymorphous form of intelligence. Although often biased and problematic, DL exhibits a machine creativity that goes beyond the anthropocentric imaginings of AI as a (mechanistic) imitation of the human norm.

People’s concerns regarding the fairness of algorithmic decision-making, coupled with its expanding utilization across various spheres of our lives underscores the need for robust measures to assess perceived fairness in standardized survey research. Existing fairness scales often suffer from inadequate content coverage, particularly in terms of Perceived Group Discrimination, and frequently employ suboptimal measurement methods, such as single-item assessments. This paper introduces the AIM-FAIR scale, a multidimensional tool grounded in classical test theory, employing Likert-scaled answering options and a reflective measurement model. Developed through four studies (n = 1777) and validated in both English and German, the scale includes 17 items across five subscales: Perceived Consistency, Perceived Equity, Perceived Group Bias, Perceived Manipulability, and Perceived (Explanatory) Transparency. Both language versions demonstrate excellent fit indices and consistent measurement invariance across diverse backgrounds, languages, and conditions. The AIM-FAIR scale offers higher ecological validity and a more comprehensive framework for evaluating fairness in ADM, enhancing cross-cultural and cross-linguistic research on AI fairness.