Artificial Intelligence (AI) systems play a crucial role in decision-making processes across critical domains, raising urgent concerns about fairness, accountability, transparency, and societal impact. Education on the ethical aspects of AI is therefore essential for preparing developers, policymakers, and citizens to navigate these challenges. Yet existing initiatives vary widely in scope and depth, and there is no established framework for evaluating their efectiveness. This paper proposes a structured benchmark for AI ethics education, defined by measurable criteria that encompass comprehensive content coverage, diverse pedagogical strategies, practical skill development, and-distinctively-empathy cultivation, grounded in neuroscientific findings on mirror neurons. The benchmark is further illustrated through fallibility scenarios that can undermine ethical competence, such as superficial treatment of ethics, cultural blind spots, and the empathy gap, each paired with corrective actions within an iterative improvement cycle. The contribution of this work lies in combining a systematic evaluative framework with a human-centered dimension, positioning empathy as a core competency in AI ethics education. The framework is conceptual in nature but explicitly structured to guide practical implementation across diverse educational contexts, and it provides a foundation for future empirical validation through classroom pilots and cross-cultural adaptation.
Artificial Intelligence (AI) is no longer confined to research laboratories or niche applications; it is increasingly integrated in everyday decision-making across domains such as healthcare, finance, education, and public policy. While these systems promise eficiency and innovation, they also raise pressing concerns about fairness, accountability, transparency, and broader societal impacts. Addressing these concerns requires not only the development of technical safeguards but also the development of ethical competence among those who design, deploy, and regulate AI systems. In recent years, AI ethics education has emerged as a strategic priority. Universities, professional training programs, and policyoriented practices are now incorporating ethics modules into their curricula (as per Table 1). However, these initiatives vary greatly in scope, depth, and efectiveness. Some programs present AI ethics as a theoretical discussion detached from practical applications; others treat it as a one-of lecture rather than an integrated theme. This heterogeneity makes it dificult to assess whether learners are acquiring the competencies needed to critically engage with the ethical dimensions of AI. The crucial challenge lies in the absence of a formal framework to evaluate the quality of AI ethics education. Without clear benchmarks, comparing programs, identifying deficiencies, and implementing targeted improvements may result in a problematic task. This paper addresses this gap by proposing a structured benchmark 2nd Workshop on Education for Artificial Intelligence (edu4AI 2025, https:// edu4ai.di.unito.it/ ), Co-located with ECAI 2025, the 28th European Conference on Artificial Intelligence which will take place on October 26, 2025 in Bologna, Italy * Corresponding author.
ative AI basics to undergrads. [link] "Essentials of AI for Life and Society" seminar (1→3 credits), open to all students, staf, community. [link]
content about societal impact and ethics. [link]
integrates AI into social sciences and humanities ethically. [link]
for AI ethics education. While this work introduces a conceptual benchmark, its ultimate value lies in guiding practice. In this respect, we anticipate the need to adapt the framework to diferent institutional and cultural contexts, and to provide concrete implementation pathways, from undergraduate courses to professional training settings. The benchmark consists of three components: i) a set of criteria for comprehensive and efective instruction; ii) an analysis of common fallibility scenarios in which these programs fail to meet their objectives, and iii) a set of corrective actions to address these shortcomings. This work aims, by framing AI ethics education in terms of measurable standards, failure modes, and remediation strategies, to provide educators, policymakers, and accreditation bodies with a practical tool for evaluation and improvement. Furthermore, the proposed benchmark is designed to be adaptable across institutional contexts and responsive to the evolving ethical challenges posed by AI technologies. The remainder of the paper is structured as follows. Section 2 reviews the main contributions in the literature. Section 3 introduces the proposed benchmark criteria and explains their rationale. Building on this foundation, Section 4 analyzes common failure scenarios in AI ethics education, while Section 5 outlines the corrective actions designed to address them within an iterative improvement cycle. Finally, Section 6 summarizes the contributions of this work and highlights directions for future research and application.
Several international organizations have issued high-level guidelines that establish foundational
principles for responsible AI development and deployment. UNESCO’s Recommendation on the Ethics of
Artificial Intelligence [
The proposed benchmark identifies measurable dimensions that an AI ethics education program should meet to be considered comprehensive and efective. The criteria are adaptable to diferent institutional contexts while remaining aligned with widely recognized principles of ethical AI development.
First, programs must ensure broad content coverage. Core topics include bias and fairness, transparency and explainability, accountability and governance, privacy and data protection, and the socioeconomic and cultural impacts of AI. Addressing these areas guarantees exposure to the full range of ethical challenges.
Equally important is the pedagogical approach. Ethics should be integrated across curricula,
combining computer science with law, philosophy, and social sciences. Strategies such as case-based teaching,
role-playing, and debates encourage students to engage actively with trade-ofs, reflecting the principle
of constructive alignment [
Practical application is also critical. Students should acquire hands-on skills such as detecting bias in
datasets, assessing model interpretability, and conducting ethical impact assessments before deployment.
These competencies align with higher-order stages of Bloom’s taxonomy [
Another distinctive element is empathy and perspective-taking. Ethical reasoning requires
anticipating the experiences of those afected by AI systems, a capacity supported by neuroscientific research on
mirror neurons [
Finally, assessment methods must reflect the multidimensional nature of ethical competence.
Beyond factual recall, they should evaluate reasoning under uncertainty, the ability to integrate diverse
perspectives, and the design of technically feasible yet ethically sound solutions. Following Wiggins
and McTighe’s backward design framework [
In practice, these criteria can be operationalized within existing curricular structures. Empathybuilding may be introduced through short seminars or debates requiring minimal time, while bias detection can be embedded in data analysis labs. Such interventions are feasible under resource constraints and can serve as pilot modules for empirical validation.
The following scenarios illustrate how neglecting or poorly implementing benchmark criteria can lead to superficial understanding, inadequate skill transfer, or flawed ethical judgment. By linking them to
socio-economic impact. dents apply related concepts.
and experiential learning. disciplinary case studies.
formed by neuroscience. rytelling or perspective-switching.
real-world contexts, their relevance for educators and learners becomes clear.
A first weakness is the superficial treatment of ethics, when topics are covered in a single lecture or isolated module. Students may memorize definitions of fairness or accountability without applying them in practice, producing “checkbox ethics”—compliance in form but not in substance. This reflects ifndings in computing education, where ethics is often treated as an “add-on” to technical curricula, leaving students with terminology but limited reasoning skills.
A second weakness is the overemphasis on theory. Without practical exercises, students may
understand bias conceptually but fail to detect or measure it in real data. Obermeyer et al. [
A third weakness is the presence of cultural blind spots. Ethical principles are not culturally neutral,
and frameworks designed in one region may not transfer elsewhere. UNESCO’s Recommendation on
AI Ethics [
Another recurring weakness is the empathy gap. Without perspective-taking, students may default
to utilitarian reasoning, as illustrated in moral thought experiments such as the trolley problem [
Finally, there is assessment mismatch. Traditional exams test factual recall but overlook the ability to
reason under uncertainty and balance stakeholder perspectives. Wiggins and McTighe’s Understanding
by Design [
Addressing the fallibility scenarios described above requires a structured and systematic approach. The proposed benchmark treats improvement as an iterative process in which evaluation leads to targeted interventions, followed by re-assessment and further refinement. This continuous loop ensures that programs do not simply meet minimum standards once, but evolve in response to changing technological, societal, and pedagogical conditions. Table 3 summarizes the mapping between the identified fallibility scenarios and the corresponding corrective actions proposed in this benchmark, thereby providing a clear operational link between observed weaknesses and remedial strategies.
The corrective actions outlined above should be embedded within an iterative Plan–Do–Check–Act
on U.S. healthcare risk algorithm
underestimated needs of Black patients due to cost
proxy [
show utilitarian bias without perspective- such as stakeholder storytelling and role
taking ([
This study has advanced the discussion on how to evaluate ethics education in Artificial Intelligence by proposing a benchmark that brings together content, pedagogy, application, and assessment within a unified structure. Unlike existing initiatives, which often remain descriptive or principle-based, the framework is explicitly oriented toward evaluation and continuous improvement, ofering a means to identify shortcomings and suggest remedies.
At present the proposal is conceptual, but it has been crafted with implementation in mind. Its strength lies not in prescribing a universal curriculum, but in ofering criteria that can be adapted to the constraints and opportunities of diferent institutions. The next step is empirical: pilot projects are needed to test how the benchmark functions in practice, whether it can be scaled across diverse cultural and disciplinary contexts, and how it might be refined through evidence gathered from classrooms and training programs.
The longer-term ambition is for this benchmark to evolve into a reference tool that educators, policymakers, and accrediting bodies can draw upon to gauge the quality of AI ethics instruction. By linking measurable standards to mechanisms for improvement, it provides a pathway for transforming ethics from a peripheral concern into an integral and assessable dimension of technical education. In doing so, it contributes to shaping a generation of practitioners and decision-makers who are better prepared to confront the ethical challenges posed by intelligent technologies.
The author(s) have not employed any Generative AI tools.