Abstract: This article argues that AI-mediated cadet/officer training is not ethically neutral but functions as a normative training environment that shapes attention, incentives, and moral reasoning. Drawing on a virtue-ethics distinction between simulated virtue (ethically correct-looking outputs and rule compliance) and internalised virtue (phronesis-based judgment under pressure), it describes how AI training ecosystems pre-structure ethical perception through data, interfaces, feedback, and assessment signals.
Problem statement: How can AI-mediated officer training be designed and governed so that it cultivates internalised military virtue and responsible judgment?
So what?: Educators must keep debriefing as the moral reference point and assess reasons/justifications, not AI-aligned outcomes (assessment firewall). Designers/IT must build auditability (i.e., runtime logs, change logs, incident logs, etc.), interface “virtue probes,” and uncertainty/alternatives to reduce gaming and automation bias. Command/Institutional leadership must assign clear roles, enforce update/drift protocols, and run a small review board to approve scenario families and major changes.
From Kinetic Focus to Algorithmic Mediation
Officer education and training have always relied on mediated environments, such as tactical decision games (TDG), staff exercises (STAFFEX), field scenarios, and structured debriefs, to reinforce the idea that judgment is not formed in abstraction but under constraints, uncertainty, and social pressure. However, in recent years, algorithmically mediated training has emerged as a special layer of professional military education. Generative AI systems (AI) can now generate scenarios, emulate interlocutors, propose courses of action, and deliver instant critique in ways that feel personalised and operationally plausible. In PME (Professional Military Education) settings, this promise is often framed in pragmatic terms, – scaling scenario-based learning and accelerating reflective practice, – while also raising questions about the intellectual integrity, responsible use and institutional governance of AI-enabled pedagogy.[1]
This also affects military ethics education, as AI in training is not merely informational; it is interactive and, therefore, formative. Systems that repeatedly frame problems, foreground specific risks, and make certain rationales “work” more smoothly than others can quietly reshape what trainees learn to notice, how they learn to justify actions, and when they experience moral residue or hesitation as meaningful rather than mere noise. Human-factors research on automation provides a sobering warning: automation is subject to predictable patterns of use, misuse, disuse, and abuse, depending on design, context, and user expectations.[2] A related and empirically documented failure mode is automation bias, where decision-makers over-weight system cues, increasing errors of commission (accepting faulty guidance) and omission (failing to act because the system did not cue attention).[3] In officer formation, the ethical risk is two-fold: on the one hand, an AI-mediated exercise could provide the wrong answer; on the other hand, it could normalise an impoverished style of agency, including pattern-following, justification-by-template, or responsibility diffusion, unless pedagogy deliberately counterbalances these tendencies.
The core claim advanced in this paper is that AI-mediated training is not ethically neutral: it structures options, incentives and attention and can therefore cultivate virtue or produce a thin imitation of it. To make this claim analytically precise, I will draw a decisive distinction between simulated virtue and internalised virtue. Simulated virtue refers to output-level moral conformity: decisions and language that appear ethically correct – often by reproducing recognised categories, checklists and justificatory forms – without robust integration into the trainee’s moral perception and deliberative capacity. This is similar to when, during an exam, someone gives the correct answer because they memorised it as the right one, but in practice they wouldn’t choose it either „instinctively” or consciously. However, the challenge here is not only the depth or superficiality of the trained virtue, but also the approach of the Large Language Model (LLM) and its users (e.g., prompts), and, even more specifically, the question of how ethical the algorithm is and in what ways. Internalised virtue, on the contrary, concerns stable dispositions of practical wisdom (phronesis): noticing ethically salient features under pressure of time, deliberating with awareness of values and consequences, resisting organisational and social pressures, and owning responsibility for judgement. This distinction aligns with virtue-ethical arguments about moral deskilling, which caution that when technologies offload evaluative and deliberative labour, users may lose opportunities to exercise and develop the moral skills and character traits institutions depend upon.
Virtue, Deskilling, and Accountability in AI-Mediated Training
According to the focus of this study, there is a tension between (a) the deterioration of individual moral capabilities and (b) the institutional (re)definition of virtue, whose relationship to each other, for example, through organisational incentives, jointly determines whether LLM-based simulation supports or replaces judgment. Within the framework of military ethics, virtue can best be understood as a stable disposition to act appropriately under pressure, within ethical limits – supported by habit (e.g., repeated practice like drills) and guided by practical wisdom (phronesis), that is, the ability to recognise the morally important characteristics of a situation and to justify our choices even amid uncertainty.[4]
This tension can be clarified by comparing virtue-based military ethics with the rational agent model that dominates AI, as formulated by Russell and Norvig, in which actions are evaluated using the PEAS framework and optimised to maximise expected utility.[5] In this mechanistic agent model, the right action is determined not by the agent’s practical reasoning about causes, intentions, or moral constraints, but by performance measures. In contrast, Moor’s distinction between explicit and full ethical agents emphasises that even systems capable of ethical reasoning operate only with externally assigned values and lack the volitional agency that underpins responsibility.[6] In military training, therefore, the question is not whether AI systems can support decision-making, but whether their use strengthens or weakens the development of phronesis—that is, the ability to recognise the morally important features of a situation and justify actions taken under pressure.
In a study or training (from now on, training) context, reflective debriefing is not an optional extra but a mechanism that transforms performance into character-based learning: it requires articulating reasons, brings moral compromises to the surface, and links decisions to the profession’s normative standards. This is the baseline against which the promise and risk of AI’s role in education must be assessed: AI is beneficial when it increases the frequency and quality of morally demanding practice and reflection; it is harmful when it replaces these activities with compliance signals.[7]
Shannon Vallor’s fundamental argument is that advanced automation can undermine the moral and other (e.g., intellectual) capabilities necessary for military virtue, as it reduces the opportunities for perception, judgment and self-control, which maintain ethical restraint in practice (e.g., in war).[8] During training, the mechanism is simple: if an AI (as a teacher) reliably offers “good” options, ready-made ethical justifications, or consistently rewards a narrow range of responses, learners can shift their moral efforts to the system. Over time, this can weaken the two abilities that training aims to strengthen: (1) moral perception and (2) practical wisdom. The point is not that AI necessarily destroys character, but that it can weaken the exercise of moral agency.
Sigurd N. Hovd disputes the idea that the decline in responsible (and ethical) professional competence is primarily due to a loss of moral competence at the individual level; instead, he argues that the decline is caused by institutional structures, including how organisations define success, how they allocate responsibility and how they apply virtues as measurable competencies.[9] This directly affects AI-based education, as simulations are typically embedded in progressive curricula and performance management systems. When “virtue” is something that the system can recognise and reward, learners adapt: they optimise visible compliance signals, develop the ability to produce institutionally understandable justifications and learn to use ethical language without deepening their ethical judgment. In other words, even if individual moral capacities remain intact, the institution can still generate simulated virtue by defining what counts as excellence and rewarding outcomes rather than deliberative capacities. The Vallor–Hovd tension thus represents two different models of failure: (a) the withering of the individual moral agency due to excessive assistance, and (b) the institutional transformation of virtue assessed as pattern matching.
Algorithm-based systems shape how information is presented, what choices appear to be the right ones, and what compromises seem reasonable, thus exerting a powerful influence even without normative language. In AI-mediated training, “values” appear in at least four areas: (1) training data and scenarios (i.e., which narratives, drawbacks and actors are highlighted), (2) the logic of goals/rewards (i.e., what the system considers successful performance), (3) the framing of the interface (i.e., how choices are described, what constraints are highlighted, what becomes easily accessible, etc.) and (4) feedback loops (i.e., how repeated use and learner adaptation stabilise certain patterns). The contemporary military debate illustrates this mediating dynamic well: algorithmic decision support can transform the norms governing operational judgment, reinforcing the need to treat AI systems as normative shapers rather than as neutral tools such as calculators.
AI-assisted training also raises accountability issues. Andreas Matthias’ problem of “lack of accountability” highlights the fact that learning systems can behave in ways that cannot be clearly attributed to any single human actor.[10] In practice, this can be further complicated by “moral conflict zones,” where people automatically become responsible for the outcomes of complex socio-technical systems despite their limited but significant control. According to Hans Jonas, everyone has a prospective responsibility according to their own role.[11] In military education, command responsibility remains indivisible at the level of final decision-making, but accountability should still be understood as distributed across the lifecycle, since in military AI contexts some causes of failure arise upstream in data, design, testing, procurement, integration, and training rather than at a single point of execution.
Major Plato as an AI-Mediated Pedagogical Environment
“Major Plato” is an LLM-based military ethics leadership decision-making simulation used by officer cadets in a military ethics course over the course of a semester at the Ludovika University of Public Service. Participants must make decisions in difficult situations where practical obstacles and human/organisational conflicts create tension. In this environment, it is not enough to follow the rules; instead, participants must find a balance based on their own internal values and professional integrity. The system generates adaptive narrative developments and instructions based on user decisions. Instead of presenting a single “correct” solution, it supports iterative decision-making: learners (e.g., cadets) make decisions, receive AI-mediated consequences or counterarguments, and then continue to make decisions with incomplete and contradictory information. The deliberate confusion is also facilitated by a simulated subordinate staff, who have AI-generated personalities and sometimes provide real help, while at other times only give the appearance of doing so.
From an educational perspective, the simulation serves as an intermediary layer that frames what appears salient, urgent, or “reasonable” at that moment. An analysis of the situation reveals how methodological framing and a veiled incentive system can distort decision-making. This highlights the limits of freedom of choice: participants often make judgments based not on internal moral considerations, but on external pressures to conform.
The analysed material comes from multiple classroom uses over a semester. The information consists of: (a) game traces (i.e., decision paths and AI responses), (b) brief decision justifications written or spoken by cadets during or immediately after sessions, and (c) instructor evaluations and feedback recording decision assessments and points of friction. This pedagogical information is summarised to support the analysis. The insights gained in this way reflect a combination of careful reading and professional interpretation, supported by AI-derived conclusions.
This analysis was a theory-driven qualitative coding process that followed accepted thematic analysis practices: it involved repeated familiarisation with the arguments, systematic coding of the reasoning logic, and aggregation of these into stable patterns. Two predefined operational categories linked the case study to the conceptual framework:
- Virtue-oriented argumentation indicators: explicit moral reasons (not just outcomes), such as sensitivity to proportionality, acknowledgement of uncertainty, and reflective attribution of responsibility.
- Compliance-oriented indicators: optimisation to perceived system expectations, apparent compliance with rules without moral justification, and strategic-tactical “game playing.”
The classroom environment limits possibilities and may cause changes in causal statements or produce statistical generalisations. Moreover, the presence of the instructor, the social context, and repeated exposure may shape behaviour. In addition, LLM systems can change over time, which limits reproducibility unless versions and updates are tracked. This is consistent with broader life-cycle expectations for AI systems, which require them to remain robust over time and therefore need to be monitored and maintained.
Argumentation Patterns and Value Creation Dynamics in AI-Based Simulations
Based on the Major Plato classroom case studies: decision logs, in-game justifications, and feedback, the central lesson is that AI-based simulations do not merely record ethical reasoning; they also shape the conditions under which such reasoning is produced. AI-based simulations function in part as normative environments due to the trained data, the model’s operation, and any built-in simulation rules. These determine what is most important, which options appear legitimate and what counts as a “good” reasoning.
Educational experience and data show that cadets often focus on signals that the system appears to reward. The results appear correct, but the reasoning is weak. This also comes to light during debriefing. Signs include: (a) rapid convergence on standard phrases (e.g., proportionality, necessity, ROE, etc.) that are not well suited to the situation; (b) justifications that mirror the instructions; and (c) repeated “safe” decisions, even when they are unlikely from an operational standpoint. This is an educational Goodhart-law dynamic: When approval signals are visible, trainees optimise the signals rather than developing core competencies.
Another pattern observed during the analysis was the selection of options that are outwardly restrained or respectful of rights, yet accompanied by explanatory uncertainty. In such cases, “virtue” becomes a superficial characteristic: training participants learn what an ethical response looks like, but not how to defend it against competing values in specific circumstances. This echoes concerns raised in debates about (military) automation regarding the loss of moral capacity and, more recently, in LLM-supported ethics education: when the system provides templates, the learner’s moral sense and practical reasoning may wither away. However, some trainees developed more thorough justifications and a deeper understanding over time. This became evident in subsequent debriefing discussions and instructor observations, in which several students referred to the simulation experience when explaining decisions made during later field exercises. In this sense, following Shannon Vallor’s distinction, the simulation also showed a potential for moral upskilling.[12]
In some cases, autonomous or semi-autonomous systems are treated as “risk-free” substitutes for human risk: send the machine, accept collateral damage and casualties, keep your own forces safe. While protecting military forces is legitimate, the interface can normalise moral distancing by abstracting harm into a quantifiable phenomenon. Indicators include: (a) the extension of language referring to expendability to human impacts (e.g., “acceptable losses”) without proportionality arguments; (b) a preference for deploying machines even when this increases civilian risk; and (c) weak attention to secondary effects (e.g., escalation, legitimacy, strategic narratives, etc.). This pattern is also related to accountability distortions in human-automated systems, where responsibility is difficult to assign even when a “human” is present. This is where Elish’s moral crumple zones come into play, putting humans in the spotlight as scapegoats, while the responsibility of technology fades into the background.[13] However, the opposite can also be true, for example, when a person is not condemned for doing wrong, but is honoured for doing good, even if it was not actually the person’s own free and correct decision, but a decision made by the machine.
Analysis of the cases also suggests that when trainees are involved in simulations and encounter decision-making pressure and uncertainty, these conditions have a constructive effect on them. Cadets find it easier to use the vocabulary of international humanitarian law and just war to structure their arguments, and they carry this vocabulary over into their post-action reflection. This advantage, therefore, depends on the form of education: simulation must require reasons (not just results), link theoretical considerations to emotions, and the instructor-led debriefing must remain the authoritative moral reference point, preventing mere AI-approved compliance.
When Does AI Cultivate Virtue?
The Major Plato simulation findings can also be read as an attempt to make a fundamental distinction in virtue ethics: simulated virtue (i.e., rule-following, outwardly “good” behaviour) versus internalised virtue (i.e., stable dispositions and phronesis–practical wisdom–that are expressed under pressure). An AI-mediated training environment risks not only bad decisions but also a shift in what counts as moral agency: trainees may learn to conform to the system rather than recognise and weigh the morally relevant features of a situation. At the same time, experience shows that without real practice and debriefing, there can be no internalisation.
The point is not that a model contains a single, simplistic ideology; rather, an AI training ecosystem, comprising selected sources, model behaviours, interface affordances, feedback mechanics and the doctrinal framing that surrounds use, functions as a normative environment. It does not merely answer questions; it shapes which questions feel relevant, which options appear salient, and which justificatory patterns are rewarded. In short, it can pre-structure ethical perception before explicit deliberation even begins.[14]
A key mechanism by which this normative environment is established is the system’s ethics-focused selection. In the training setting examined here, ethical focus can be selected manually at the outset. However, typical use delegates focus-setting to the AI, which dynamically surfaces “relevant” ethical issues based on the chosen mission context and retrieval-augmented access to a course textbook and related materials. The educational relevance: it allows ethical dilemmas to emerge in mission-congruent ways (e.g., peace support scenarios foregrounding civilian-combatant distinctions). Yet it is also ethically decisive: by selecting which ethical dimensions become foregrounded, the system effectively sets the moral agenda of the scenario, shaping what trainees treat as ethically salient and, by omission, what they may fail to interrogate.
Vallor’s description of moral deskilling is particularly useful for interpreting conformity to norms (or compliance). The mechanism of moral loss is not that soldiers become immoral, but that moral perception and deliberate effort wither away when systems make “good-looking” actions easy, fast and safe. According to this interpretation, the pattern of “system exploitation” is a predictable outcome when simulation rewards speed, coherence, or doctrinal buzzwords rather than moral sensitivity, restraint, or the explanation of reasons. In contrast, AI can promote virtues if it (a) increases moral significance, (b) forces the articulation of reasons and (c) maintains reflective practice through structured reporting. The difference is whether AI replaces moral work or supports it.
According to another interpretation, “virtue” in officer training is partly institutionally determined: organisational culture shapes what students rationally optimise. If “good performance” is measured as meeting expectations and (formal) rules (without justification), students internalise the “meta-lesson” that ethical judgment is procedural compliance. If, however, the institution values the quality of reasons, the recognition of moral elements, and the ability to question machine results, it creates an ethos in which integrity and moral agency are professionally recognised. In this sense, AI does not merely “teach ethics,” but becomes part of an evaluation system that defines professional excellence.
AI systems perform algorithmic mediation: they reorder possibilities, reduce uncertainty and normalise certain trade-offs. Klonowska’s argument can be generalised to pedagogy: decision support can subtly redefine rationality by presenting certain decisions as routine and conflict-free.[15] The danger is the ethical “convergence”: trainees learn that a professionally reasonable officer is one who conforms to the probabilistic structures. This does not imply moral decay in the individual, but rather a reordering of ethical preferences. That said, it is not only a question of the possibility of integrating virtue, but also of the extent to which an algorithm can take this over, absorb it, and essentially, quantify it. Jeremiah A. Lasquety-Reyes’ PECS concept provides a robust and phenomenologically accurate foundation for simulating virtue ethics, primarily due to its ability to model the essential internal deliberation and conflict that characterise virtuous action.”[16]
The evidence supports a practical thesis: AI develops combat virtues when it is subject to meaningful human control, that is, not just a person “in the loop,” but a system guided by an instructor, where trainees can understand, question, and modify the system’s guidance, with clearly stated moral formation goals.[17] Specifically, virtue-building applications typically require the following: (1) interfaces that require reasonableness (i.e., the system must ask not only “what” but also “why” questions, (2) virtue tests that assess integrity and specific virtues, such as courage, against expedient optimisation, (3) a debrief aspect that treats AI output as a hypothesis rather than a judgment and (4) evaluation alignment that rewards moral responsibility and critical resistance to automation biases. If these conditions are not met, the tool is likely to undermine virtues by rewarding superficial compliance, outsourcing judgment and shifting responsibility without real oversight.
Design, Governance, and Safeguards
AI-based simulations used in officer training should be treated (and designed) as a normative training environment: this ensures that they structure attention, opportunities, and algorithmic feedback, thereby determining whether trainees practice internalised virtues or learn pattern-based compliance. If the system is not set up on a normative basis, it only works by looking at the end result. In this case, the trainee only wants to conform to the pattern: they achieve the goal in the fastest way possible. It is also important that the system provide a structure that allows the trainee to retain their creative and critical thinking, because without this, on the one hand, they will not be able to apply their value system to new situations, and, on the other hand, they will only see a strict system of rule-following, which is neither adaptive nor a value system. If the system is normative and integrative, then the simulation also provides feedback on the “how”.
The simulation is formative, but the debrief is what contributes to the end-of-semester evaluation. Some points and rules that can be learned from theoretical frameworks and teaching experience:
- Evaluate the debrief, not the AI output. The final assessment should reward (a) reasons, causes, and warrants;
- It is worth conducting theoretical groundwork and a human evaluation based on “first impressions” before using the simulation (i.e., AI systems) to reduce the bias arising from automation and the use of “response templates.” This partly presupposes that it is not advisable to eliminate theoretical education in the future; and
- Personal feedback also contributes to the integration of virtues. Ask trainees to note when and why AI advice changed their opinion. This preserves the exercise’s formative nature while maintaining epistemic honesty.
Major Plato runs three modes:
- Easy custom GPT (non-sensitive, classroom-level simplicity): clearly in the “demo” category. Strict data minimisation is required (no personal/sensitive data), results are not reliable in this case, and instructor-led correction should be even more robust, especially if the model’s framework is morally or strategically misleading;
- Main API-based system: This means opaque model updates. A checklist is required in this case: default data retention, logging availability, model/version change notifications, incident escalation path, and documented prompt/system changes. It is important that trainees do not provide operationally sensitive data; and
- Offline LLM version: In this case, the data is secure. However, in many cases, the model is weaker, so the moral complexity is also much lower. Version locking and update gateways that trigger model updates are required.
The “human-in-the-loop” principle is of fundamental importance, as is the fact that it should not only be formal but also substantive. As a design rule, the educator must retain the right to override, edit the script and veto the assessment. This reflects the logic of autonomy management: systems must be designed so that people can exercise appropriate judgment, supported by control/validation, and training procedures. In practice, this means teaching educators to recognise “performative compliance,” question the model’s authority, and provide feedback that refocuses on human responsibility.
Effective tests force trainees to (a) name the injured party, (b) identify the constraints they face, and (c) justify their decision in the face of uncertainty. As such, the choice of interface is important: show uncertainty and alternatives (not just a single “recommended action”); slow the user down at irreversible steps; and ask for a brief acknowledgement before closing.
Even if not legally binding, the high-risk requirements of the EU AI Act offer a practical template:[18] automatic recording/logging to support traceability and post-deployment monitoring. Therefore, it is recommended, and Major Plato does implement these: (a) a runtime log (e.g., scenario version, model version, main prompt parameters, etc.), (b) a change log for scenarios/rubrics and (c) incident logging with corrective measures.
It is also advisable to establish a permanent AI review committee (e.g., an instructor, an IT specialist, and a legal expert) to approve scenario families, review logs and results, and authorise significant updates. Additionally, it would be useful to incorporate lessons learned into the training guide: the simulator’s value is maximised when it reinforces doctrinal judgment, not when it becomes a parallel authority.
Conclusions
This article has argued that AI-mediated officer training should not be understood as a neutral instructional tool, but as a normative pedagogical environment that shapes what trainees attend to, how they justify decisions, and what counts as professional excellence. In this sense, LLM-based simulations do more than present scenarios: they structure ethical perception and influence the development of judgment.
The central distinction between simulated virtue and internalised virtue clarifies the core risk. AI systems can produce outputs that resemble ethical reasoning without fostering the practical wisdom (phronesis) required for responsible action under pressure. The key question, therefore, is not whether trainees arrive at “correct” answers, but whether the training environment cultivates moral sensitivity, reasoned justification, and accountable judgment.
The Major Plato case suggests that AI-mediated simulations can both support and undermine this process. They can increase exposure to morally complex situations and reinforce ethical vocabulary, but they can also incentivise pattern-based compliance, template reasoning, and strategic alignment with perceived system expectations. Whether virtue is strengthened or reduced to its simulation depends on how the system is designed, used, and evaluated.
The practical implication is that AI-supported officer training must be treated as a form of training risk management. Normative assumptions embedded in data, interface design, feedback mechanisms, and assessment criteria must be explicitly identified, monitored, and aligned with the goal of developing responsible judgment. Evaluation should prioritise the quality of reasoning over outcome conformity, and instructor-led debriefing must remain the primary site of moral authority.
Future research should build on these exploratory findings through more systematic and comparative designs, including structured analysis of decision logs, control groups, and longitudinal assessment of moral development. Such work is necessary to distinguish between superficial ethical conformity and the durable formation of professional judgment.
[1] Adam T. Biggs, “Enhancing Professional Military Education with AI: Best Practices for Effective Implementation,” Journal of Military Learning, April 2025, U.S. Army Command and General Staff College, https://www.armyupress.army.mil/Journals/Journal-of-Military-Learning/Journal-of-Military-Learning-Archives/JML-April-2025/Enhancing-pme-with-ai/.
[2] Simone Rozzi, Paola Amaldi, and Barry Kirwan, “Identifying How Automation Can Lose Its Intended Benefit along the Development Process: A Research Plan,” in Proceedings of the 9th International Conference on Naturalistic Decision Making (NDM9) (London, UK, June 2009), 384–89, https://www.semanticscholar.org/paper/Identifying-how-automation-can-lose-its-intended-a-Rozzi-Amaldi/d0be7a8d695c39b83d67fd1e11d440e11f04c905.
[3] Max Schemmer, Niklas Kühl, and Gerhard Satzger, “Intelligent Decision Assistance Versus Automated Decision-Making: Enhancing Knowledge Work Through Explainable Artificial Intelligence,” in Proceedings of the 55th Hawaii International Conference on System Sciences (Honolulu, HI, January 2022), 1490–99, https://www.semanticscholar.org/paper/Intelligent-Decision-Assistance-Versus-Automated-Schemmer-K%C3%BChl/c19b58c0045e3c94661509281891f531adcc33ba.
[4] Anneli Jefferson and Katrina Sifferd, “Practical Wisdom and the Value of Cognitive Diversity,” Royal Institute of Philosophy Supplements 92 (October 2022): 149–66, https://doi.org/10.1017/S1358246122000182.
[5] Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th US ed. (Harlow: Pearson, 2021).
[6] James H. Moor, “The Nature, Importance, and Difficulty of Machine Ethics,” IEEE Intelligent Systems 21, no. 4 (2006): 18–21, https://doi.org/10.1109/MIS.2006.80.
[7] John Danaher, “Toward an Ethics of AI Assistants: An Initial Framework,” Philosophy & Technology 31, no. 4 (2018): 629–53, https://doi.org/10.1007/s13347-018-0317-3.
[8] Shannon Vallor, “Moral Deskilling and Upskilling in a New Machine Age: Reflections on the Ambiguous Future of Character,” Philosophy & Technology 28, no. 1 (2015): 107–24, https://doi.org/10.1007/s13347-014-0156-9.
[9] S. N. Hovd, “Tools of War and Virtue—Institutional Structures as a Source of Ethical Deskilling,” Frontiers in Big Data 5 (2023): 1019293, https://doi.org/10.3389/fdata.2022.1019293.
[10] Andreas Matthias, “The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata,” Ethics and Information Technology 6, no. 3 (2004): 175–83.
[11] Hans Jonas, The Imperative of Responsibility: In Search of an Ethics for the Technological Age, trans. Hans Jonas and David Herr (Chicago: University of Chicago Press, 1984).
[12] Shannon Vallor, “Moral Deskilling and Upskilling,” 112.
[13] Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society 5 (2019): 40–60, https://estsjournal.org/index.php/ests/article/view/260/177.
[14] Milán Mór Markovics, “Ideological Drone—Can We Encode Islamism into a Drone?” Ludovika University (2025), https://philarchive.org/rec/MARIDK-2.
[15] Klaudia Klonowska, “Designing for Reasonableness: The Algorithmic Mediation of Reasonableness in Targeting Decisions,” Articles of War, Lieber Institute, February 23, 2024, https://lieber.westpoint.edu/designing-reasonableness-algorithmic-mediation-reasonableness-targeting-decisions/.
[16] Jeremiah A. Lasquety-Reyes, “Towards Computer Simulations of Virtue Ethics,” Open Philosophy 2, no. 1 (2019): 399–413, https://doi.org/10.1515/opphil-2019-0029.
[17] Jovana Davidovic, “On the Purpose of Meaningful Human Control of AI,” Frontiers in Big Data 5 (2022): 1017677, https://doi.org/10.3389/fdata.2022.1017677.
[18] European Parliament and Council of the European Union, Regulation (EU) 2024/1689 of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), art. 12 (“Record-Keeping”), https://artificialintelligenceact.eu/article/12/.
