Care Ethics for AI Governance in Higher Education: Applying the Jones AI Ethics Framework

Institutional AI policy borrowed the fairness, accountability and transparency vocabulary of commercial AI ethics, then added an academic integrity frame aimed at student misuse. Jones argues that both can record a disparate outcome as bias while leaving the relations that produced it unexamined, and turns Tronto's phases of care into five questions a governance body can put to one tool, on behalf of specific students.

Key Findings

• The borrowed vocabulary has blind spots. A fairness audit can show that proctoring flags Black students at higher rates, but it does not ask how being flagged is experienced, who answers for it, or what changes once the institution knows.
• Five components, each carrying a question. Attentiveness, Responsibility, Competence, Responsiveness and Trust derive from phases of care and are operationalized through the MSJCC, so a moral quality becomes something a committee can ask and record.
• Responsibility sits with the institution that deployed the tool, not the vendor or the instructor who noticed the pattern, because the institution chose the system for students who did not choose it.
• Trust names where solidarity becomes visible. It asks whether rules bind their authors as they bind students.
• Five recommendations follow, each mapped to components and actors: relational ethics in governance, course policy built with students, equity audits before and after deployment, care ethics in AI literacy, and student capacity to judge institutional AI.

Why the borrowed vocabulary falls short

Universities converged on a shared language quickly. Dabis and Csáki (2024) analyzed early generative AI policy at 30 universities in the top 500 of the Shanghai Ranking and found common emphases on accountability, human oversight, transparency and inclusiveness, while a review of 84 AI ethics guidelines found transparency, justice and fairness among the most cited principles (Jobin et al., 2019). The vocabulary arrived with its assumptions about where harm shows up and who is positioned to notice it.

Principle-based ethics also travels poorly into practice: principles barely influence practitioners (Hagendorff, 2020) and stay too abstract to constrain decisions (Munn, 2023). In universities they meet procurement that sometimes accepts vendor assurances instead of independent vetting (Ali et al., 2024) and rarely lets students refuse tools such as online proctoring (Coghlan et al., 2021). The harm evidence is specific: proctoring flags Black students and students with darker skin tones at higher rates, attributed in part to facial detection failures (Yoder-Himes et al., 2022); detectors misclassify writing by non-native English speakers far more often than writing by native speakers (Liang et al., 2023); success models are less accurate for racially minoritized students, and common bias mitigation has not closed the gap (Gándara et al., 2024). Group parity metrics do not ask how a student experiences being classified (Birhane, 2021).

The academic integrity frame narrows this further: a review of generative AI policies at 20 leading universities found the primary problem they represent to be the possibility that students will not submit original work, with equal access mentioned but given less weight (Luo, 2024). Whether a deployed system treats students equitably then becomes secondary, and the students most exposed to differential effects are those it least registers.

The five components and the questions they ask

Attentiveness comes from caring about, the phase in which a need is recognized, and is operationalized through counselor self-awareness: whose presence and circumstances does a tool register, and what assumptions do decision-makers bring to that judgment? It also reaches the conditions in which students meet the tool: in a survey of 1,106 undergraduates at one public university in New Jersey, 95% reported home broadband, yet three in five reported at least one connectivity challenge, and students with inadequate connectivity were more likely to take incomplete grades (Katz et al., 2025). An attentive institution asks what a signal may be recording before asking what it says about the student.

Responsibility comes from taking care of, the point at which recognition becomes an obligation to act, and is operationalized through client worldview: who must act on what attentiveness reveals, and do they understand affected students' circumstances? Because responsibility is distributed across faculty, leaders, student affairs, procurement, governance and vendors, Jones locates it with the institution that chose to deploy the tool. Competence comes from care-giving, the phase in which a need is met through skilled action, and asks whether those who select and act on a tool can use it well with these students. Faculty report limited AI literacy and broad interest in professional development (Mah & Groß, 2024), and policy frameworks treat training as an operational requirement (Chan, 2023).

Responsiveness comes from care-receiving, the phase in which the recipient's response tells the caregiver whether the need was met: how do students experience the tool, and does what they report change institutional decisions? That requires channels reaching procurement, configuration or discontinuation. Trust comes from caring with, where care becomes a shared structural commitment, and asks whether the same rules bind those who write them and those who live under them. Trust erodes when policy frames generative AI chiefly as a student risk while faculty and administrative uses face no equivalent expectation of disclosure.

How the framework was built, and what it adds

The framework is a four-stage conceptual synthesis drawing on Jaakkola's (2020) templates for theory synthesis and theory adaptation, with no new empirical data. It began from teaching AI ethics in counselor education, where principle-based frameworks addressed fairness at system level but said little about how tools affect relationships between professionals and the people they serve. Jones reads UNESCO's (2021) Recommendation as addressed to governments and project teams rather than to an institution deciding whether one tool serves the students in its care, and sets the components alongside its commitments, with oversight answering to Responsibility, awareness and literacy to Competence, and diversity to Attentiveness and Responsiveness.

The second stage chose care ethics, and chose Tronto over dyadic formulations such as Noddings's because Tronto extends care to institutions. The third added the operational layer, since care ethics names moral qualities but not competencies: each component was mapped to the MSJCC domain whose aims most closely match that phase, because the MSJCC already integrate self-awareness, worldview, relational practice, structural advocacy and intersectional analysis. The fourth tested them against the empirical literature, checking whether each documented harm could be described as a failure of one or more components.

Prior work addresses design and development rather than adoption: Tavory's (2024) application of Tronto's phases to AI in mental health assigns duties to developers and regulators, and care-centered value-sensitive design (van Wynsberghe, 2013) addresses design. Jones claims three contributions: the ethical agent becomes the institutional actor who adopts and configures a tool after it is built, intersectional analysis becomes required in every component, and the unit of analysis becomes one tool, in one context, with specific students.

Two illustrative scenarios

The first is a composite, not a case. A midsize university adopts an automated proctoring system that flags face-detection failures and audio anomalies. After a semester, faculty notice that Black students and students with darker skin tones are flagged more often. The institution calls the tool an industry standard, leaves review to instructors, and does not tell affected students about the concern. Yoder-Himes et al. (2022) studied a system the authors say was used by at least 1,500 universities and found women with the darkest skin tones far more likely to be flagged than darker-skinned men or lighter-skinned students of either sex. An evaluation across 14 courses and 244 examination attempts found automated decisions differed from human proctors' in 74 attempts, approximately 30% (Tweissi et al., 2022). The scenario engages Responsibility and Trust. Manual review reduces consequences without examining the disparity, so the framework treats it as grounds to revisit procurement, audit flagging by student group and consult affected students.

The second concerns AI-text detection in a graduate professional program, where a licensed detector scores each submission, three submissions cross the referral threshold, two of those three students speak English as an additional language, and policy treats the score as sufficient to open an investigation. Liang et al. (2023) found detectors consistently misclassified writing by non-native English speakers, and the accuracy evidence is no better: a widely used detector identified AI-generated content in 91% of submissions built to evade it but flagged only 54.8% of the AI-generated content they contained, while academic staff referred 54.5% of those submissions for misconduct (Perkins et al., 2024), and Kofinas et al. (2025) found markers could not reliably distinguish AI-assisted from student-only work. This engages Attentiveness, because the process never examined whose writing the detector was calibrated to recognize, and Trust, because the cost falls on the student who must rebut it. A framework-informed response would stop letting detector scores open proceedings, audit referrals by student group, and redesign Assessment so AI use becomes a pedagogical question, the shift Giray et al. (2025) describe from detection as policing toward trust as a foundation.

What this means for practice

• Instructors. Before acting on a flag or a detector score, ask what circumstances the signal might be recording, and treat manual review as mitigation rather than remedy, since case-by-case review leaves the pattern intact.
• Administrators. Put the five institutional questions into procurement review, name an owner for the answers before a contract is signed, and apply the same disclosure expectations to faculty AI use as to students.
• Faculty developers. Build care ethics, intersectional research on algorithmic harm and case material into AI literacy programming, so competence includes judgment about whom a tool was built to see.
• Student affairs and governance bodies. Create a reporting route reaching procurement, configuration or discontinuation, and make equity audits and post-deployment monitoring conditions of adoption.

Limitations

• The framework is conceptual and untested; whether decision-makers apply the components consistently and whether their use improves outcomes are open questions.
• Its operational layer comes from counselor education, and the MSJCC were written for counselor-client relationships, so their translation to institution-student relationships is a hypothesis rather than a demonstrated fit.
• Most supporting studies are from U.S. institutions and several policy analyses examined highly ranked universities, so transferability to community colleges, minority-serving institutions and fully online programs is unestablished.
• The two scenarios are composites built from published evidence, not empirical case data: they show how the framework could apply, in hypothetical applications, and not that it works.

Citation

Jones, M. S. (2026). Care Ethics for AI Governance in Higher Education: Applying the Jones AI Ethics Framework. International Journal of AI in Pedagogy, Innovation, and Learning Futures, 1(4).