Generative AI in Values Education: Affordances, Ethical Risks and Implications for Teaching

Synthesis: Chen (2026) argues that a generative model can be a moral interlocutor but not a values educator, because the educator's position requires answerability, exemplarity and address, and a model can satisfy none of them. Five affordances, from generating moral content to a tacit curriculum of machine authority, each silently replace a pedagogical relation, and the risks internal to them are ordered rather than independent. The claim is conceptual, not empirical: it offers a criterion for evaluating any substitution of a machine for a teacher.

Key Findings

1. The position, not the output, is the test. Three conditions define the educator's position: answerability, exemplarity, and address, and none of them is a dilemma-solving skill.
2. Five affordances are distinguished. Generating moral content, simulating perspectives, Socratic questioning, feedback on moral reasoning, and a tacit curriculum of machine authority.
3. Each affordance replaces a relation. The simulated perspective replaces the stakeholder who can be wronged; the Socratic interlocutor replaces the questioner with a stake in the answer.
4. Risks form four ordered families. Epistemic, relational, axiological, and agentic risks are causally connected, with agentic risk as the end point of the sequence.
5. Explanation is not answerability. A model can produce an explanation of its output on request, but the explanation comes from the same process that produced the claim.
6. Exemplarity is not closable by better training data. A model can describe integrity with precision, but there is nothing it has declined in order to be integrious.
7. Policy is itself a curriculum. Guidance that specifies disclosure requirements but is silent on what counts as a good argument teaches procedure rather than judgment.

Values education as a relational practice

Chen distinguishes values education from three neighboring activities: it is not the transmission of correct positions, because a student can recite a position and still be unable to hold it under pressure; it is not socialization into local norms, because the practice should let a student examine the norms she has inherited; and it is not compliance training, although compliance training often borrows its vocabulary. What distinguishes it is an outcome that is a disposition to judge well, and a disposition is not an item of information that can be transferred from one mind to another. The practice is therefore organized around encounters rather than content: a case is discussed not because it contains the answer but because discussing it puts a student in a position where she must take a view and answer for it. This reframing shifts the question asked of a tool from what it produces to what position it occupies. A textbook and a teacher may contain the same propositions and are not educationally equivalent, because a student stands in a different relation to each. It also means a substitution of a machine for a person must be evaluated as a substitution of a relation rather than of a function, which reorders the usual debate about generative AI in higher education.

The five affordances

The five affordances are not a taxonomy of tools, since the same tool can occupy several positions in a single session. As a generator of moral content, the model supplies dilemma cases, scenario variations, prompts, and assessment items at low cost, replacing the teacher who selects and frames cases; thirty versions of a dilemma with different stakeholders can be produced in an afternoon. The risk is not only fabrication but default: a generated case has already decided which considerations count as considerations, so it carries an implicit moral ontology presented as a neutral prompt. As a simulated perspective, the model occupies the stakeholder who can be wronged, offering an interlocutor who holds a position and never tires or takes offense; the risk is the cognitive form of encounter without its vulnerability, and the persona effect means the framing of the agent changes how students argue. As a Socratic interlocutor, the model supplies unlimited questioning, restatement, and pressure on inconsistency, replacing the questioner with a stake in the answer; the risk is dialogue without answerability and no ground for refusal. As a reader of moral reasoning, the model gives consistent assessment of structure, reasons, and counter-arguments, replacing the reader whose judgment must be earned; the risk is that the model's standard becomes the standard of good reasoning. The fifth position, a tacit curriculum of machine authority, is rarely named as a position at all: students learn from fluent, immediate answers that moral questions have answers of that kind.

The four families of risk

The epistemic family covers what students come to accept as true or well reasoned, and has four varieties: fabrication, implicit moral ontology, homogenization, and the substitution of fluency for judgment. Homogenization follows from the same model being available to everyone, so the range of arguments students encounter narrows toward the range the model produces, and a class of thirty students may arrive with thirty versions of the same three positions. The relational family covers losses to the positions of teacher and student: the other who cannot be wronged, the tutor who is always available and never tired, and the repositioning of the teacher as a supervisor of tool use rather than a person who embodies a commitment. The axiological family concerns value pluralism, where the interlocutor's lack of a position to defend turns pluralism into a list of options, which is precisely what relativism looks like from the inside. The agentic family concerns responsibility: if a model's advice shapes a student's judgment, no one is answerable for that advice, and the student holds a decision whose grounds were produced elsewhere. The families are ordered rather than independent, and the ordering matters for how they should be researched and taught. Treated as independent, they invite independent remedies, such as better retrieval for fabrication and better detection for integrity; treated as ordered, they invite a single question, namely who is answerable to this student for what is being claimed, a question that cannot be answered by improving the tool.

What this means for practice

• Instructors. Put generated material in front of students as an object to interrogate, never as the answer to a moral question, using a three-step task: the model produces a dilemma with a decision, students name the moral register the case assumes, and students rewrite the case so a different register is at stake.
• Instructors. Assess a documented decision and its retrospective account rather than vocabulary tests satisfied by well-formed reasoning.
• Instructors. Hold machine-generated positions open as disagreements, so the absence of an author becomes the lesson.
• Curriculum designers. Keep the interlocutor and educator distinction visible in every design decision that depends on it, and treat institutional policy as a curriculum question: a policy that specifies disclosure requirements but is silent on what counts as a good argument is not neutral.
• Faculty developers. Build teachers' repertoire for conducting disagreement, including recognizing the moral register of a text and distinguishing a reason from a rationalization.

Limitations

• The article is a conceptual theoretical analysis and reports no empirical study: there is no sample, no setting, no intervention, and no outcome data, so the claims are conceptual and normative rather than tested.
• The three conditions of the educator's position are derived from the concept of the practice, not from a survey of teachers, so their derivation is analytic rather than empirical.
• The four families of risk are analytically distinct and causally connected by argument, not by evidence: the ordering that makes agentic risk the end point is asserted from the conceptual analysis rather than demonstrated.
• Four problems remain open and are assigned to future work: whether the cognitive form of moral encounter practiced with a simulation transfers when vulnerability returns, whether a long relation with a system that remembers a student gradually acquires properties of address, how judgment can be assessed at scale, and whether an institutional culture that treats argument as something to be evaluated can be designed.
• The framework delivers a criterion rather than a measure, and a criterion cannot be tabulated, so it offers no instrument for comparing models or classrooms.

Citation

Chen, Y. (2026). Generative AI in Values Education: Affordances, Ethical Risks and Implications for Teaching. Education Insights, 3(10).