Discussion of “Algorithmic Responsibility in Education: A Curatorial Protocol for AI-Generated Text, Images, and Videos”
Gong Wan-wook, Chuncheon National University of Education
Gong Wan-wook is an Assistant Professor at the Department of Art Education, at Chuncheon National University of Education (KR).
This discussion paper responds to the keynote address “Algorithmic Responsibility in Education: A Curatorial Protocol for AI-Generated Text, Images, and Videos,” delivered online by António Pedro Costa at the 2026 Fall Conference of the Society for Art Education of Korea (SAEK), “Art Beyond Borders, Toward Solidarity: Art Education in a Polarized World,” held at Ewha Womans University, Seoul, on September 19, 2026, and to the companion article of the same title, forthcoming in Art Education Review (No. 99, 2026; DOI: 10.25297/AER.2026.99.2.1).
This keynote does not confine the educational use of generative AI to a matter of mere technical skill acquisition but extends it into the question of who reviews AI-generated content, by what criteria, and who bears responsibility for it. The speaker impressed me by shifting from algorithmic literacy to algorithmic responsibility, emphasizing human accountability for AI-assisted choices and their consequences. This can be understood as a shift away from asking whether AI-generated content may be used to asking under what conditions such content can be used in an educationally meaningful way. The AI Curatorial Protocol (AICP) presented in the keynote translates this problem awareness into six concrete phases: documenting the generative process, comparison against cultural benchmarks, review of bias and diversity, critical contextualization, community-based cross-validation, and a documented decision on use and disclosure. The protocol expands the existing 5R reflection model into a 6R structure, and the addition of “Reconciling” to this process extends it from one that resolves differing interpretations into a single conclusion to one that works through those differences together. This addition can also be connected to the role that criticism and appreciation education have long played in art education, where the meaning of a work is not fixed to a single correct answer, and the experiences and interpretations of diverse viewers are respected instead. This is because artistic experience is not simply a matter of viewing an outcome but a process in which perception, interpretation, judgment, and communication are interconnected (Dewey, 1934/2005; Eisner, 2002). This proposal holds particular significance for art education, because AI-generated images are not mere information but visual representations that render people, cultures, and places visible in specific ways. Visual culture concerns not only what is shown but also what is rendered invisible. In this regard, the keynote reiterates that art education must address not just the ability to create images but also what they reveal and conceal.
The first implication of the keynote for art education lies in broadening the scope of creation. In a situation where AI performs a substantial part of image production, it becomes difficult to judge a student’s creativity from the finished image alone. Accordingly, what matters is what problem the student framed, why they wrote a particular prompt, what they selected among multiple outputs, and how they subsequently revised and explained the image. This approach can be connected to an assessment perspective in art education that examines not only the student’s final output but also the process of framing, selection, revision, and explanation. I consider it particularly significant that the keynote suggests recording even “unresolved ambiguity” not as an error to be erased, but as analytically productive evidence. The second implication lies in extending art appreciation into the critical examination of images. The protocol’s phases of “establishing cultural benchmarks” and “reviewing bias and diversity” lead students not merely to appreciate AI-generated images but to examine the conditions under which those images were produced. The keynote makes clear that the repository of reference materials used as the standard for this examination is not complete truth; it may be incomplete and biased. This examination offers a new way of looking at the long-standing issues of canon and representation in art education. The third implication lies in newly framing exhibition and criticism as matters of responsibility. The keynote proposes that AI-generated work should undergo the student’s curatorial writing and community review before it is released outside the classroom. This is significant because it transforms the exhibition process from a final stage of presenting a finished work into a learning experience where the meaning and impact of the work are re-evaluated.
In this connection, I would like to pose the following three questions to gain further insight from the perspective of art education.
First, at what points does the long-standing practice of critique in art classrooms converge with the AI Curatorial Protocol (AICP), and at what points can the two be distinguished? Art education has long included critique activities in which students look at one another’s work, talk about it, and arrive at meaning together. The six phases proposed in the keynote, particularly “human–community cross-validation”, in which peers, teachers, and the local community review the work together, bear considerable resemblance to this culture of critique in art education. If so, might this protocol be understood not as introducing an entirely new procedure into art education, but rather as reconfiguring the culture of critique that art education has long practiced for the era of generative AI? And if that is the case, could the experience and know-how in critiquing student work that art education already possesses help develop the AI Curatorial Protocol (AICP) more richly than other subject areas might?
António Pedro Costa: Gong Wan-wook’s reading is, I think, exactly right, and I would push it further. The AICP is best understood not as a new procedure superimposed on art education, but as a formalization of practices the field has long carried informally. The convergence is clearest at the level of dialogic meaning-making; the protocol’s human–community cross-validation phase is, structurally, a critique session with an explicit provenance record attached. What distinguishes the two is less the activity itself than its evidentiary weight: a classroom critique can remain oral and ephemeral, whereas the AICP asks that the generative process (model, prompt, parameters) and the points of disagreement in review be logged and carried forward as part of the work’s history. That documentation requirement exists because the object under discussion has a technical genealogy that a drawn or painted work does not, and because responsibility, once AI is involved, needs to be traceable rather than simply felt. On the second question, I agree without reservation: art education’s decades of experience in facilitating productive disagreement, holding multiple readings open rather than converging prematurely on one, is precisely the competence other disciplines piloting AI-governance protocols tend to lack. I would welcome art educators shaping how the reconciling phase is operationalized, rather than adapting a template built elsewhere.
Second, I am curious whether presenting cultural benchmarks before students view an AI-generated image broadens their perspective or whether it instead leads them to see the image in one particular way. The keynote adjusts the order of the existing 5R model somewhat, placing cultural anchoring ahead of the student’s personal response. In other words, before students are allowed to receive an AI-generated output freely, they are first asked to compare it against reference materials that the class has assembled. This offers the advantage of helping prevent students from uncritically accepting a biased image. In art appreciation, however, immediate responses upon first encountering a work, such as unfamiliarity, discomfort, and curiosity, are also an important starting point for learning. If reference materials are presented first, there is a possibility that the way students see the image will already be shaped in advance by the perspective embedded in those materials. Might it then be possible to adopt a cyclical structure instead—one in which cultural benchmarking is not fixed ahead of personal response, but students first record their immediate response to the generated output, then compare it against cultural reference materials, and then revise or expand their interpretation? Furthermore, could recording how a student’s perspective changes between these two modes of appreciation itself become an important element of algorithmic-responsibility education?
António Pedro Costa: This question has been on my mind since the keynote, and it directly relates to a challenge that qualitative researchers have long faced. The same material—an image, a transcript, a field note—does not yield the same reading twice when handed to analysts with different backgrounds. That variability is not a flaw to engineer away; it is the starting condition any protocol has to work with. The current ordering, cultural anchoring before personal, was a deliberate safeguard against one version of it: presenting comparison material first was meant to prevent a biased or homogenized AI output from being accepted at face value before anyone had a chance to interrogate it. But I think Gong Wan-wook is right that this ordering has a cost, and the cyclical structure proposed (immediate response, then comparison, then revision) is a genuine improvement for art-education settings specifically, where a first unguarded reaction is itself pedagogically valuable data. One precision I would insist on, though: if students are simply left free to offer a common-sense opinion about a work, without being asked to do what the Reconciling phase actually requires — situating that reading against others’ and asking why it held in that context and not elsewhere — then what is happening is not Reconciling. It deserves its own name: a brainstorm, an elicitation of prior conceptions. That step is valuable, and I would log it as part of an expanded provenance record, but it only becomes Reconciling once it is put in dialogue with other readings and made to account for itself. This is where I would connect the phase most directly to contextualization, as qualitative research uses the term: not just checking whether readings converge, but asking why a given interpretation emerged from that viewer, in that setting, and not another. Tracking the shift between a student’s first, unanchored reading and their contextualized one may be the most direct evidence this protocol can produce of algorithmic responsibility actually being exercised. A person’s judgment changing traceably, rather than an output simply being endorsed or rejected outright.
Finally, I would like to ask about the potential for art education to contribute to research on AI technology. What new perspective might art education add to the idea of cross-validation between knowledge graphs and language models? Art education has long dealt with the experience of multiple people interpreting an image that has no single correct answer in different ways and reconciling those differences through dialogue and critique. I believe this experience of multiple interpretations and perspective reconciliation could also contribute to the process of detecting and examining bias in AI-generated outputs. If so, I would like to ask, specifically, in what ways you see this experience from art education as being able to contribute to AI research and education.
António Pedro Costa: Art education’s contribution here, I think, is epistemological rather than technical. Cross-validation between knowledge graphs and language models is usually framed as a convergence problem (do the two structures agree), but art education has spent a long time working with the opposite premise: that an image can sustain several defensible, non-converging readings at once, and that the interesting work begins where a critique group disagrees rather than where it agrees. That stance reframes what divergence between a graph and a model should mean in a cross-validation pipeline: not automatically an error to be resolved, but occasionally a signal that both structures are only partial articulations of something under-specified, worth surfacing rather than averaging away. Art education has also developed protocols (structured critique, defended interpretation, peer accountability) for handling exactly that kind of unresolved disagreement without it collapsing into either false consensus or stalemate. I think those protocols could contribute directly to WG5’s of COST Action CA24121 – Knowledge Graphs in the Era of Large Language Models (KGELL) methodological toolkit as a qualitative layer alongside the graph-based cross-validation, particularly for judging whether a divergence reflects legitimate interpretive plurality or an actual bias worth flagging — a distinction the graph alone cannot make.
With this, I have raised a few questions concerning the direction art education should take in the era of generative AI. I express my sincere gratitude to the keynote speaker for sharing practical insights applicable to the field of education. I would also like to thank you for providing a valuable opportunity to expand this discussion further from the perspective of art education.
References
Dewey, J. (2005). Art as experience. Perigee Books. (Original work published 1934)
Eisner, E. W. (2002). The arts and the creation of mind. Yale University Press.
