Michalinos Zembylas of the Open University of Cyprus, writing in an open forum contribution to the journal, contends that emotion-reading AI does something far more consequential than measure feelings. It reconfigures students’ emotional autonomy and risks reproducing what he calls structural affective injustice. Drawing on affect theory, the philosophy of emotion, and critical AI studies, the paper builds a case that when algorithms classify and evaluate emotional life, they do not neutrally observe it. They actively shape which emotions count as legitimate, which are flagged as problems, and which ways of feeling are treated as normal in the first place.

The technical premise of affective computing dates back to Rosalind Picard’s foundational work at MIT in 2000, which proposed that machines could recognize and respond to human affect. Modern implementations typically rely on facial expression analysis, voice prosody, posture detection, and physiological signals, feeding machine learning models trained to map these inputs onto categories such as happiness, confusion, or frustration. Yet the scientific foundations of this mapping are contested. Psychologist Lisa Feldman Barrett’s research, cited in the paper, challenges the idea that emotions have universal, readable facial signatures, suggesting instead that emotional expressions are highly context-dependent and culturally variable. If the underlying science is shaky, the paper asks, what exactly are these systems measuring, and whose interests does the measurement serve?

Zembylas’s central conceptual innovation is the notion of emotional imperialism, a term he adapts from recent philosophical work by Archer and Matheson. The idea refers to the imposition of universalized, largely Western emotion norms through algorithmic infrastructures. Affective AI systems, he argues, privilege positivity, legibility, and self-regulation: students should appear engaged, optimistic, and emotionally manageable. Feelings that are ambiguous, culturally specific, relational, or simply uncomfortable, such as righteous anger, grief, or collective melancholy, get marginalized or pathologized. The algorithm becomes an invisible arbiter of emotional propriety, enforcing a narrow emotional vocabulary as if it were a universal standard.

This connects to a broader critique of what scholars call affective economies, the ways emotions circulate through social and material systems rather than residing inside individual minds. Following Sara Ahmed’s influential account, Zembylas treats affect and emotion as socially, culturally, and politically mediated practices rather than discrete internal states. In that framing, an emotion-recognition system is not a window into a student’s psyche but a piece of affective infrastructure, a term borrowed from human geography, that channels emotional life in particular directions. When a dashboard rewards visible enthusiasm and penalizes withdrawal, it reshapes the emotional climate of the classroom itself, teaching students what they must perform to be seen as thriving.

The risks are not evenly distributed. The paper draws on documented evidence that emotion recognition technologies perform unevenly across demographic groups. The landmark Gender Shades study by Joy Buolamwini and Timnit Gebru revealed substantial accuracy disparities in commercial facial classification systems, with darker-skinned women misclassified at far higher rates. More recent work by McInerney and Keyes examines how race and disability are configured within emotion recognition technology, suggesting that neurodivergent students and students of color may have their emotional expressions systematically misread. A child whose cultural norms discourage direct eye contact, or whose autistic expression of concentration looks like disengagement to a Western-trained model, could be flagged, corrected, or reported, all under the banner of objective measurement.

Zembylas also warns of paternalistic emotional governance. Affective AI promises to enhance well-being, but well-being defined by whom? The paper situates these tools within a longer history of emotional capitalism and what critics have called the happiness industry, in which positive affect becomes a measurable commodity and a target of management. In schools, this logic echoes what Ben Williamson has analyzed as psycho-policy, the use of psychological data to steer educational governance. Systems like classroom behavior apps already translate social-emotional learning into quantifiable metrics. Affective AI extends this datafication inward, turning the interior texture of feeling into a stream of actionable data, and turning students into objects of continuous emotional surveillance, a dynamic that resonates with Shoshana Zuboff’s account of surveillance capitalism.

Against these dangers, the paper proposes an ethical framework built on four guiding principles. Relational responsibility insists that emotional life is constituted in relationships and that AI systems must be accountable to those relationships rather than to abstract performance metrics. Opacity, drawing on Édouard Glissant’s philosophy of relation, defends the right of students to keep parts of their emotional lives illegible to institutions, resisting the assumption that everything felt must be captured. Non-instrumentalization demands that students’ emotions never be treated merely as resources for optimizing engagement or profit. And cultural multiplicity requires that systems accommodate diverse, culturally situated forms of feeling rather than enforcing a single emotional standard. Together, these principles shift the question from whether emotion AI works to whether it should be allowed to govern feeling at all.

The framework also reframes autonomy itself. Rather than the liberal ideal of a self-contained individual making independent choices, Zembylas adopts a relational account of autonomy, influenced by feminist ethics of care from scholars such as Virginia Held and Joan Tronto. On this view, students’ emotional agency depends on supportive social conditions, and technologies that undermine those conditions, even while claiming to support well-being, may erode the very autonomy they promise to enhance. The paper’s engagement with critical pedagogy, from Paulo Freire to contemporary decolonial critiques of AI in education, signals that this is ultimately a political argument about power, not merely a technical checklist for responsible design.

The paper is candid about the challenges ahead. Operationalizing principles like opacity within commercial EdTech ecosystems that depend on data extraction will be difficult, and existing AI ethics guidelines, as Jobin and colleagues’ global survey showed, often converge on vague principles without enforcement mechanisms. Zembylas acknowledges that his framework is a provocation and a starting point rather than a finished policy instrument. But its timing is pointed: as schools worldwide adopt adaptive learning platforms and emotion-aware tutoring systems, the window for asking foundational questions is closing. The article’s contribution is to insist that the debate over affective AI in education must move beyond accuracy benchmarks and privacy compliance toward a deeper reckoning with whose emotions are recognized, whose are erased, and who holds the power to define what a well-regulated student should feel.

Subject of Research: Ethical frameworks for affective artificial intelligence in education, focusing on affective injustice and emotional imperialism

Article Title: Toward an ethical framework for affective AI in education: challenging affective injustice and emotional imperialism

Article References: Zembylas, M. (2026). Toward an ethical framework for affective AI in education: challenging affective injustice and emotional imperialism. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03344-8

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03344-8

Keywords: affective AI, education, emotion recognition, emotional imperialism, affective injustice, AI ethics, EdTech, surveillance, relational autonomy, critical pedagogy, affect theory, algorithmic bias

The research, published on 19 February 2026 as volume 3, article 11 of the journal, rests on a structured literature review and conceptual analysis rather than new empirical data. Pesek draws on governance theory and on critical scholarship concerning digitalisation, platformisation, and datafication to build an analytical framework capable of describing what actually happens when AI enters schools. The core insight is that AI behaves as a systemic and transversal actor. Unlike a textbook or a learning management system, it does not sit neatly inside a single layer of the education system. It operates across the boundary between centralised regulation and decentralised educational practice, shaping decisions at the ministry level and in individual classrooms simultaneously, often without either policymakers or teachers fully noticing.

To make this complexity tractable, the paper introduces a conceptual distinction between three layers of educational AI. The first layer consists of foundational AI infrastructures, essentially the large AI models on which everything else depends. The second layer is AI content, the generated material that flows into lessons, assessments, and learning resources. The third layer comprises AI tutors, the interactive systems that increasingly mediate the relationship between learners and knowledge. Pesek argues that these three layers demand different governance responses, because they concentrate power, risk, and accountability in different places. Treating them as a single undifferentiated category of “AI tools,” he suggests, is one of the reasons existing policy frameworks have struggled to keep pace.

The analytical heart of the paper is a reconfigured hybrid governance model. Hybrid governance refers to arrangements in which authority is shared between state regulators, commercial platform providers, local institutions, and professional educators, rather than residing in any single actor. The study shows that AI has reconfigured these hybrids in ways that were not anticipated by earlier digital education policy. When a school adopts an AI tutoring system, for example, decisions about what counts as mastery, how feedback is phrased, and which learning pathways are offered are no longer made solely by curriculum authorities or teachers. They are partly encoded in models trained on data that reflect commercial priorities and statistical patterns far removed from any school’s local context. Governance, in other words, has been partially delegated to infrastructure.

From this analysis, Pesek identifies four key governance risks. The first is the homogenisation of learning processes, in which widely deployed AI systems push diverse classrooms toward similar rhythms, formats, and definitions of success, narrowing the space for pedagogical diversity. The second is intensified surveillance, as AI-mediated platforms continuously capture data on how students read, write, hesitate, and solve problems, extending the datafication of education into its most intimate moments. The third risk is blurred accountability: when an AI system produces a misleading explanation, an unfair assessment, or a harmful recommendation, it is often unclear whether responsibility lies with the developer, the platform, the school, or the teacher. The fourth risk is the erosion of student and teacher agency, as both groups find their decisions increasingly shaped by opaque recommendation engines rather than by professional judgement or learner choice.

These risks are not hypothetical. The paper situates them within a well-documented policy trajectory. International organisations such as the OECD and UNESCO have repeatedly highlighted both the promise and the governance gap of educational technology, from the OECD’s Digital Education Outlook 2023 to UNESCO’s 2021 guidance for policy-makers on AI and education. Critical scholars including Neil Selwyn and Ben Williamson have long warned that the digitalisation of schooling concentrates power in platform companies and reconfigures educational purposes around data and measurement. Pesek’s contribution is to connect these strands into a single governance argument: the problem is not merely that AI may be used badly, but that the structures through which education systems steer themselves are no longer adequate to the technology they now contain.

The proposed response is a reconfigured hybrid governance approach that differentiates responsibilities across system levels and across AI functions. At the level of foundational AI infrastructures, the paper points toward regulatory oversight of the models themselves, including transparency about how they are trained and evaluated. At the level of AI content, accountability mechanisms need to establish who is answerable when generated material is inaccurate, biased, or inappropriate for a given age group. At the level of AI tutors, governance must protect the integrity of the learner-teacher relationship and ensure that interactive systems support, rather than supplant, pedagogical relationships. Crucially, the framework insists that these responsibilities cannot all be discharged centrally; they must be distributed across national regulators, intermediate authorities, school leaders, and classroom professionals in a deliberate and explicit way.

The paper also advances concrete policy recommendations aimed at operationalising this approach. These include regulatory oversight tailored to the three-layer distinction between AI models, AI content, and AI tutors; accountability mechanisms that make responsibility traceable across the chain from developer to classroom; and explicit protection of educational purpose and professional autonomy, so that teachers retain the authority to interpret, adapt, and when necessary refuse AI-mediated recommendations. The underlying principle is that governance should safeguard democratic values and human-centred education while still allowing education systems to harness the genuine benefits of AI, such as personalised support and reduced administrative burden. The alternative, the paper implies, is a drift in which governance defaults to the incentives of platform providers rather than the purposes of public education.

What makes the study distinctive is its refusal to treat AI in education as a purely technical or ethical dilemma to be solved at the level of individual classrooms. By foregrounding governance as the central analytical and policy concern, it reframes the debate: the question is not only whether an AI tool helps students learn, but who holds power over the systems reshaping learning, who answers when those systems fail, and how the distribution of authority can be redesigned deliberately rather than left to market forces. For policy-makers, school leaders, and educators confronting the accelerating arrival of AI in education, the message is that the deepest challenges are structural. The technology has already reconfigured the architecture of educational governance; the task now is to reconfigure it back, with accountability, transparency, and human agency as explicit design goals rather than afterthoughts.

Subject of Research: Governance of artificial intelligence in education systems

Article Title: Rethinking Education Governance in the Age of AI

Article References: Pesek, I. (2026). Rethinking Education Governance in the Age of AI. Frontiers of Digital Education, 3(1), Article 11. https://doi.org/10.1007/s44366-026-0085-z

Image Credits: AI Generated

DOI: 10.1007/s44366-026-0085-z

Keywords: AI in education, education governance, hybrid governance, accountability, datafication, platformisation, AI tutors, surveillance, teacher agency, education policy, EdTech, governance risks