Can Machines Perform Emotional Labor in Education? Tracing Debates in Early Philosophy of Artificial Intelligence

Synthesis: Can an artificial system perform the emotional labor of teaching, or only reproduce its outward signs? Noah Khan argues that the second does not justify the first. Drawing on debates about machinic emotion between 1950 and 1996 — Ziff, Smart, and Putnam on whether feeling requires biological life; Weiss, Scriven, and Danto on whether relation, honesty, or language can evidence it; Dyer, Ortony et al., and Arbib on whether cognition-first computation captures embodied affect — the article reconstructs three justificatory questions about ontology, evidence, and embodiment. Its contribution is a shared vocabulary that separates representation, simulation, and mimicry from warranted claims that a system has emotion and performs emotional labor, and that lets educators state which premise a benchmark, design feature, or classroom exchange actually settles.

Key Findings

1. The article reconstructs debates on machinic emotion from 1950 to 1996, closing the year before Picard's 1997 Affective Computing consolidated the field.
2. Ziff's nine-premise case that only living creatures can literally have feelings is answered by Smart's identity theory and Putnam's behavioral criterion, leaving three rival ontologies in play.
3. Weiss, Scriven, and Danto show that love, technically guaranteed honesty, and linguistic compliance each fail on its own to establish inner feeling.
4. Dyer's three models and Ortony et al.'s 1988 account treat emotion as produced by cognition, while Arbib's 1992 review argues that visceral and precognitive processes shape appraisal.
5. The conclusion is a burden of justification rather than an impossibility claim: representation, simulation, and mimicry underdetermine emotional labor.

The gap between convincing performance and emotional labor

Reviews of generative AI in education report broad skepticism among researchers that systems can care for students or demonstrate emotional understanding, yet the counter-evidence is easy to find: ChatGPT has outperformed humans on an emotional-awareness evaluation, has been rated as an empathetic psychoeducational tool, and has produced responses judged more emotionally engaged than those in a therapist-response dataset. Khan asks what would license the move from the second kind of finding to the first.

His answer turns on a set of distinctions. A model of emotion is a formal representation of selected emotional relations; a simulation runs a process designed to reproduce some of those dynamics; mimicry is observable resemblance in output with no claim about the process behind it. Having emotion means possessing states properly characterized as emotional, and performing emotional labor means regulating such states under institutional feeling rules in order to shape another person's emotional experience. A system can model, simulate, and mimic without doing either.

Hochschild's account of emotional labor includes surface acting, so the argument cannot rest on a demand for perfect sincerity. Even while acting, a teacher is an emotional subject with feelings to manage, occupies a role carrying responsibilities, and remains answerable for how the display shapes a pedagogical relation. Khan adds an authenticity condition drawn from Bialystok and Kukar: the educational value of apparent care can depend on students' reasonable perception that it is not wholly fabricated, which concerns the actor's capacity for emotion and responsible relation rather than congruence of feeling and display.

Three justificatory questions, 1950–1996

The first question is ontological: what kind of entity could be a bearer of feeling? Paul Ziff's "The Feelings of Robots" (1959) answers with nine premises and the curt conclusion that "a robot would behave like a robot," since only living creatures can literally have feelings. J. J. C. Smart replies within months, invoking Genesis and physicalism to blur the artifact/organism boundary and arguing that sensations simply are brain processes. Hilary Putnam then separates structure from behavior: structural criteria suit plants, but for entities that behave, judgments of life should prioritize the organization of behavior as described by psychological theory. That move, Khan notes, prefigures the behavioral proxies of affective computing.

The second question concerns evidence within a relation. At the 1960 New York University symposium published as Dimensions of Mind, Paul Weiss argues that a private self is disclosed through love and sympathy, so that a machine we could love would have to be granted human nature and human powers. Michael Scriven counters with a thought experiment: wire a robot so that lying is impossible, and its avowals of feeling become trustworthy. Arthur Danto attacks the data source itself, showing that whether English changes or not, linguistic compliance establishes no criteria for inner feeling — a dilemma that anticipates emotionally fluent large language models.

The third question is about constitution. Michael Dyer's BORIS, OpEd, and DAYDREAMER models treated emotion as emerging from general cognitive processes of retrieval, planning, and reasoning over goals and beliefs, a stance formalized in Ortony, Clore, and Collins's The Cognitive Structure of Emotions (1988). Michael Arbib's review charges that this cognition-first account excludes visceral sensation and facial expression, and that bodily, precognitive, and hormonal processes shape how events are construed in the first place.

Why a preconsolidation window matters

Khan marks 1950–1996 because 1996 is the last year before Picard's Affective Computing (1997) canonized the field and folded disparate arguments into one paradigm. This is not a history of progress, and Khan does not claim that today's researchers inherit these positions consciously. Its value is that incompatible criteria for attributing emotion remained unusually visible before a single paradigm normalized one set of them.

The methodological wager is that a present-day metric may silently presuppose an answer to one of the three questions: an emotion-recognition tool that classifies frustration is already committed to some account of what frustration is, and a generative tutor's reassurance presumes something about whether apt language evidences a subject. Recovering the earlier disagreement makes that presupposition available for scrutiny.

The frustrated student and the levels of claim

The recurring example is a tutor that responds patiently when a student repeatedly fails a mathematics problem. A Ziffian position rejects emotional attribution at the first step, because the system is not alive. A Smart-style position permits attribution in principle but demands evidence about the processes the system instantiates. A Putnamian position asks about the organization of the tutor's behavior across changing circumstances rather than a single reassuring answer. All three can agree the response is useful while disagreeing about what produced it.

Khan uses that agreement to separate usefulness from warranted attribution and to resist a binary in which rejecting the stronger claim means declaring the system worthless. A tool may support a teacher's emotional work by identifying patterns, preparing alternative explanations, or absorbing repetitive communicative demands. Whether that helps depends on accuracy, context, transparency, and institutional use — questions anthropomorphic claims can obscure. The same distinctions underwrite critical AI literacy: students need vocabulary for asking whether a system detected a pattern, represented an appraisal relation, simulated a process, mimicked an expression, or had and regulated an emotion.

What this means for practice

• Instructors and designers should describe what a system does at the level the evidence supports — detected a pattern, generated empathic language, simulated an appraisal — and reserve "cared" or "performed emotional labor" for claims that also defend a subject of feeling, evidence for its state, and a form of responsibility the system can bear.
• Evaluators should ask which dimension of emotion a metric includes and which it treats as dispensable, since adding a camera, microphone, or physiological monitor enlarges the data available to an appraisal model without instantiating the bodily organization an embodied account treats as constitutive.
• Teacher educators can teach the three questions directly, so students encountering first-person, emotionally fluent systems can distinguish proxy performance from the ontological and ethical claims inferred from it.

Limitations

• The article is conceptual and historical: it reports no empirical evaluation of any contemporary affective system, no classroom data, and no student or teacher sample.
• The historical selection is interpretive rather than exhaustive; Khan presents three reconstructed debates and does not claim they exhaust the philosophy of emotion.
• The article declines to decide whether artificial systems can have emotions, so it identifies the burden of justification without settling whether it can be met.
• The canonical debates say little about the gendered distribution of emotional labor or the power to define normal emotional expression, an omission Khan flags as serious for any proposal to automate care.
• The analysis stops at 1996, before affective computing's consolidation, so its categories are not tested against contemporary architectures, benchmarks, or the experiences of students encountering them.

Citation

Khan, N. (2026). Can Machines Perform Emotional Labor in Education? Tracing Debates in Early Philosophy of Artificial Intelligence. Educational Theory.