A companion to Raising Machines, a six-part series about the parenting metaphor in AI. This article asks what that metaphor leaves in the background: the work of care. Sources checked through 13 October 2026.

The Tutor Has a Button; Responsibility Does Not

Imagine a university buying an AI tutor for its introductory physics course. It answers questions at midnight, offers another explanation without sounding annoyed, and ends its messages with encouragement. The procurement document calls it personalised support. The interface calls itself your learning companion.

The university also has a teaching assistant whose contract covers five hours a week. Someone has to check the tutor’s explanations, deal with complaints, update the course material, and help the student who has spent an evening learning a confident mistake. None of those tasks has a button.

This is a hypothetical institution, not a report about a particular deployment. It isolates a question that performance benchmarks do not answer: has the university increased its capacity to care for students, or moved the work somewhere less visible?

The first Raising Machines article asks what we owe the being whose values we shape. The article on gods, slaves, children, and tools asks what duties each metaphor carries. Care ethics adds a third question: who actually carries those duties?

“We are raising AI” sounds like an allocation of responsibility. Until the work and authority are assigned, it is only the beginning of one.

Care Has More Than One Verb

In Moral Boundaries (1993), Joan Tronto distinguishes four phases: noticing a need, taking responsibility for meeting it, doing the practical work, and attending to the response of what receives the care. Their corresponding ethical qualities are attentiveness, responsibility, competence, and responsiveness. Her terms are caring about, taking care of, care-giving, and care-receiving [[1]].

The separation matters. A person can notice a need and leave someone else to deal with it. A manager can accept responsibility without providing the resources. A worker can carry out a task competently while the recipient’s response never reaches anyone who can change the arrangement.

These are distinctions within a practice, rather than a claim that care always proceeds through four tidy boxes. Applied to our tutor, they let us trace how a student’s difficulty becomes somebody’s job, and whether that job can actually be done.

Tronto’s later Caring Democracy (2013) adds caring with: organising caring needs and their provision consistently with democratic commitments to freedom, equality, and justice. That is a further political requirement, not another name for the fourth phase [[2]].

Nothing in this framework makes women naturally responsible for care. Its feminist use here is to question how work and authority get distributed, including distributions that a story about a benevolent parent can make seem obvious. A university can adopt the vocabulary without changing the division of labour at all.

The Machine Arrives With Other People’s Work Inside It

AI development provides an empirical reason to take the division of labour seriously.

Nithya Sambasivan and colleagues interviewed 53 practitioners working on high-stakes AI in India, the United States, and East and West African countries. They documented data cascades: data problems that compound into downstream failures. In their sample, 92% reported at least one cascade. The authors connect these problems to practices and incentives that undervalue data work. The participants were purposively and snowball sampled; the percentage describes those interviews, not a representative estimate for all AI projects. The applications included health and conservation, rather than being a study of contemporary chatbots [[3]].

The result supplies a mechanism worth examining: work treated as secondary can still determine whether the celebrated system works.

Ding Wang, Shantanu Prabhat, and Sambasivan approached the work from another direction. Their 2022 study interviewed 25 data annotators in India, alongside 10 industry experts and 12 AI/ML practitioners. Most annotators worked on image or video tasks for the automotive industry. Participants described tight scrutiny, increasing targets, work pressure, and limited career progression. The authors also recognise advantages of formal employment over platform work, including benefits. Their finding is about the organisation of particular workplaces, not that every annotator everywhere experiences identical conditions [[4]].

Neither study establishes that all data annotation is care work. Reading it through care ethics is an interpretation: some tasks sustain the conditions under which other people can act safely and well, and the people doing them also have needs the institution must take seriously.

Calling a system a child can concentrate attention on the relationship between a developer and a model. It can leave the annotator, evaluator, teacher, support worker, and affected user outside the family photograph. That is a possible limitation of the metaphor, not an experimentally measured effect of using the word child.

First Notice the Need, Then Name the Responsible Institution

Return to the physics tutor. The following is a proposal for analysing its deployment, not a validated checklist for educational effectiveness.

The first phase asks what students need. “More access to explanations” is a plausible answer, but so are accessible material, a place to ask embarrassing questions, help identifying a misconception, and reliable contact with a teacher. These needs may require different interventions. The university should investigate them with students, including students who do not use the proposed tool. Usage logs cannot represent a person whose barrier prevents them from logging in.

The need should remain open to revision. In her 2010 paper on caring institutions, Tronto argues for attention to purpose, power, and the particular needs of different people. She warns against treating needs as fixed, or excluding recipients from judgement because they lack professional expertise [[5]].

In this example, attentiveness means that students can explain why an answer is unhelpful without having to diagnose its technical cause. It also means that a department tests whether its chosen problem really exists. Buying a solution does not establish the need for it.

The second phase names responsibility. A vendor can be responsible for the service it supplies; a department remains responsible for how it uses that service in teaching. The course leader needs authority to suspend the tutor, an agreed route for correcting course material, and a person responsible for handling complaints. Contractual boundaries should be visible to the people expected to work across them.

Responsibility without the capacity to act is an assignment of blame. If a teaching assistant must monitor failures but cannot obtain fixes or stop the service, the institution has given them the problem while retaining the decision.

This is where an ethics discussion becomes a budget discussion.

Care-Giving Needs Hours; Care-Receiving Needs a Route Back

The third phase concerns practical competence. In the hypothetical course, people need paid time to check explanations, test accessibility, maintain approved materials, answer escalated questions, and assess whether the tutor supports the intended learning. Accuracy on a benchmark is relevant evidence for some of that work. It does not allocate the work or settle whether the whole arrangement serves this course.

A deliberately simple calculation shows the scale. Suppose a course has 1,000 students, 10% require one human follow-up each week, and each follow-up takes 20 minutes. Those assumed values require about 33.3 hours of direct follow-up per week: 100 cases multiplied by a third of an hour. That excludes preparation, coordination, accessibility work, and maintenance. It is a workload illustration, not an estimate of an actual escalation rate or a staffing recommendation.

Now suppose the university has purchased unlimited automated conversations but allocated five hours for human follow-up. Under those assumptions it has funded the interface more generously than the repair process. A sympathetic tone cannot close the gap.

The comparison still needs a baseline. Students already need help in a course without an AI tutor, and a useful tutor might reduce some of that workload. Our arithmetic cannot establish that automation increases it. I would compare the whole arrangement with feasible alternatives under the same budget: additional tutorials, better course materials, or a tutor with adequately staffed follow-up. The relevant outcomes include what students learn, who can obtain help, and how much paid and unpaid work each arrangement requires.

The fourth phase asks whether the help met the need. Students require a route back to a person, including a way to challenge an answer and request another form of support. The university should examine understanding and access alongside satisfaction. A student may be pleased with an explanation before discovering that it taught them the wrong principle.

Response is also more than a thumbs-down button. A complaint that produces a ticket nobody has time to investigate is information without responsiveness. In this proposed arrangement, someone reviews patterns, reports what changed, and checks whether the change helped. Students should be able to give feedback without making every conversation available for unrelated uses.

The same questions apply to workers. Can an evaluator flag an ambiguous instruction without being penalised for failing to produce a quick label? Can support staff report an impossible workload? Their response is evidence about the conditions of provision, not merely a disruption of provision.

Caring With Includes People Who Do Not Own the Product

Tronto’s democratic addition changes who gets a say. Her 2013 argument treats the allocation of caring responsibilities as a political matter and asks who is included in making those allocations. She also discusses privileged irresponsibility: arrangements that let some people treat the work supporting their lives as somebody else’s concern [[2]].

Applied to the tutor, this suggests giving students and staff an effective role in defining its purpose, the acceptable uses of their data, and the conditions under which it should be withdrawn. Purchasing the tool should not settle those questions permanently. People who decline it need a usable way to participate in the course.

The workers supplying training and evaluation deserve consideration too. For procurement, the proposal is to ask about paid time, task design, worker representation, and routes for disputing instructions or payments, rather than treating labour as an invisible input. A university may not be able to inspect every upstream workplace. That is a limit on what it can verify, not a reason to count the question as answered.

Participation costs time. If a department invites students and precariously employed staff to govern the system entirely through unpaid meetings, it recreates part of the burden it claims to recognise. Representation needs resources and a defined influence on decisions. An advisory group whose recommendations can be ignored without explanation is a weak response to a problem of power.

These are normative proposals drawn from the framework. The interview studies do not demonstrate their effectiveness, and a successful arrangement would still need evaluation. The framework does make the question more precise: which decisions have become easier for one group because another group must absorb the consequences?

A Female Voice Is Not a Theory of Care

There is a familiar shortcut from care to femininity, and from femininity to service. It deserves its own boundary.

In 2019, UNESCO’s I’d Blush If I Could criticised gendered voice-assistant design and recommended ending female voices as the default, exploring a neutral machine gender, and designing responses to discourage gender-based abuse. Those are the report’s historical concerns and recommendations. They are not a description of every assistant available in 2026, or a controlled experiment measuring how much a particular voice changes users’ attitudes [[6]].

The conceptual problem survives without a current product claim. An interface can stage patience, deference, and availability as feminine while giving us no reason to think that its users are being cared for well. It can also help users while imposing unreasonable demands on the workers behind it. These are logically compatible outcomes; whether either occurs in a particular deployment requires evidence.

Care is not identical to satisfying a command. In the tutor example, doing what the student wants could mean supplying an answer that prevents them from practising. Refusing could support learning, or merely enforce a bad policy. The right question is what response meets the educational need and who can contest that judgement.

Neither a motherly voice nor a refusal settles it.

The Parenting Metaphor Has a Defence

The strongest reply is that parenting already includes all this. Parents notice needs, arrange resources, do practical work, listen to responses, and seek help. Calling AI our children may be intended to make developers accept lasting responsibility instead of selling something and walking away.

That is a defensible use of the metaphor. This argument does not depend on proving that family language makes people careless. The objection is about what the metaphor specifies: it identifies a responsible relationship without identifying the institutions and workers that make responsibility possible.

There is also a tension in copying the parental distribution of authority. A parent may make some decisions for a young child who cannot yet make them. Adult students and workers do not become children because a company calls its model one. They require standing in decisions that affect them. Tronto’s institutional argument is useful precisely because it makes power explicit instead of assuming that a well-intentioned family arrangement can simply be enlarged [[5]].

The distinction is between a reason to accept responsibility and an account of how responsibility should be exercised. The parenting metaphor can offer the first. Care ethics presses for the second.

Duties to People Do Not Wait for Machine Standing

There remains the question raised by the series: might an AI itself become someone to whom care is owed?

This article does not resolve that question. Repairing a system can be care for the people who depend on it without establishing that the system has interests of its own. Conversely, identifying existing human interests does not prove that machines could never have morally relevant ones. Maintenance, care for users, and possible care for the machine are distinct claims.

An argument for machine standing would need to explain which properties matter and why the system has them. A fluent request for kindness cannot settle that argument by itself. Nor should a debate about future machine interests postpone the duties to the students, annotators, teachers, and support staff whose interests are already part of the deployment.

The practical result is modest. A caring interface is one possible component of a caring arrangement. Whether the arrangement deserves the description depends on needs, competence, working conditions, response, and authority.

Back in the hypothetical department, the tutor still has a button. The follow-up work has a budget, a named team, and an appeal route—or it remains somebody’s unpaid evening.

The metaphor cannot fill in the rota.

Literature, source attributions, and the illustrative workload calculation checked through 2026-10-13. The university example and its numerical inputs are hypothetical; the institutional proposals are normative arguments.

References

[1] Tronto, J. C. (1993). Moral Boundaries: A Political Argument for an Ethic of Care. Routledge. Chapter 5, “An Ethic of Care”. Publisher’s chapter abstract: https://doi.org/10.4324/9781003070672-8

[2] Tronto, J. C. (2013). Caring Democracy: Markets, Equality, and Justice. New York University Press. Especially pp. 22–23, 34–35, 46–65, and 104–105. ISBN 9780814782781. https://nyupress.org/9780814782774/caring-democracy/

[3] Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., & Aroyo, L. M. (2021). “Everyone wants to do the model work, not the data work”: Data cascades in high-stakes AI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, 15 pages. https://doi.org/10.1145/3411764.3445518 (author-hosted paper: https://storage.googleapis.com/gweb-research2023-media/pubtools/5936.pdf)

[4] Wang, D., Prabhat, S., & Sambasivan, N. (2022). Whose AI dream? In search of the aspiration in data annotation. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, Article 582, 16 pages. https://doi.org/10.1145/3491102.3502121 (author manuscript: https://arxiv.org/abs/2203.10748)

[5] Tronto, J. C. (2010). Creating caring institutions: Politics, plurality, and purpose. Ethics and Social Welfare, 4(2), 158–171. https://doi.org/10.1080/17496535.2010.484259 (paper: https://e-docs.eplo.int/phocadownloadpap/userupload/aportinou-eplo.int/Creating%20Caring%20Institutions%20Politics%20Plurality%20and%20Purpose.pdf)

[6] West, M., Kraut, R., & Chew, H. E. (2019). I’d Blush If I Could: Closing Gender Divides in Digital Skills Through Education. UNESCO and EQUALS Skills Coalition. https://unesdoc.unesco.org/ark:/48223/pf0000367416 (UNESCO’s dated account of the recommendations: https://www.unesco.org/en/articles/id-blush-if-i-could-world-hit)

Changelog

  • 2026-10-13 — Publication Schedule: Prepared for release on 2 November 2026, after the six-part Raising Machines series. All cited works predate that release; the literature cutoff above records the actual review date.
  • 2026-10-13 — Comparative Evaluation: Added the need to compare the complete deployment with existing provision and feasible alternatives; the workload illustration alone cannot establish an increase caused by AI.
  • 2026-10-13 — Research Review: Distinguished care-ethics interpretation from the findings of two interview studies, marked the university deployment and workload inputs as hypothetical, and separated existing human duties from the question of machine moral standing.