When Care Becomes an Algorithm: The Hidden Ethics of Teaching, Triage, and Dignity

Hatched by MGH

Jul 12, 2026

10 min read

2 views

74%

The question beneath both stories

What if the hardest part of helping people is not the teaching, but deciding how much help they need?

That question sits quietly beneath two very different scenes. In one, a person learns a personal care skill through video prompting, a step by step method that breaks a task into repeatable, visible actions. In the other, a machine learning model helps classify the intensity of an intervention plan, so a person receives a level of support that is closer to what their situation actually calls for. On the surface, one is intimate and manual, the other statistical and automated. But both are trying to solve the same problem: how to convert human need into a form that support systems can actually respond to.

That problem is bigger than disability services or autism treatment. It is one of the central design challenges of modern care. If support is too vague, people get overlooked. If it is too rigid, people get misclassified. If it is too individualized without structure, it becomes inconsistent and hard to scale. If it is too standardized, it risks flattening the person into a category. The tension is not between technology and humanity. It is between precision and dignity.

The real unit of care is not the person, but the moment

Most systems pretend that a person is a stable thing with a stable level of need. In reality, need changes by context, by task, by time of day, by confidence, and by environment. A woman learning to shave her legs may need one kind of support at the start, another once she understands the sequence, and almost none once the routine becomes familiar. A child or adult receiving behavioral support may not need the same intensity across every domain of life, or even across every week.

This is where the deeper insight appears: effective support is not a permanent label, it is a sequence of moments.

Think of a restaurant kitchen. A chef does not ask, “Is this dish easy or hard?” in the abstract. The chef asks, “What does this dish need right now?” A sauce may need reduction, a protein may need rest, a plate may need to be reset. Good support works the same way. It observes the next obstacle, not just the overall identity of the person.

That is why video prompting matters. It reduces a complex task into visible moments that can be mastered one by one. And that is why machine learning matters. It can help systems notice patterns in patient data that indicate what level of support is likely appropriate. Both are attempts to make care more situationally intelligent.

The mistake is to treat intelligence as the opposite of compassion. In fact, compassion without specificity often becomes guesswork. The more exactly a system can distinguish one moment from another, the less it has to rely on vague assumptions about who someone is.

Dignity is not served by treating everyone the same. It is served by noticing what each person needs, at the right level, at the right time.

Why mastery often starts with narrowing the world

There is a seductive myth in education and care: that the best help is the most open ended help. But many skills are learned not by broad freedom, but by carefully constrained attention. Video prompting works because it shrinks the world. It removes improvisation long enough for a learner to see the shape of the task. A sequence that once felt overwhelming becomes manageable because each step becomes visible, literal, and repeatable.

This same principle applies to support planning. When a system can sort available information into a smaller number of meaningful categories, it can make a decision that is clearer, faster, and often more equitable than a loose human hunch. The point is not that an algorithm replaces judgment. The point is that it can reduce ambiguity long enough for judgment to become more reliable.

Here is a useful mental model: support has three layers.

Visibility: Can the person see what to do next?
Calibration: Is the amount of help matched to the actual level of need?
Fadeability: Can support be reduced as competence grows?

Video prompting excels at visibility. Machine learning can improve calibration. A well designed care system needs both, because people do not only need access to tasks, they need access to the right amount of structure around those tasks.

Consider learning to ride a bicycle. A child may begin with training wheels, then move to a parent holding the seat, then to a few practice runs on a quiet path, and eventually to independent riding. The support is not a single act of help. It is a carefully staged disappearance of help. Good systems are not the ones that intervene forever. They are the ones that know how to leave.

This is why the best intervention is often not maximal intervention. It is the smallest effective intervention.

The danger is not automation, but misrecognition

Skeptics often ask whether machines should be involved in caregiving decisions at all. That is the right question, but it is often framed too bluntly. The real danger is not that a model makes a recommendation. The danger is that the recommendation becomes detached from the person it claims to represent.

A model can only classify what it can measure. A video prompt can only teach what it can sequence. Both are partial representations of a life. If a system mistakes the representation for the reality, then it becomes dangerous, even when it is accurate on paper.

This is the central ethical issue: support systems should never confuse predictability with personhood.

Imagine two students. One appears “high functioning” in a classroom, but is exhausted by transitions and melts down at home. Another struggles to communicate in conversation, but independently completes many self care tasks with visual supports. A one size fits all judgment would miss both. A nuanced system would ask not, “What category are they in?” but “What context changes the answer?”

That is also why the phrase “gold standard” needs caution. Standards are useful because they reduce arbitrary decision making. But standards can also become blinders if they are treated as complete truths rather than as starting points. The best standard is one that remains open to revision when it meets a real person.

This is where the pairing of these two ideas becomes especially revealing. Video prompting teaches us that learning improves when information is made concrete and sequential. Machine learning teaches us that classification improves when enough structured data reveals patterns hidden from casual observation. Yet both methods remind us of a limit: the closer we come to precision, the more we must guard against forgetting the person behind the pattern.

The opposite of personalization is not standardization. The opposite of personalization is abstraction without accountability.

A framework for humane precision

If we want support systems that are both scalable and dignified, we need a way to think about them beyond the simple choice between human judgment and automation. Here is a practical framework: the Four Cs of humane precision.

1. Clarify the task

Before deciding how much help someone needs, define the task in concrete terms. “Daily living skills” is too broad. “Shaving legs safely in the shower” is specific enough to teach, measure, and improve.

Why this matters: vague goals produce vague support. The more exact the task, the more accurately support can be matched.

2. Classify the intensity

Not all needs require the same level of intervention. Some people need comprehensive support, some need focused support, and some need only prompts or occasional checks. Data can help identify which bucket is most appropriate, but only if the categories are meaningful and used carefully.

Why this matters: intensity mismatches waste resources and can frustrate the person receiving support.

3. Sequence the learning

Skills are not transferred all at once. They are acquired in steps, often with models, prompts, practice, and fading assistance. A video prompt works because it respects this sequence.

Why this matters: the brain learns patterns more reliably when tasks are chunked into observable parts.

4. Fade the support

Every good support plan should ask a final question: how does this help become less necessary over time? If support cannot be faded, it may be compensating for a design problem rather than building capability.

Why this matters: the aim is not dependency. The aim is increased agency.

This framework applies well beyond disability services. Managers use it when onboarding employees, teachers use it when scaffolding classroom skills, and families use it when helping children manage routines. The common mistake is to give too much help too long, or too little help too soon. Humane precision sits between those extremes.

The deeper synthesis: technology should not replace care, it should make care more legible

The most powerful connection between these two ideas is this: both reveal that care fails when it is illegible.

A person who needs help may not know how to ask for it clearly. A clinician may not know how to estimate it accurately. A caregiver may not know how to teach it consistently. A system may not know how to allocate resources fairly. Video prompting makes a task legible to a learner. Machine learning makes patterns legible to a system. In both cases, the point is not mere efficiency. It is to reduce the gap between what is needed and what is offered.

This reframes the purpose of technology in care. Technology should not be judged only by whether it is novel or automated. It should be judged by whether it increases the legibility of need, the match between support and situation, and the eventual independence of the person.

A useful analogy is eyeglasses. Glasses do not improve vision by seeing for you. They improve vision by making the world readable again. Good care technology should function similarly. It should not take over the person’s life. It should sharpen the interface between the person and the world.

That is a much higher standard than efficiency. It asks whether the system helps someone become more capable, more understood, and less trapped by invisible barriers.

Key Takeaways

Ask what moment needs support, not just what person needs support. Needs vary by task, context, and time.
Use the smallest effective intervention. The best support is often the one that can be reduced as skill grows.
Treat classification as a tool, not a truth. Models can improve decisions, but they should not replace contextual judgment.
Make tasks visible and teachable. Break complex skills into clear steps before trying to measure or scale them.
Design for fading. If support does not lead toward more independence, it may be maintenance, not empowerment.

Conclusion: the future of care is not more or less technology, but better matching

It is tempting to imagine the future of support as a choice. Either humans care, or machines classify. Either we teach with patience, or we optimize with data. But that binary is too crude for the real challenge.

The real challenge is matching, matching the right help to the right person, in the right amount, at the right moment, and then knowing when to step back. That is the hidden common ground between a video that teaches a daily skill and a model that sorts treatment intensity. Both are attempts to make support more exact so that dignity is less dependent on guesswork.

If we learn anything from this intersection, it should be this: care is not merely an act of kindness. It is an engineering problem with moral stakes. The systems we build decide whether people are overwhelmed, under supported, or genuinely equipped. The future belongs to designs that can hold both truth and tenderness at once.

And perhaps that is the deepest shift of all. The goal is not to make people fit systems more neatly. It is to build systems that can finally see people clearly enough to respond well.