Human-in-the-Loop, What It Actually Means When a Learner Is in the Room

Somewhere along the way, "human-in-the-loop" got flattened into a job description: the person who clicks approve on whatever the machine spits out. That framing is convenient for the seller. It is dangerous for the learner. And it is wrong.

This piece exists to say, plainly, what human-in-the-loop should mean when the human in the room is a nine year old, a first year teacher, a nurse three months into a new EHR, or a team lead trying to keep a project honest. The short version: the person is not the reviewer of the model. The person is the model. The machine is the tool that helps that person notice where their own map of the world is drifting away from what the world is actually doing.

The frame this site works from

The whole IamHITL front is grounded in active inference at a conceptual level (Class E, standard formulation in Parr, Pezzulo, and Friston, 2022). In plain language: a mind is constantly predicting what happens next, comparing the prediction to what actually happens, and quietly updating the map of the world when the two disagree. Predict, compare, update. That loop is the whole thing.

When we say "the human is the generative model," we mean it in exactly that technical sense (Class B, this is what the code and the classroom protocols both assume). The learner already carries an internal model of how reading works, how a sentence should sound, how a friend should treat them, how a manager should give feedback. New information does not get poured into an empty head. It gets tested against that existing model, and either accepted, refused, or, most interestingly, causes a small repair.

A machine (a tutoring system, a scoring rubric, a checklist, a language model) can be part of that loop. It cannot replace the loop. It can only help the person notice a prediction error more clearly, or offer a candidate way to update. The final update, the change to the actual map, happens inside the person. Nowhere else.

UNI, the science front of this family, is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.

Three commitments that turn HITL from slogan into practice

1. Consent, before anything else

If a person cannot say no to a suggestion the machine (or the curriculum, or the manager) is offering, they are not in the loop. They are downstream of it. Consent is not a checkbox at the top of the session. It is a live posture, renewed as the work moves, honored the moment it is withdrawn.

In practice this looks small. A student can say "not that example, pick another." A patient can ask "what happens if I do not follow this recommendation." A team member can flag a workflow and have the flag counted as data, not as noncompliance (Class C, this is how our own intake webhook and repair-loop logs are configured). The pattern is the same in every setting: the person keeps the veto, and the system treats a veto as information rather than as friction to be optimized away.

2. Repair, not correction

Correction happens to a wrong answer. Repair happens to a relationship, including the relationship between a person and their own map of the world. When a prediction fails, most schools and most workplaces respond as if the person is the defect. Mark it wrong, redo it, move on. That is exactly the moment the loop breaks.

A repair loop, in the sense this site uses it, is different. The failure is treated as a signal that the model needs to update, not that the person needs to feel small. The next move is to sit with the surprise, look at what the model actually predicted and what the world actually did, and choose a small, testable adjustment. Then run the world again and see if the surprise gets smaller (Class B, this is the observable pattern our classroom protocols track). This is trauma-informed, non-clinical work. It is not clinical care, and it makes no clinical claim about a person's trauma history. It simply refuses to add more harm by turning every prediction error into a personal verdict.

3. Shared inference, on the record

A loop is shared. That means the machine's guesses, the teacher's guesses, and the learner's guesses are all on the same table, all labeled, all inspectable. Nobody's map is sacred. Nobody's map is disposable either.

This is where transparency stops being marketing and starts being pedagogy. If the learner cannot see what the system predicted, why it predicted it, and what would have changed the prediction, the learner cannot use the system to sharpen their own model. They can only obey it or ignore it, and both of those are how minds get smaller over time. Our own transparency page and receipts explorer exist for exactly this reason: the record is the shared table.

What this rules out

Framed this way, several popular claims about "AI in the classroom" or "AI at work" become unusable, and honestly should.

A system that grades a student and offers no path to inspect the grade is not in the loop with them. It is over them. A tutor that hides the reasoning behind a recommendation is not helping the learner refine a model. It is asking the learner to substitute the machine's model for their own. A workplace tool that logs vetoes as productivity loss is not learning from its users. It is punishing them for signaling. None of that is human-in-the-loop. It is human-under-the-loop, with a friendlier logo.

Where our own work stops being sure, we say so. Whether a specific classroom protocol reliably reduces the intensity of a repair episode across many students, over months, is (Class U) not yet demonstrated at the scale we would need to make a general claim. We are gathering the evidence, and we will show the falsifier when we publish it (Class F). If the evidence does not hold, we will say that too.

Where this sits in a wider conversation

Others in the active-inference community have started mapping the terrain around this same posture. AJ Maren's resource map Where to Start with Active Inference, A Resource Map for 2026 lists several pathways into the field; in our voice, it is a useful pointer to the surrounding work, not a claim of agreement about ours. A separate essay of hers on how the machine-intelligence landscape has shifted gives field-level context for why a natural, biology-honoring frame matters right now; again, we cite it as context.

A quiet definition, to close

Human-in-the-loop, when a learner is in the room, means this. The person holds the model of the world. The machine offers guesses, evidence, and structured surprise. The person keeps the veto. The failures are read as invitations to repair, not as verdicts. The reasoning is on the record. The people around the learner (teacher, parent, manager, clinician) share the table, not the throne.

That is the whole discipline. It is small. It is slow. It is enough.