This position stands on The Parameterisation Gap as its floor — it is the enforcement-sibling. Where the gap says declare the reduction, this says: when the parameterised thing is a person, declaration is not enough — refuse to accumulate it.
A question under a discipline's founding verb
User research has a founding verb: to know the user. We talk about knowing the user the way we talk about knowing a city or a craft — as something you get better at, get closer to, eventually get right. I have taught this verb. I have asked students to build personas, draw journey maps, fill in the goals-and-frustrations box. The method works; it lets a designer act for people they will never meet.
But there is a question the verb hides, and it is worth asking plainly: is the user the person?
I do not think it is. And the gap between the two is not a flaw in the research to be tightened with better interviews. It is the same gap the parameterisation gap already named — the permanent distance between what can be parameterised and what is lived — now lodged in the founding verb of a discipline rather than in a machine's pipeline.
The user is not the person
"User" is a smaller word than it looks, and the smallness is the point of the abstraction. Three properties separate it from "person", and a careful practice keeps all three in view.
The user is the person minus everything the system has no use for. The subtraction is the point of the abstraction — and the thing the discipline forgets.
| Property | What it means |
|---|---|
| Role-relative | You are a user only in relation to a system being used. "User" exists from the maker's side of the glass. The person exists with no system in view. |
| Reductive by design | The user is the person minus everything the system has no use for. The subtraction is not a defect — but the discipline tends to mistake the remainder for the whole. |
| Teleological to know | You know a user in order to — design, convert, retain, serve. You know a friend as an end; you know a user to act upon. |
The persona is a profile wearing a face
Here is the uncomfortable part, and it is why this matters more in research than in a recommender. A recommendation engine's profile is cold and visibly partial — nobody mistakes a click-log for a soul. The persona is warm. It has a name, a face, a pull-quote, a goals-and-frustrations box. It performs completeness.
And the warmth is exactly what hides the reduction. User research, the practice that prides itself on empathy, may be the more refined site of the conflation, precisely because its output feels like a person.
The recommender's profile is cold and visibly partial. The persona is warm — and the warmth is what hides the reduction.
Declaration is not enough
The parameterisation gap makes a claim and then makes a choice. The claim: the lived always exceeds the parameterisable, because no map is the territory. The choice: the honest response is to declare the gap — every tool names what it can capture and admits what it cannot. Honesty about the gap is the architecture's deepest feature there.
When the thing being parameterised is a person, I have come to think declaration is not enough. A profile you confess is still a profile. Naming the reduction does not stop the reduction from being treated as the person the moment it is kept — a persona that persists across a project, gets refined against feedback, gets reconciled into a canonical type, is a map being mistaken for territory, and the disclaimer in the footnote does not change what the team acts on.
Honesty says: I reduced you, and I am telling you so. Refusal says: I will not hoard the reduction at all. A different posture toward the same gap.
The architecture: refuse to accumulate
What does it mean, concretely, to refuse to accumulate a person? It is an architecture, not a sentiment, and it is exact.
No persistent model of the person — a structural prohibition, not a privacy setting. There is no profile table, no embedding-of-you, no cluster you belong to that survives the session. The place where a model of the person would live does not exist in the schema.
The profile is a per-query artefact the person authors. Instead of a machine extracting a model of you from your behavioural exhaust, you compose — in the open, editable, visible — a description for this purpose, now: what you want, what you already know, what you are avoiding. You hold it up like a mask and put it down after. Across sessions the masks may contradict one another, and nothing reconciles them, because there is no canonical you to reconcile them into.
Qualification reads the mask, never the traces. The reading layer is confined to what you declared, barred from inferring what you did not say. A forbidden-inference list validates its input the way the forbidden-string list validates the narrator's output: no persistent preference vector, no demographic guess, no cluster membership, no predicted next action.
No learning from overrides. When you overrule a result, that correction does not flow back to refine a model of you. This is the move a behavioural recommender structurally cannot make, because learning-from-your-corrections is how it builds the profile it then mistakes for you. The correction changes this query's authored mask — which you hold — never a stored model, which the machine would hold.
The cut is narrow and exact: rich models of things, no model of persons. The system can be as sophisticated as you like about the corpus; the prohibition is only on modelling the person.
The grounding: map is not territory
It would be tempting to ground all of this in something dramatic — to say the machine can never reach you because it is a machine, not living matter. I do not think that is right, and it is worth being careful, because the wrong grounding gives the whole argument away.
If the residue were unreachable because of the machine's substrate, then a sufficiently rich organic or analogue computer could, in principle, close the gap — and the honest answer to "the profile is not the person" would become "we will build better hardware." The gap would be a limitation waiting to be solved.
It is not. The residue is unreachable because no rendering of a person is the person — map is not territory — and that holds for any rendering by any device. The reduction is not a property of silicon; it is a property of representation. That is why the refusal is permanent and not provisional: there is no better machine that earns the right to keep a model of you, because the problem was never the machine's quality.
What this does not escape
The authored mask is still a representation. You are still rendered to the machine in reduced form when you compose it. The architecture refuses the worst form of the error — a machine-built, persistent profile mistaken for the person — but it cannot dissolve representation as such. The reduction is merely authored by you, and denied accumulation.
That irreducible residue — that any rendering of a person remains a shadow — is not a problem the architecture solves. It is the seam the architecture leaves honest. Over-claiming here would be its own conflation: pretending the mask is not a reduction would repeat, one level up, exactly the error the whole position is about.
Not a critique — a practice
The corrosive version of this argument — user research is a lie — is easy to make and easy to shrug off, because it leaves the practitioner nowhere to stand. I am not making it. The abstraction is the condition of the work, not a defect in it; you cannot know ten million strangers as persons, and the reduction is what lets anyone design for anyone they will never meet.
The version that holds is constructive: here is what the user hides from you, and here is the practice that keeps the person visible behind the user. Teach the method and its limit. Build the tool that reads the mask you authored and forgets it after. Hold the gap open instead of papering it over with a convincing artefact.
The challenge is not "stop reducing." It is "stop forgetting you reduced" — and stop keeping what you would have to forget.
External corroboration
Magnifica Humanitas (Leo XIV, 15 May 2026) is the most authoritative external diagnosis of this conflation — and it offers no architecture; the diagnosis is its, the cure is here. §10 names "the pretense that a single language… can translate everything, including the mystery of the person, into data." §100: simulated care builds "only the appearance" of a relationship. §128: "a person's future is not calculable" — the reason this architecture refuses to learn from overrides. §178: "restoring to individuals… the ability to decide how [their data] is used" — the requirement the user-authored mask and structural forgetting fulfil. Cited for the diagnosis only: this position keeps its grounding map-is-not-territory, borrows "the mystery of the person" as a phrase rather than a theology, and does not adopt the encyclical's restrain-AI posture — it builds AI right.
Summary
"Knowing the user" is the parameterisation gap lodged in the founding verb of a discipline. The user is not the person — it is role-relative, reductive by design, and known only in order to act. The persona's warmth hides the reduction better than any click-log does.
Standing on the published gap, this position makes one move: when the parameterised thing is a person, declaring the reduction is not enough — the honest response is to refuse to accumulate it. No persistent model, a mask the person authors, a forbidden-inference list, no learning from overrides; rich models of things, no model of persons. The grounding is map-is-not-territory, not the machine's substrate — which is what makes the refusal permanent rather than a wait for better hardware.
The user is not the person. The honest architecture is the one that refuses to forget the difference — and refuses to keep what it would have to forget.