Contestability Presupposes a Decision: The Sorting That Escapes Every Remedy
METHODS · WORKING PAPER — Governance. Programme: Access Classes & Routing Rights. Empirical and analytical; anchored to dated, citable law and literature as of July 2026. This paper engages an established research field and does not claim to originate it.
There is a mature body of research on how to make AI systems contestable, and a body of law that gives people rights to challenge automated decisions about them. Both are built on an assumption so natural it is rarely stated: that a decision was made. This paper is about what happens to the person for whom none was. Before any decision, an automated step has often already ranked, scored, filtered, or priced them — and if that step narrows what reaches them without ever producing a decision anyone can point to, then every contestation mechanism designed to protect them addresses something that, in law and in design, never occurred. Not a decision made badly. A decision never made at all, about a person who was simply never shown.
The claim is narrow and, we think, correct: the sorting layer sits beneath the threshold of every remedy currently available, and it does so structurally rather than by oversight. This is the accountability gap that routing rights exist to name.
What the contestability literature established
Contestable AI by design is a real and productive field, and it should be credited precisely. It holds that AI systems can be built to be open and responsive to human intervention across their whole lifecycle — not only after an automated decision, but during design and development. Alfrink, Keller, Kortuem and Doorn synthesised the literature into a design framework whose features include built-in safeguards against harmful behaviour, interactive control over automated decisions, explanations of system behaviour, human review and intervention requests, and tools for scrutiny by decision subjects and third parties. Almada put the design goal plainly: a contestable system should let users and third parties effectively seek human intervention in a given automated decision. Hirsch and colleagues framed contestability as humans challenging machine predictions, allowing intervention before those predictions are put into force; Vaccaro and colleagues treated it as a deep system property of joint human–machine decision-making and a form of procedural justice. More recent work distinguishes contestability as a construct in its own right — the actionable capacity to challenge, influence, and seek redress over AI-driven decisions — separate from explainability or recourse. The field has moved from theory into practice, with design research conducted alongside local governments deploying algorithmic decision-making in public administration.
Nothing in this paper disputes that work. Its argument depends on it being right.
Why a decision-anchored remedy cannot reach the sorting layer
Read the definitions again with attention to their object. Every one of them takes a decision as the thing to be contested: intervention in a given automated decision; challenging a prediction before it is put into force; redress over AI-driven decisions. This is not a defect of the field. It is a boundary condition, and it is shared by the law.
The clearest illustration is the most citizen-favourable authority available. In SCHUFA (C-634/21, December 2023), the Court of Justice of the European Union confronted exactly the upstream problem: a credit agency produces a score, a bank refuses the loan, and the agency says it merely performed a preparatory act. The Court refused that reading, holding that where a third party draws strongly on the probability value, the establishment of that value is itself automated individual decision-making under Article 22 — and it did so explicitly because the restrictive reading would create a lacuna in legal protection. The concept of a decision, the Court said, has broad scope.
That is the law reaching upstream, and it matters. But observe how it reaches. Article 22 requires three cumulative conditions: a decision, based solely on automated processing including profiling, producing legal effects or similarly significant effects. SCHUFA extends the first condition upstream by attribution — the score becomes a decision because a downstream decision drew strongly on it and it played a determining role. Remove the downstream decision and the attribution has nothing to attach to. The Court’s own limit confirms this: where a lender does not place significant weight on the score, the agency’s activity falls outside the prohibition.
Colorado’s SB 189 has the same architecture. Its obligations attach to automated decision-making technology that materially influences a consequential decision. Same anchor, same requirement: identify the decision, then trace the influence back.
So the doctrine, at its most generous, says: find the consequential decision, then follow it upstream. The sorting layer’s distinctive property is that there is often nothing at the other end to find.
What escapes, and why it escapes structurally
Three properties place a routing event beneath every threshold at once.
No decision is ever identified. A job advertisement is not shown to you. A case is never escalated to the queue where it would have been reviewed. A listing does not surface. Nobody decided to reject you, because nobody considered you. There is no adverse outcome, no determination, no act to attribute — and therefore nothing for SCHUFA‘s attribution to reach and nothing for Colorado’s “materially influence” to influence.
The effect does not clear the threshold, individually. Article 22 requires legal or similarly significant effects; being ranked eleventh rather than third rarely qualifies, and the fact that it happens to you across a hundred systems, every day, is invisible to a test applied one decision at a time. The harm of routing is cumulative and distributional; the remedy is individual and episodic.
No single step plays a determining role. SCHUFA turned on a score that determined the outcome in almost all cases. Modern routing distributes the narrowing across many steps — retrieval, filtering, ranking, personalisation — none of which is individually determinative and all of which are jointly decisive. A doctrine that requires a determining role will find none, precisely where the narrowing is most thorough.
Add a fourth, from the design side: contestability by design can only be built into a system whose outputs are decisions its subjects can perceive. You cannot design an intervention request for someone who does not know they were sorted, or an explanation of a determination that was never rendered. The mechanisms are sound; there is nothing for them to attach to.
The law is not silent — but what it gives is not this
Honesty requires stating what does exist. The Digital Services Act reaches the ranking layer directly: platforms must set out, in plain language, the main parameters of their recommender systems, and the very large ones must offer at least one option not based on profiling. That is a genuine intervention in the sorting layer, and it should not be waved away.
But look at what it provides. It provides system-level transparency: how the machinery works, in general, described in terms of service; and an alternative feed, offered as a setting. It does not provide individual routing notice — you are never told that you were sorted, in this instance, in this way. It provides no route by which you might contest your own routing, no reasons for your own narrowing, no reconsideration of the sorting itself. It tells you about the sieve. It tells you nothing about which side of it you landed on, or why, or how to be re-sieved.
That distinction — between knowing the parameters of a system and knowing what it did to you — is the entire distance between disclosure and decision-accountability, arriving in the layer that precedes any decision.
Reading it through admissibility
The Institute’s method asks whether a system has earned standing to enter a consequential decision chain. Routing systems present a peculiar case: they enter the chain upstream of the point at which the law recognises a decision, and therefore enter without ever being admitted. Nothing calls a ranking engine a decision-maker, so nothing tests its standing. It shapes who reaches the decision without ever being subject to the discipline the decision carries.
Applied here, no record, no standing has an unusual edge. The sorting layer typically produces a rich record — of what was ranked, on what features, in what order — and none of it reaches the person the ranking was about. The record exists, in the operator’s systems, and confers no standing on anyone else. This is the record dial reading Ceremonial from the citizen’s side: complete documentation, zero reconstructability by the person with the most at stake. (How this dial drifts over time is treated in the Futures strand: see Signposts*, Dial 4 —* The Right to Know You Were Routed*.)*
What a routing right would have to require
If the decision cannot serve as the trigger, something else must. We propose the trigger be narrowing.
A routing event occurs where an automated step materially narrows the set of outcomes reachable by an identifiable person — whether or not any decision about that person is ever made. Where a routing event occurs, four things follow.
1. Notice that narrowing occurred. Not that “an AI system was used,” which is the Article 50 disclosure and tells the person nothing about themselves. Notice that this person’s option set was narrowed by an automated step, at a moment when it still matters.
2. A record of the road not taken. This is the evidentiary object the field does not currently have, and it is the hard one. Contesting a decision requires knowing the reasons for the outcome you received. Contesting a routing requires knowing what you were not shown — the counterfactual the sorting removed. Without it there is nothing to argue about: the person can allege only that something might have been withheld, and the operator can answer that nothing was.
3. A route to be routed differently. The remedy for a bad decision is a re-decision. The remedy for a bad routing is a re-routing: reconsideration at the sorting layer, by a party who can actually change how the person is sorted. Sending a routing complaint to a decision-appeal process is a category error — there is no decision to appeal.
4. Standing before the downstream decision, not after. A routing right exercised after the consequential decision has been made is a request for compensation, not a remedy. The narrowing must be contestable while the options it removed can still be restored.
The honest tension
Each of these is expensive, and two of them may be partly unattainable.
Notice on every ranking is noise, and noise is a form of silence: a person notified of every filter, score, and sort that ever touched their file learns nothing. The trigger must therefore be material narrowing of a consequential option set — and this paper does not supply the test that draws that line. It only insists that the line is drawn at narrowing, not at decision.
The record of the road not taken is the hardest requirement, and it collides with real interests, not merely inconvenient ones. Disclosing what a person was not shown may disclose what others were; it invites gaming by those who learn the sieve; and it runs into the same trade-secret wall the CJEU itself respected in SCHUFA, where the Court confirmed the agency need not reveal the mathematical formula weighting the data. A counterfactual record that is systematically unfaithful — reconstructed after the fact, unverifiable against anything — would be worse than none, because it would look like accountability.
And some narrowing is good. Triage by urgency, spam filtering, and relevance ranking serve the people they sort. The argument is not that sorting is illegitimate. It is that sorting which materially narrows a person’s consequential options should not be the only exercise of power over them that carries no obligation at all — merely because no one called it a decision.
What this paper does not settle
Three questions, held open. First, what constitutes material narrowing of a consequential option set needs an operational test, and this paper argues only that the trigger belongs there. Second, whether a faithful counterfactual record can be constructed at acceptable cost — and verified independently — is an open technical and legal problem on which the whole scheme turns. Third, the contestability literature has developed a collective pathway, contesting the design of systems, datasets, and models rather than individual outcomes; whether routing rights are better vindicated collectively than individually is a serious question this paper does not resolve, and the answer may well be that the individual right is unworkable without the collective one.
FAQ
Doesn’t the GDPR already cover this? The CJEU said a credit score can be a decision.
It said so where a third party draws strongly on the score and it plays a determining role in a decision with legal or similarly significant effects. That reaches upstream by attributing an upstream act to a downstream decision. Where no downstream decision about the person is ever made, there is nothing to attribute to — and that is the ordinary condition of the sorting layer.
Isn’t this what contestable AI research already addresses?
That research is sound and this paper depends on it. But its object is the decision: intervention in a given automated decision, challenging a prediction before it is put into force, redress over AI-driven decisions. Where no decision is rendered and the person is never shown, there is nothing for those mechanisms to attach to.
Doesn’t the DSA already regulate ranking?
Partly, and genuinely: main recommender parameters must be disclosed in plain language, and the largest platforms must offer an option not based on profiling. That is system-level transparency and a setting. It is not individual notice that you were sorted, nor any route to contest your sorting.
What is the trigger, if not a decision?
Narrowing. Where an automated step materially narrows the set of outcomes reachable by an identifiable person — whether or not any decision about them is made — a routing event has occurred, and it should carry notice, a record of what was narrowed away, a route to be routed differently, and standing exercisable before the downstream decision.
