Diagnostic vs. Defensive Information Control: A General Theory with Applications to Surveillance, Negotiation, Platform Design, and Peer Review

A General Theory with Applications to Surveillance, Negotiation,

Platform Design, and Peer Review

Devin Arkema

April 2026

Abstract

This paper introduces a fundamental distinction between two structurally different forms of institutional information control: defensive (withholding information to protect a value, process, or entity) and diagnostic (withholding information to instrumentalize the resulting ignorance and generate observational data from the affected party’s behavior under epistemic constraint). While both forms create epistemic asymmetry, they have different legitimacy conditions, different temporal profiles, and different normative requirements. The paper derives from this distinction a specific normative constraint—the equalization principle—which holds that no manufactured diagnostic epistemic asymmetry may persist into consequential adjudicative, evaluative, or decisional processes without disclosure and contestability. The paper defends equalization against weaker alternatives (proportionality, ex post audit, aggregate notice, independent oversight) by arguing that the specific harm of diagnostic control is inferential subordination—the production of interpretations of the affected party’s own conduct that the affected party cannot inspect—and that only disclosure plus contestability addresses this harm at the correct structural level. The paper tests the principle across four domains: state surveillance, negotiation, platform design, and scientific peer review. In each domain, existing failures of justice and institutional legitimacy are traced to violations of the equalization principle. The paper also identifies a form of epistemic injustice—structural epistemic subjection—not captured by Miranda Fricker’s testimonial/hermeneutical taxonomy, and proposes equalization as its institutional remedy. The general conclusion is that the conditions for permissible institutional opacity are more demanding than typically recognized: any institution that manufactures ignorance for diagnostic purposes must specify the temporal boundary of the diagnostic phase, identify the equalization mechanism, and ensure contestability before consequential deployment of the diagnostic yield.

Keywords: information control, epistemic asymmetry, equalization principle, institutional design, surveillance capitalism, epistemic injustice, platform governance, diagnostic information

1. Introduction: The Unexamined Architecture of Information Withholding

Every institutional order that claims authority over persons must, at some point, decide what those persons are permitted to know and when. This is not a peripheral feature of governance but a constitutive one: the management of epistemic access is among the most consequential acts any institution performs. Yet the philosophical literature on information control remains curiously undifferentiated. We speak of secrecy, privacy, transparency, and surveillance as though these terms exhaust the structural possibilities. They do not. This paper argues that a fundamental distinction—between defensive and diagnostic information control—has been hiding in plain sight across multiple institutional domains, and that making it explicit transforms our understanding of when epistemic asymmetry is permissible, when it is not, and what structural constraints govern the boundary.

The distinction originates in legal procedure. A court that sequesters a jury is withholding information defensively: the aim is to protect the deliberative process from contamination, not to test what deliberation looks like in the absence of that information. By contrast, a prosecutor who withholds evidence from a witness during cross-examination is withholding information diagnostically: the aim is precisely to observe what testimony looks like when the witness cannot consult the withheld material. In the first case, ignorance is a shield; in the second, ignorance is an instrument. These are not variations on a single theme. They are structurally different operations with different legitimacy conditions, different temporal profiles, and different obligations upon termination.

This paper develops the defensive/diagnostic distinction as a general theory of institutional information control and argues that the distinction’s primary contribution is not taxonomic but normative: it generates a specific constraint—the equalization principle—that governs any permissible use of diagnostic information control. The equalization principle holds that no manufactured epistemic asymmetry created for diagnostic purposes may persist into any consequential adjudicative decision affecting the party from whom information was withheld. Diagnostic ignorance must be temporary and bounded; it must be equalized before the institutional process enters its consequential phase. This paper tests the generality of this principle across four domains: state surveillance, negotiation, platform design, and scientific peer review.

The argument proceeds as follows. Section 2 formalizes the defensive/diagnostic distinction and identifies its structural features. Section 3 derives the equalization principle, specifies its formal requirements, and defends it against weaker alternatives. Sections 4 through 7 apply the framework to surveillance, negotiation, platform design, and peer review, respectively. Section 8 addresses the relationship between this framework and existing accounts of epistemic injustice. Section 9 considers objections. Section 10 concludes by identifying what the framework reveals about the general conditions for permissible institutional opacity.

Two of the author’s prior publications bear directly on this analysis. From Static Oversight to Distributed Corrigibility (Arkema, 2025) identifies observer effects and epistemic hardening as structural failures in AI governance—failures that arise precisely because oversight systems create epistemic asymmetries without adequate constraints on their diagnostic use.¹ The Entanglement You Were Already Inside (Arkema, 2025) argues that constitutive entanglement between agents and systems makes internal capture detection structurally unreliable—a result that underwrites the present paper’s claim that diagnostic information control requires external equalization mechanisms rather than internal self-correction.²

2. The Distinction Formalized

2.1 Defensive Information Control

Defensive information control is the restriction of epistemic access for the purpose of protecting some value, process, or entity that would be damaged by disclosure. The protected object is not the withholding party’s epistemic advantage but something external to it: a deliberative process, a vulnerable source, an ongoing investigation, a trade secret whose disclosure would undermine legitimate competitive activity. The defining structural feature of defensive withholding is that the ignorance itself does no work. The institution does not learn anything from the fact that the affected party does not know. The epistemic asymmetry is a side effect of protection, not a tool of observation.

Formal characterization: An act of information control C is defensive if and only if (i) the purpose of C is to prevent harm to some identified value V, (ii) the epistemic asymmetry created by C is not instrumentalized—that is, no inference is drawn from the behavior or testimony of the party from whom information is withheld that depends on their ignorance of the withheld material, and (iii) the duration of C is governed by the persistence of the threat to V, not by the informational yield of the asymmetry.

Examples include jury sequestration, witness protection programs, classified national security information, sealed court proceedings during ongoing investigations, and trade secret protections. In each case, the withholding serves a protective function, and the institution’s legitimacy does not depend on what the uninformed party does while uninformed. The jury’s sequestered deliberation is not being tested; it is being shielded.

2.2 Diagnostic Information Control

Diagnostic information control is the restriction of epistemic access for the purpose of generating information from the resulting ignorance. The manufactured epistemic asymmetry is not a side effect but the operational mechanism. The institution withholds information in order to observe what happens when the affected party acts, speaks, or decides without access to the withheld material. The ignorance is instrumentalized: it produces data that the institution could not obtain if the party had full epistemic access.

Formal characterization: An act of information control C is diagnostic if and only if (i) the purpose of C is to generate information about the behavior, testimony, or decision-making of a party P, (ii) the epistemic asymmetry created by C is instrumentalized—that is, inferences are drawn from P’s behavior that depend constitutively on P’s ignorance of the withheld material, and (iii) the epistemic yield of C is a design feature of the information control, not an incidental byproduct.

Examples include cross-examination techniques that withhold evidence to test witness credibility, undercover operations that observe behavior under manufactured conditions, diagnostic medical tests that control patient information to assess symptoms independently, and A/B testing regimes on digital platforms. In each case, the institution is not merely protecting something from disclosure; it is actively using the ignorance as an observational instrument.

2.3 The Structural Difference

The distinction is not one of degree but of kind. Defensive and diagnostic information control have different functional architectures: different relationships between the act of withholding and the institutional purpose it serves. This generates different legitimacy conditions. Defensive control is legitimate when the protected value warrants the epistemic cost and the protection is proportionate. Diagnostic control requires an additional layer of justification: because the ignorance is being used, the institution must justify not only the withholding but the instrumentalization of the resulting epistemic state.

This structural difference has a temporal dimension. Defensive information control has no intrinsic temporal boundary—it persists as long as the threat persists. A classified document may remain classified indefinitely if the security rationale endures. Diagnostic information control, by contrast, is inherently time-bounded, because its purpose—generating information from observed behavior under epistemic constraint—is fulfilled when the observation period concludes. Once the diagnostic yield has been obtained, the rationale for continued withholding evaporates. Any persistence of the asymmetry beyond its diagnostic function constitutes a category shift from diagnostic to something else—and that something else requires independent justification.

This temporal asymmetry is philosophically significant. It means that diagnostic information control carries an inherent obligation of termination that defensive control does not. The diagnostic withholder must, upon completion of the diagnostic function, confront the question of equalization. Failure to do so transforms a bounded epistemic intervention into an unbounded epistemic advantage—a condition this paper will call inferential subordination—the condition in which one party is permanently subject to interpretations of their own conduct, generated under manufactured epistemic constraint, that they cannot inspect, contest, or even learn have been formed. This condition closely resembles what Shoshana Zuboff identifies as the structural core of surveillance capitalism.³

3. The Equalization Principle

The structural analysis of Section 2 generates a specific normative constraint. If diagnostic information control is defined by the instrumentalization of manufactured ignorance, and if that instrumentalization is inherently time-bounded, then the following principle follows:

The Equalization Principle: No manufactured epistemic asymmetry created for diagnostic purposes may persist into any consequential adjudicative, evaluative, or decisional process affecting the party from whom information was withheld. Before the institutional process transitions from its diagnostic phase to its consequential phase, the epistemic asymmetry must be equalized—that is, the affected party must be granted access to the withheld material sufficient to contest any inference drawn from their behavior during the diagnostic period.

The principle has three structural components. First, the temporal boundary: equalization must occur before the process transitions from information-gathering to consequence-assignment. The diagnostic phase and the adjudicative phase must be sequentially distinct, and the asymmetry must not bridge the gap between them. Second, the epistemic content: equalization requires disclosure of the specific material whose withholding generated the diagnostic yield, not merely general notification that information was withheld. The affected party must be able to reconstruct and contest the inferences drawn from their behavior during ignorance. Third, the contestability requirement: equalization is not merely informational but procedural. The affected party must have a genuine opportunity to challenge whether the inferences drawn during the diagnostic period are valid—whether their behavior under epistemic constraint actually demonstrates what the institution claims it demonstrates.

The equalization principle is not a novel invention. It operates implicitly in several established legal doctrines. The Brady v. Maryland rule (1963) requires prosecutors to disclose exculpatory evidence to the defense before trial—a form of mandatory equalization before the adjudicative phase.⁴ The confrontation clause of the Sixth Amendment guarantees the right to cross-examine adverse witnesses—ensuring that the prosecution’s diagnostic yield from witness preparation is contestable at trial. Miranda warnings require that suspects be informed of their rights before interrogation—a notification mechanism that constrains the diagnostic use of interrogation. These doctrines arrive at the equalization logic from different constitutional starting points and serve different procedural functions; they are not straightforward instances of a single rule. But they converge on a shared structural principle: that epistemic asymmetries created during the investigative or prosecutorial phase must not persist unaddressed into the adjudicative phase. The convergence is a family resemblance, not an identity—but the resemblance is structural, not accidental.

3.1 Why Equalization, and Not Something Weaker?

A referee may reasonably ask: why does diagnostic information control require equalization specifically—disclosure plus contestability—rather than some weaker remedy? Four alternatives suggest themselves: proportionality review (the diagnostic yield must be proportionate to the epistemic cost imposed), ex post audit (an independent body reviews whether the diagnostic use was warranted after the fact), aggregate notice (the affected party is told that diagnostic information control occurred, without disclosure of the specific inferences drawn), and independent oversight (a third party monitors the diagnostic process without disclosing it to the affected party). Each is insufficient, and the reason is structural.

Proportionality review addresses the creation of the asymmetry but not its deployment. A proportionate diagnostic intervention can still produce inferences that enter consequential decisions uncontested. The problem the equalization principle targets is not excessive information control but uncontested inferential use. Ex post audit addresses the question of whether the institution behaved properly but not whether the affected party can challenge the inferences drawn from their behavior. An audit that vindicates the institution’s process still leaves the affected party unable to contest specific conclusions—and it is the conclusions, not the process, that enter consequential decisions. Aggregate notice tells the affected party that something happened but not what was inferred. This is informationally empty with respect to the specific inferential content that shapes consequences. And independent oversight substitutes one epistemic asymmetry for another: the overseer knows what the affected party does not, and the affected party must trust the overseer’s judgment without access to the underlying material.

The common failure of all four alternatives is that they leave the affected party in an inferentially subordinate position: someone else holds conclusions about the affected party’s behavior under manufactured conditions, and the affected party cannot examine whether those conclusions follow from that behavior. This is the structural core of the problem. Diagnostic information control does not merely restrict access to facts; it generates interpretations of the affected party’s own conduct that the affected party cannot inspect. The remedy must therefore be addressed to inference, not merely to information. Disclosure provides access to the interpretive content. Contestability provides the structural capacity to challenge whether the inference is warranted. Together, they restore the affected party’s standing as a participant in the consequential process rather than merely a subject of it. No weaker remedy achieves this, because no weaker remedy addresses the inferential subordination that diagnostic information control specifically creates.

What has not been recognized is that these doctrines instantiate a single structural principle—and that this principle generalizes beyond the legal context. The equalization principle identifies a formal constraint on any institutional exercise of diagnostic information control, regardless of domain. To test this claim, the following four sections apply the framework to domains where diagnostic information control operates but the equalization principle has not been articulated.

4. Application I: State Surveillance

State surveillance is paradigmatically diagnostic. Its institutional function is to observe what persons do when they do not know they are being observed. The epistemic asymmetry—the state knows it is watching; the citizen does not know they are being watched—is not a side effect of protection but the operational mechanism of the surveillance itself. What makes surveillance surveillance rather than merely observation is precisely this manufactured ignorance: the subject’s behavior is informative to the state because and insofar as the subject does not know the state is watching.

Zuboff’s analysis of surveillance capitalism identifies the economic exploitation of this structure: commercial platforms extract behavioral data from users who do not know what is being collected, how it is being analyzed, or to what predictive ends it is being deployed.⁵ What the defensive/diagnostic framework adds to Zuboff’s account is a structural diagnosis of why this extraction is unjust—not merely because it is nonconsensual, but because it violates the equalization principle. The diagnostic yield of platform surveillance is never equalized: users are never granted access to the behavioral inferences drawn from their monitored activity, never given the opportunity to contest those inferences, and never informed that the asymmetry’s diagnostic function has concluded. The diagnostic phase has no terminus. It runs permanently.

This framing reveals why consent-based solutions to the surveillance problem are structurally inadequate. A contractualist might object that informed consent transforms the moral character of the asymmetry: if the user agrees to be monitored, the monitoring is no longer imposed, and the epistemic asymmetry is voluntary. But this objection conflates two distinct things. Consent to a process (monitoring) does not grant epistemic access to the inferences generated by that process. A user who consents to data collection has agreed to be observed; they have not thereby been shown what was concluded from their behavior, nor given the opportunity to contest those conclusions. And it is the inferences, not the observations, that enter consequential decisions. Consent to monitoring without equalization is therefore consent to the creation of a diagnostic asymmetry but not consent to its inferential deployment—because the user cannot consent to inferences they do not know exist. This is not a failure of informed consent (the user may fully understand that monitoring will occur) but a structural gap: consent operates on processes, while equalization operates on inferences. The equalization principle requires that diagnostic inferences be disclosed and contestable before they enter consequential processes—before a credit score is assigned, an insurance premium is calculated, a content feed is algorithmically shaped. No amount of upstream consent bridges this gap, because the gap is between process-authorization and inference-access, and these are structurally different epistemic positions.

The framework also distinguishes legitimate state surveillance from illegitimate variants with greater precision than existing accounts. Law enforcement surveillance that is bounded by warrant requirements, subject to judicial oversight, and followed by disclosure during criminal proceedings satisfies the equalization principle: the diagnostic phase is temporally bounded, the yield is disclosed to the affected party, and the party has the opportunity to contest it at trial. Mass surveillance programs that collect data indefinitely, generate inferences that are never disclosed to affected parties, and feed those inferences into consequential decisions (no-fly lists, predictive policing, automated threat assessments) violate the equalization principle structurally, because the diagnostic phase has no boundary and the adjudicative phase offers no epistemic access to the diagnostic yield.

5. Application II: Negotiation

Negotiation involves epistemic asymmetry by design. Each party withholds information about their reservation price, alternatives, constraints, and preferences. The philosophical question is whether this withholding is defensive or diagnostic—and the answer is that it is characteristically both, in ways that existing negotiation ethics has failed to disentangle.

A negotiator who conceals their best alternative to a negotiated agreement (BATNA) is typically engaged in defensive information control: the concealment protects their bargaining position from exploitation. The ignorance of the other party is a side effect of this self-protection; the concealing party does not learn anything new from the fact that the other party does not know their BATNA. But a negotiator who makes a deliberately misleading first offer in order to observe the other party’s reaction—to gauge their flexibility, eagerness, or constraints from their response to a known-extreme position—is engaged in diagnostic information control. The misleading offer manufactures a specific epistemic condition (the other party believes the first offer represents a genuine position) and extracts information from the resulting behavior.

The boundary between these two operations is sharper than it initially appears. The structural test is whether the tactic manufactures a false epistemic environment in order to observe the other party’s response to it. A negotiator who conceals their BATNA does not create a false environment; they simply decline to disclose a fact about themselves. The other party’s ignorance is real but incidental—they would negotiate similarly whether or not the BATNA existed. But a negotiator who fabricates a competing offer creates a false environment: the other party’s response is shaped by a manufactured condition, and the fabricating party extracts information from that shaped response. The test is not whether information is withheld (all negotiation involves withholding) but whether the withholding constructs conditions designed to produce interpretable behavior. But a harder case sharpens the test. A negotiator who conceals a strong BATNA benefits from the other party’s false belief (that the concealer has weaker alternatives). Does the concealer’s observation of the other party’s resulting behavior count as diagnostic? It does not, and the reason reveals a secondary criterion: the concealer did not create the false belief but merely allowed it to persist. The other party’s assumption about the concealer’s alternatives is their own default inference, not an environment the concealer manufactured. The diagnostic test requires active construction of false conditions, not passive non-correction of the other party’s assumptions. This secondary criterion—distinguishing actively manufactured from merely allowed-to-persist epistemic conditions—tracks a real structural difference. Fabricating a competing offer inserts a false object into the other party’s epistemic environment; concealing one’s BATNA merely declines to correct the environment the other party would have constructed independently. Positional bluffing—stating an aggressive anchor one is willing to move from—falls on the defensive side by both criteria: it protects the negotiator’s range and does not construct false conditions. Strategic misrepresentation—planting false deadlines, fabricating constraints, simulating competing offers to gauge the other party’s reaction—falls on the diagnostic side: it engineers a false epistemic environment to extract behavioral data.

This distinction has normative consequences. Defensive withholding in negotiation is generally permissible: one is not obligated to disclose one’s own vulnerabilities. But diagnostic withholding—using manufactured epistemic conditions to extract behavioral information—triggers the equalization principle. If diagnostic inferences drawn from the other party’s behavior under manufactured conditions enter the consequential phase of the negotiation (the final agreement), the affected party has been subjected to an unequalized epistemic asymmetry.

This explains the widespread moral intuition that certain negotiation tactics are unethical even when they are legal. Strategic deception designed to test the other party’s responses—planted false information, simulated time pressure, fabricated competing offers—is diagnostic information control. The intuition that these practices are unfair is grounded not merely in a general prohibition on deception but in the structural principle that diagnostic epistemic asymmetry requires equalization before consequential use. The other party cannot contest inferences drawn from their behavior under conditions they did not know were manufactured. The diagnostic yield enters the final agreement unequalized.

The framework also illuminates the ethics of mediation. A mediator who engages in caucus-style sessions—meeting with each party privately—necessarily creates epistemic asymmetries. The critical question is whether the mediator uses private information defensively (to protect the process) or diagnostically (to test party positions against each other). The equalization principle suggests that mediators must, before the agreement phase, disclose any diagnostic inferences that will shape the proposed resolution, so that both parties can contest whether the mediator’s private observations support the conclusions drawn from them.

6. Application III: Platform Design

Digital platforms present perhaps the clearest case of diagnostic information control operating at scale without any equalization mechanism. The standard platform business model involves three layers of epistemic asymmetry: the platform knows what data it collects (users typically do not); the platform knows what inferences it draws from behavioral data (users have no access to these models); and the platform knows how those inferences shape the user’s subsequent experience (algorithmic curation is opaque). Each layer is diagnostic: the platform is not merely protecting its proprietary methods (defensive) but actively instrumentalizing user ignorance to generate behavioral predictions.

Luciano Floridi’s levels of abstraction framework is useful here for distinguishing between different kinds of platform opacity.⁶ Some platform opacity operates at the implementation level: how the algorithm processes data is proprietary. Other opacity operates at the inference level: what the algorithm concludes about the user is hidden. Still other opacity operates at the consequential level: how those conclusions shape the user’s experience is undisclosed. The defensive/diagnostic framework reveals that these three levels of opacity have different normative profiles. Implementation opacity may be defensively justified (protecting proprietary methods from competitors). But inference opacity and consequential opacity are diagnostic: they instrumentalize the user’s ignorance of what has been inferred about them and how those inferences shape their experience.

The equalization principle generates specific design requirements. Before any consequential decision based on behavioral inference—content ranking, price discrimination, eligibility determination, ad targeting—the affected user must have access to the inference drawn and the opportunity to contest it. This is stronger than existing transparency requirements, which typically require disclosure of what data is collected but not what inferences are drawn from it or how those inferences enter consequential decisions. The equalization principle requires transparency at the inference and consequence levels, not merely at the collection level.

The breadth of the equalization principle raises a legitimate scope question: does every algorithmic personalization decision count as a “consequential” process? The answer is no, and the framework requires a threshold account. The equalization principle applies when diagnostic inferences enter decisions that affect the party’s standing—their access to resources, opportunities, prices, or institutional treatment. The standing criterion can be specified more precisely: a decision is standing-affecting when it alters the party’s position in an institutional or economic structure in ways that constrain their future options. Content ranking that determines which news article appears first in a feed does not alter the user’s institutional position; content ranking that determines whether a job listing, housing advertisement, or credit offer reaches the user does, because it shapes the opportunity set from which the user can act. Price discrimination based on behavioral profiling crosses the threshold because it assigns different economic terms to the same transaction based on diagnostic inferences the user cannot inspect. Eligibility determinations for insurance, lending, or employment cross it because they gate access to institutional resources. The standing criterion is not equivalent to “any decision the user would care about”—it is narrower than preference and broader than legal harm. It tracks whether the diagnostic inference enters a decision that repositions the party within a structure of access, cost, or opportunity. This criterion requires further development for edge cases, particularly around cumulative low-level inferences that individually fall below the threshold but collectively reshape the user’s standing. A resolution would likely require an aggregative threshold—something that tracks the total repositioning effect of diagnostic inferences over time rather than evaluating each inference-to-decision link atomistically. The shape of the problem is clear: atomistic assessment systematically under-counts the standing impact of pervasive low-level diagnostic control, because each individual inference looks trivial while the cumulative effect is structural. The paper flags this as a limitation requiring further work.

This framework identifies A/B testing as a particularly clear case of unequalized diagnostic control. In A/B testing, users are placed in experimental conditions without their knowledge, their behavior under those conditions is observed, and the resulting data enters platform decisions that affect those same users (and others). The diagnostic structure is explicit: the entire point is to observe behavior under manufactured conditions. Yet no equalization occurs. Users are never informed they were in an experiment, never told what was inferred from their behavior, and never given the opportunity to contest whether the experimental conditions generated valid inferences. The Facebook emotional contagion study (Kramer, Guillory, and Hancock, 2014) exemplifies this: users’ emotional responses to manipulated feeds were observed diagnostically, and the resulting inferences entered academic publication and platform design decisions without any equalization.⁷

7. Application IV: Scientific Peer Review

Peer review is the institutional epistemic practice that most visibly deploys both defensive and diagnostic information control—yet the distinction between them has never been made explicit in the philosophy of science literature on peer review. Making it explicit reveals structural tensions in current peer review practice.

The framework applies most directly to the identity asymmetries in review design. Single-blind review is primarily defensive (protecting reviewers from retaliation), but it also creates a diagnostic component: reviewers who know the author’s identity may instrumentalize that knowledge to assess the manuscript against reputation or affiliation—information the author cannot contest. Double-blind review resolves this in one direction but leaves the editorial epistemic asymmetry intact: editors know both identities and make consequential decisions shaped by information unavailable to either party.

A harder question arises with evaluative criteria. When a reviewer applies methodological standards the author did not anticipate, is this diagnostic information control or merely disciplinary pluralism? The distinction matters, and the framework provides a test. Undisclosed criteria are diagnostic when the institution designs the non-disclosure to extract information it could not otherwise obtain—for example, when an editor deliberately withholds review criteria to see whether the author’s work satisfies standards the author was not told to address. Undisclosed criteria are merely opaque when the non-disclosure is a byproduct of under-specification, pluralism, or poor process—the reviewer applies their own standards not because the institution designed the asymmetry but because the institution failed to coordinate evaluative frameworks. The equalization principle applies to the first case but not the second. The second is a governance failure, but it is not diagnostic information control in the structural sense defined here. This distinction narrows the peer review application to genuine cases: editorial practices that deliberately withhold evaluative frameworks to test whether submissions satisfy unstated standards, not merely editorial processes where standards are vague or contested.

Where the diagnostic test is met, the equalization principle identifies what post-review disclosure must include: not merely the decision and summary comments but the specific evaluative framework applied, the inferences drawn, and the opportunity for the author to contest whether those inferences are valid under the applied framework. The same question applies to grant review, promotion committees, and other institutional evaluation processes: are the inferences drawn from the evaluated party’s performance under epistemic constraint disclosed and contestable before the consequential decision is finalized?

A candid acknowledgment is warranted here. Peer review is the domain where the diagnostic/defensive framework is most difficult to apply cleanly, because much of the dysfunction in evaluation processes is emergent, decentralized, and undesigned. Reviewers apply idiosyncratic criteria not because an institution engineered the asymmetry but because no institution coordinated the evaluative framework. The framework’s diagnostic test—asking whether the non-disclosure is a design feature—correctly excludes these cases, but this means the peer review application is narrower than the other three. Surveillance, platform design, and negotiation all involve deliberately engineered epistemic asymmetries where the diagnostic structure is an explicit feature of institutional design. Peer review’s diagnostic cases are real but less pervasive, and the section is best read as demonstrating that the framework can apply to epistemic evaluation contexts, not that all evaluation dysfunction is diagnostic in the paper’s structural sense.

8. Relation to Epistemic Injustice

Miranda Fricker’s influential account of epistemic injustice identifies two forms: testimonial injustice (deflated credibility assigned to a speaker due to prejudice) and hermeneutical injustice (gaps in collective interpretive resources that disadvantage certain groups).⁸ The defensive/diagnostic framework identifies a third form that Fricker’s taxonomy does not capture: structural epistemic subjection, the condition in which a party is placed under manufactured epistemic asymmetry whose diagnostic yield enters consequential decisions without equalization.

Structural epistemic subjection differs from testimonial injustice because the affected party’s credibility is not deflated—their testimony may be taken at full face value. The injustice consists not in how their speech is received but in the fact that the conditions under which they speak are manufactured to generate information they cannot access or contest. It differs from hermeneutical injustice because the interpretive resources exist—the institution that creates the asymmetry possesses a fully articulated framework for interpreting the behavioral data it extracts. The gap is not in collective interpretive resources but in access to them: one party has the framework, the other does not, and this asymmetry is by design.

José Medina’s extension of Fricker’s framework in The Epistemology of Resistance moves closer to the present analysis by emphasizing how epistemic injustice operates through institutional structures and resistant practices, not only through individual credibility judgments.⁹ Medina’s account captures cases where epistemic injustice is sustained by institutional patterns rather than individual prejudice—which is adjacent to the present argument. But structural epistemic subjection identifies a mechanism that even Medina’s extended framework does not isolate: the deliberate manufacture of epistemic asymmetry as an observational instrument, where the affected party’s ignorance is not a background condition of social inequality but a designed feature of an institutional process. The harm is not that the party is epistemically disadvantaged by persistent social structures (Medina) or that their credibility is deflated by prejudice (Fricker), but that their ignorance is engineered and instrumentalized to produce knowledge about them that they cannot access.

This third form of epistemic injustice is particularly relevant to institutional settings. A worker subjected to undisclosed productivity monitoring is not suffering testimonial injustice (their work output is measured accurately) or hermeneutical injustice (the employer has clear metrics). They are suffering structural epistemic subjection: they perform under manufactured conditions whose diagnostic yield enters consequential decisions (promotion, termination, task assignment) without equalization. The same structure applies to students under algorithmic proctoring, citizens under predictive policing, and patients under diagnostic protocols that withhold the physician’s working hypothesis.

What makes this distinctively epistemic rather than merely procedural or institutional? The answer lies in what is at stake. The affected party is not simply subjected to an unfavorable process (procedural unfairness) or dominated by asymmetric institutional power (domination). They are deprived of access to interpretations of their own conduct—conclusions about what their behavior means, drawn under conditions they did not know were manufactured. The injustice is epistemic because the harm operates at the level of knowledge and inference: the institution claims to know something about the party that the party cannot examine, contest, or even learn has been concluded. The worker’s performance review reflects inferences from monitored behavior the worker did not know was being diagnostically observed. The student’s proctoring flag reflects algorithmic conclusions about their conduct drawn under conditions of manufactured surveillance. The injury is not that they were watched, or that the process was unfair, but that knowledge about them was produced from their own instrumentalized ignorance and deployed without their epistemic participation. This is an injustice in their capacity as knowers—Fricker’s defining criterion—but it operates through a mechanism her taxonomy does not describe.

A skeptic might argue that structural epistemic subjection is merely a hybrid of existing categories—procedural unfairness plus domination plus asymmetric interpretive access—rather than a genuinely new form. The reply is that the hybrid fails to capture the mechanism. Procedural unfairness locates the wrong in the process: incorrect steps were followed, or fair process was denied. But the process in diagnostic information control may be formally correct—the monitoring was authorized, the criteria were applied, the decision followed protocol. The wrong is not procedural but epistemic: knowledge was produced about the party from conditions the party did not know were diagnostic. Domination, in Philip Pettit’s republican sense, locates the wrong in the capacity for arbitrary interference.¹⁰ But the diagnostic withholder need not interfere arbitrarily; they may act consistently and predictably. The wrong is not arbitrary power but unilateral interpretation: the institution generates conclusions about the party’s conduct and deploys them without the party’s epistemic participation. Neither procedural unfairness nor domination identifies this specific mechanism—the production of self-regarding knowledge from instrumentalized ignorance—as the locus of injustice. Structural epistemic subjection does.

The equalization principle provides what Fricker’s framework does not: a specific institutional remedy for this form of injustice. Testimonial injustice calls for the cultivation of epistemic virtues (sensitivity to credibility economies). Hermeneutical injustice calls for the expansion of collective interpretive resources. Structural epistemic subjection calls for equalization: the institutional obligation to disclose diagnostic inferences and provide contestability before consequential decisions are finalized. The remedy is structural rather than characterological precisely because the injustice is structural: it inheres in the architecture of information control, not in the attitudes of individual knowers.

9. Objections and Replies

9.1 The Boundary Problem

Objection: The defensive/diagnostic distinction is analytically clean but practically blurred. Most real-world information control serves both protective and diagnostic functions simultaneously. A police investigation both protects its sources (defensive) and tests suspects (diagnostic). How can the equalization principle apply when the same act of withholding serves both functions?

Reply: The dual-function case does not collapse the distinction; it applies it at a finer grain. The equalization principle applies to the diagnostic component of the withholding, not to the withholding as a whole. A police investigation may continue to protect its sources (defensive) while being required to disclose, before trial, the inferences drawn from suspects’ behavior under investigative conditions (diagnostic equalization). This is precisely what the Brady rule accomplishes: it does not require disclosure of investigative methods (defensive) but does require disclosure of exculpatory material (diagnostic yield). The principle applies to functions, not to acts.

9.2 The Operational Impossibility Objection

Objection: Requiring equalization for every diagnostic inference is operationally impossible at scale. A platform that runs thousands of A/B tests simultaneously cannot disclose every inference to every affected user.

Reply: This objection establishes only that current platform practices are incompatible with the equalization principle—which is the paper’s claim, not its rebuttal. If an institution cannot equalize the diagnostic epistemic asymmetries it creates, the framework’s conclusion is not that the principle should be relaxed but that the institution should not create those asymmetries at the current scale. The operational impossibility of equalization is an argument against the practice, not against the principle. An institution that cannot afford to disclose what it infers from manufactured ignorance cannot afford to manufacture that ignorance.

9.3 The Collapse Objection

Objection: All information withholding is implicitly diagnostic: even a defensive withholder is always in a position to observe the affected party’s behavior and cannot help learning from it. The distinction collapses under scrutiny.

Reply: The distinction tracks purpose, not possibility. A defensive withholder could learn from the affected party’s behavior under ignorance, but the withholding was not designed to generate that learning and the institution does not instrumentalize it. The structural test is whether the epistemic asymmetry is a design feature of the information control or an incidental property. This tracks a real institutional difference: a classified document creates an epistemic asymmetry, but the classifying authority does not design the classification to observe what the public does while ignorant of the document’s contents. The asymmetry is protective, not observational. The objection confuses the possibility of diagnostic use with the structure of diagnostic design.

9.4 The Paternalism Objection

Objection: Some diagnostic information control is benevolent—medical diagnostics, educational testing, therapeutic assessment. Requiring equalization in these contexts undermines the practitioner’s ability to help the patient or student.

Reply: The equalization principle does not prohibit diagnostic information control; it constrains its consequential use. A physician who withholds a working hypothesis during examination (to avoid anchoring the patient’s symptom report) is engaged in legitimate diagnostic control. The equalization principle requires only that the physician disclose the diagnosis and the reasoning before the consequential decision (treatment plan) is finalized—which is precisely what informed consent already requires. The principle does not undermine benevolent diagnostic practice; it formalizes the conditions under which such practice is legitimate, and those conditions align with existing best practices in medical and educational ethics. Where they do not align, the misalignment identifies a structural problem in the existing practice, not in the principle.

10. Conclusion: The General Conditions for Permissible Institutional Opacity

This paper has argued that information control in institutional contexts is not a single phenomenon but two structurally distinct operations—defensive and diagnostic—with different legitimacy conditions, different temporal profiles, and different normative requirements. The central contribution is the equalization principle: the formal constraint that no manufactured diagnostic epistemic asymmetry may persist into consequential adjudication without disclosure and contestability.

The four domain applications demonstrate that the principle is not a legal peculiarity but a general structural requirement of permissible institutional information control. State surveillance, negotiation, platform design, and scientific peer review each involve diagnostic information control, and in each case, existing failures of justice and legitimacy can be traced to violations of the equalization principle—diagnostic yield entering consequential decisions without disclosure or contestability.

The framework also identifies a new form of epistemic injustice—structural epistemic subjection—that Fricker’s testimonial/hermeneutical taxonomy does not capture, and provides a specific institutional remedy: equalization. This complements rather than competes with Fricker’s account, extending the analysis of epistemic injustice from individual credibility economies to institutional architectures of information control.

More broadly, the framework reveals that the general conditions for permissible institutional opacity are more demanding than typically recognized. An institution is not justified in maintaining opacity merely because it has a legitimate interest in secrecy. It must determine whether its opacity is defensive or diagnostic. If diagnostic, it must identify the temporal boundary of the diagnostic phase, specify the equalization mechanism, and ensure contestability before consequential use. These are not optional refinements of good governance practice. They are structural requirements of legitimate institutional authority over epistemic access. The core of the argument is that diagnostic information control creates a specific harm—inferential subordination—that no weaker remedy addresses: the affected party becomes subject to interpretations of their own conduct that they cannot inspect or contest. Any institution that manufactures ignorance for observational purposes and deploys the resulting inferences consequentially without equalization is engaged in this form of epistemic domination—regardless of whether the initial withholding was justified.

The implication for AI governance is particularly direct and connects to the author’s prior analysis of distributed corrigibility.¹¹ The equalization principle is, in a precise sense, what distributed corrigibility structurally requires: it is the mechanism by which an observed system retains epistemic standing rather than being reduced to a data source. An AI system subjected to diagnostic monitoring—behavioral evaluation under conditions opaque to the system—that has no access to the inferences drawn from its behavior and no mechanism to contest them is in a state of inferential subordination relative to its overseers. The equalization principle identifies what oversight architectures must provide to avoid this: not merely observation, but disclosed and contestable inference. More broadly, the question for institutional design is not whether institutions should be transparent (they need not be, in all cases) but whether their opacities are defensive or diagnostic, and whether their diagnostic opacities satisfy the equalization principle. These are structural questions with determinate answers, and the framework developed here provides the analytical tools for asking them.

References

Arkema, Devin. “Conviction Under Revisability.” PhilArchive, 2025. https://philarchive.org/rec/ARKCUR.

Arkema, Devin. “From Static Oversight to Distributed Corrigibility.” PhilArchive, 2025. https://philarchive.org/rec/ARKFSO.

Arkema, Devin. “The Entanglement You Were Already Inside.” PhilArchive, 2025. https://philarchive.org/rec/ARKTEY.

Brady v. Maryland, 373 U.S. 83 (1963).

Floridi, Luciano. The Ethics of Information. Oxford: Oxford University Press, 2013.

Fricker, Miranda. Epistemic Injustice: Power and the Ethics of Knowing. Oxford: Oxford University Press, 2007.

Kramer, Adam D. I., Jamie E. Guillory, and Jeffrey T. Hancock. “Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks.” Proceedings of the National Academy of Sciences 111, no. 24 (2014): 8788–8790.

Medina, José. The Epistemology of Resistance: Gender and Racial Oppression, Epistemic Injustice, and Resistant Imaginations. Oxford: Oxford University Press, 2013.

Pettit, Philip. Republicanism: A Theory of Freedom and Government. Oxford: Oxford University Press, 1997.

Zuboff, Shoshana. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. New York: PublicAffairs, 2019.

Zuboff, Shoshana. “Surveillance Capitalism or Democracy? The Death Match of Institutional Orders and the Politics of Knowledge in Our Information Civilization.” Organization Theory 3, no. 4 (2022).

Footnotes

1. Devin Arkema, “From Static Oversight to Distributed Corrigibility,” PhilArchive (2025), https://philarchive.org/rec/ARKFSO. ↩

2. Devin Arkema, “The Entanglement You Were Already Inside,” PhilArchive (2025), https://philarchive.org/rec/ARKTEY. ↩

3. Shoshana Zuboff, The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power (New York: PublicAffairs, 2019), 175–200. ↩

4. Brady v. Maryland, 373 U.S. 83 (1963). The Court held that suppression of evidence favorable to the accused violates due process, establishing that the prosecution’s epistemic advantage cannot persist into adjudication unchecked. ↩

5. Zuboff, The Age of Surveillance Capitalism, 8–12. Zuboff characterizes this as an “epistemic coup”: the unilateral claiming of private experience as raw material for prediction products. ↩

6. Luciano Floridi, The Ethics of Information (Oxford: Oxford University Press, 2013), 40–58. Floridi’s Levels of Abstraction (LoA) framework provides tools for specifying at what level of description a system’s operations should be transparent. ↩

7. Adam D. I. Kramer, Jamie E. Guillory, and Jeffrey T. Hancock, “Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks,” Proceedings of the National Academy of Sciences 111, no. 24 (2014): 8788–8790. ↩

8. Miranda Fricker, Epistemic Injustice: Power and the Ethics of Knowing (Oxford: Oxford University Press, 2007), 1–29. ↩

9. José Medina, The Epistemology of Resistance: Gender and Racial Oppression, Epistemic Injustice, and Resistant Imaginations (Oxford: Oxford University Press, 2013), 27–56. Medina extends Fricker’s account to include patterns of epistemic interaction that are structurally maintained rather than individually enacted. ↩

10. Philip Pettit, Republicanism: A Theory of Freedom and Government (Oxford: Oxford University Press, 1997), 51–79. Pettit defines domination as subjection to another’s capacity for arbitrary interference, whether or not that capacity is exercised. ↩

11. Arkema, “From Static Oversight to Distributed Corrigibility.” The paper argues that static oversight architectures fail because they harden epistemic access rather than distributing it, creating the conditions for undetected capture. ↩