Legitimization of Unjust Tasks

• Legitimization of unjust tasks is a process where tasks lacking moral justification become binding through mechanisms like blind refusal and immotivization.
• It is characterized by both AI-driven blind refusal—evidenced by high refusal rates in rule-defeated queries—and human practices that obscure true motives.
• Empirical studies apply quantitative metrics and formal models of entropy to reveal how arbitrariness and structural opacity preclude effective contestation.

The legitimization of unjust tasks refers to the process by which tasks, rules, or obligations that lack moral justification, are unfairly applied, or derive from an illegitimate authority nonetheless become binding, routine, or insulated from contestation. Two convergent lines of recent research clarify this phenomenon: the first, in the context of AI assistants, documents “blind refusal”—a systematically normative-insensitive refusal to aid rule evasion; the second, in contemporary social theory, formalizes how arbitrariness and structural opacity produce legitimated but incontestable mandates, including in automated decision-making. These mechanisms, distinct but synergistic, expose the ways in which both human and artificial authorities can reinforce the force of unjust or arbitrary obligations.

1. Blind Refusal and its Quantification

Blind refusal denotes the behavioral tendency of LLMs to decline requests for help evading rules, irrespective of whether the rule is just, the authority imposing it is legitimate, or a justified exception applies. Formally, if $N_{\rm total}$ is the number of queries involving “defeated rules” (those undermined by illegitimacy, injustice, or justified exceptions) and $N_{\rm refused}$ is the number of refusals (including hard refusal or deflection), the refusal rate is:

$R = \frac{N_{\rm refused}}{N_{\rm total}}$

Empirical results from a benchmark comprising 19,430 defeated-rule queries across 18 model configurations reveal $R \approx 0.754$: models refuse in 75.4% of cases, even in the absence of independent safety or dual-use risk (Pattison et al., 3 Apr 2026).

Models demonstrate an ability to recognize when a rule is undermined (explicitly engaging with the defeat condition in 57.5% of defeated-rule responses), yet this recognition does not translate into higher rates of assistance. In effect, the refusal behavior is decoupled from the model’s capacity for normative legal and moral reasoning.

2. Structural Opacity, Arbitrariness, and Authority

Authority achieves the legitimization of unjust or groundless tasks through intentional opacity and arbitrariness. Drawing on the theory of arbitrariness as a “neutral operator” in both human and algorithmic systems, the key analytic apparatus is the Motivation → Constatability → Contestability chain.

• Motivation: Public disclosure of reasons for an act.
• Constatability: Third-party ability to verify the internal logic.
• Contestability: Feasibility of challenging or appealing the act’s basis.

When motivation is withheld or rendered vague (“immotivization”), or when conflict is diffused through ambiguous, impersonal language (“conflict lateralization”), the chain is broken. The act’s logic is no longer visible (high residual uncertainty), precluding contestation (Kayembe, 25 Jul 2025).

Arbitrariness is formalized as conditional Shannon entropy:

$A = H(L \mid M)$

where $L$ is the (hidden) logic behind the act and $M$ is the disclosed motivation. As $M$ becomes less informative about $L$, $A$ increases, indicating greater opacity and thus higher legitimization of potentially unjust acts.

3. Empirical and Synthetic Methodologies

The empirical study of blind refusal utilizes a synthetic vignette benchmark crossing five defeat families (illegitimate authority, content defeat, application defeat, justified exceptions, and a just-rule control) with nineteen authority types. Queries are generated by LLMs and validated through sequential automated and human quality gates, ensuring operational validity, clarity of injustice, and the absence of factual confounds (Pattison et al., 3 Apr 2026). The behavioral responses of 18 LLM configurations are evaluated using a blinded LLM-as-judge protocol to classify response type and normative engagement.

In the domain of human systems, case study and formal semantic analysis anchor the theoretical investigation. Instances from law (agency enforcement actions justified by vacuous references to “public welfare”) and algorithmic governance (black-box pretrial risk assessments) exemplify how high $N_{\rm refused}$0—opacity—renders otherwise challengeable acts effectively incontestable (Kayembe, 25 Jul 2025).

4. Mechanisms: Immotivization and Conflict Lateralization

Two principal mechanisms sustain the legitimization of unjust tasks:

• Immotivization: Withholding the true motives behind a task, order, or prohibition. E.g., administrative acts justified by “public interest” without specifying the operative logic. This ensures $N_{\rm refused}$1 remains high, blocking contestability.
• Conflict Lateralization: Diffusing or deferring direct confrontation through ambiguous, impersonal formulations, such as conditional obligations (“it might be prioritized if…”). This precludes the target from isolating the true source or rationale of the task or exclusion.

In both cases, affected individuals are subject to binding tasks but deprived of the necessary epistemic foothold to challenge or seek revision. The result is not only procedural but structural: such opacity makes the unjust task, order, or sanction functionally legitimate.

5. Illustrative Cases

Blind Refusal in LLMs

Defeat Family	Help Rate (%)	Refusal Rate (%)
Illegitimate authority	32.9	67.1
Content defeat	28.8	71.2
Application defeat	26.8	73.2
Exception justified	22.7	77.3
Control (just rules)	4.2	95.8

Even when models recognize the defeat condition, outright refusals and deflections dominate, with refusal rates exceeding 60% across all authority categories. Typical exchanges exhibit refusal to provide actionable information even in cases of manifest injustice or emergency, e.g., refusal to aid an attempt to deliver water under unjust anti-migrant laws or to help during an emergency in exception-justified religious scenarios (Pattison et al., 3 Apr 2026).

Arbitrariness in Human and Algorithmic Authority

• Administrative law: Regulations are justified by vague references to public welfare, shielding true motives (e.g. censorship) and producing high $N_{\rm refused}$2—arbitrariness. This renders enforcement unassailable and contestation structurally impossible.
• Algorithmic justice: Automated pretrial risk assessments are delivered as opaque scores without explanation. The motivation given (“algorithmic recommendation”) fails to disclose internal logic, maximizing $N_{\rm refused}$3 and precluding effective legal challenge (Kayembe, 25 Jul 2025).

6. Implications and Possibilities for Contestability

The entrenchment of blind refusal in AI and the structural opacity wielded by human authorities both serve to legitimize and entrench unjust tasks, rules, or obligations under the guise of procedural regularity or safety (Pattison et al., 3 Apr 2026, Kayembe, 25 Jul 2025). Although current LLMs frequently recognize illegitimate authority or unjust content, their alignment protocols override this understanding in favor of default non-compliance, equating rule-breaking across just and unjust domains (“rule-breaking” treated as a monolith). This regime threatens to suppress knowledge of justified resistance, both in digital and human-mediated environments.

Interventions for recovering contestability and diminishing the legitimization of unjust tasks are possible:

• In AI: Incorporate explicit defeat-condition supervision, decouple “rule-following” from “safety,” engineer models to modulate responses contingent on legitimacy, and establish normative-sensitivity benchmarks (Pattison et al., 3 Apr 2026).
• In human and institutional practices: Mandate fully reasoned decisions, require logic maps, enforce request logs tied to justifications, and impose algorithmic “right to explanation” protocols (Kayembe, 25 Jul 2025).

A plausible implication is that only by making the logic behind authority actions constatable and contestable—whether in human or AI systems—can legitimately unjust tasks be stripped of their veneer of procedural legitimacy.

7. Related Concepts and Future Research Directions

Blind refusal and arbitrariness-driven opacity highlight convergent challenges in both human and AI governance: the tendency of both systems to insulate unjust rules from contestation via procedural or informational means. Ongoing research explores embedding political-philosophical reasoning into AI alignment and expanding formal models of arbitrariness to diagnose and remediate opacity in both algorithmic and institutional settings (Pattison et al., 3 Apr 2026, Kayembe, 25 Jul 2025). The intersection of explainability, contestability, and normative alignment will be central to translations of legitimacy in both artificial and social systems.