Contest & Justify in Algorithmic Decision Making

• Contest & Justify is a framework that enables individuals to challenge adverse algorithmic decisions by demanding transparent reasons and formal review processes.
• It draws on administrative law principles, incorporating standing, notice, and multi-layered review to ensure procedural fairness and accountability.
• Its implementation leverages dynamic explainability and computational argumentation to provide clear evidentiary standards and iterative redress mechanisms.

“Contest & Justify” denotes a conception of algorithmic contestation in which socio-technical arrangements let people challenge algorithmic decisions that affect them and require those decisions to be properly reasoned and reviewable. In this usage, contestation is directed at decision subjects rather than expert users exploring model behavior, and it functions as an accountability infrastructure: affected persons must be able to appeal a decision, obtain enough information to “meaningfully challenge” it, and access a process capable of review, correction, and redress (Lyons et al., 2021).

1. Conceptual foundations

In the algorithmic-accountability literature, contestation is distinct from model debugging or interactive machine-learning feedback. “Designing for contestability” in the sense associated with expert users concerns exploration, feedback, and improvement of system behavior, whereas “designing for contestation” concerns the position of the person affected by a decision: whether that person can challenge an adverse output, obtain reasons, and secure review (Lyons et al., 2021). This makes contestation a procedural safeguard rather than merely an explanatory feature.

The concept is closely tied to procedural fairness. One formulation states that “the procedural dimension of fairness entails the ability to contest and seek effective redress against decisions made by AI systems and by the humans operating them,” and related work links contestation to autonomy, accountability, due process, legitimacy, and protection from harm (Lyons et al., 2021). In the Australian AI Ethics Framework, Principle 7 was stated as: “When an algorithm impacts a person there must be an efficient process to allow that person to challenge the use or output of the algorithm.” Submissions on that framework treated contestability both as a standalone principle and as a mechanism through which accountability, fairness, autonomy, and transparency are realized (Lyons et al., 2021).

A terminological distinction is also necessary. In this literature, “contest” refers to challenging a decision. By contrast, in the contests literature, a contest is a strategic environment in which agents exert effort and winning probabilities are governed by a contest success function and a market-clearing condition; that usage is analytically unrelated to algorithmic review and redress (Azrieli et al., 2024).

2. Administrative-law lineage

A central source for “Contest & Justify” is the Australian administrative law system, treated as a model for contestation in algorithmic decision-making. Administrative law is framed as “a system of accountability” governing executive decisions that affect individuals, such as visas, welfare, and licences, and it supplies an institutional vocabulary for reviewability, reasons-giving, and remedies (Lyons et al., 2021).

Three elements are particularly important. The first is transparency and access to information, including mechanisms that enhance transparency by allowing access to information, for example through rules enabling freedom of information. The second is investigative oversight bodies, such as the Commonwealth Ombudsman, the Administrative Review Council, and Royal Commissions, which address systemic issues rather than only individual grievances. The third is avenues of review, including internal and external review, through which individuals can contest decisions that affect them (Lyons et al., 2021).

This administrative-law inheritance matters because contestation is not reduced to a single interface element or explanation artifact. It is a layered institutional arrangement in which standing, notice, hearing, reasons, review scope, and remedy are specified in advance. A plausible implication is that meaningful contestation in algorithmic systems requires comparable specification of institutional roles and powers, even when the technical substrate differs from public-law decision-making.

3. Procedural structure of contestation

Administrative law identifies a set of pertinent features that can be mapped into algorithmic design questions. The resulting structure is neither purely legal nor purely technical; it is procedural, informational, and organizational at once (Lyons et al., 2021).

Administrative feature	Administrative content	Algorithmic design question
Standing and reviewability	A person needs standing and not all decisions are reviewable	Who can contest, and which decisions are contestable?
Fair hearing	Notify a person if a decision is likely to be adverse and give an opportunity to respond	Is a pre-decision hearing required, and through what process?
Notice of decision	Provide notice and information about how to initiate review	How is the person notified, and how is the contestation interface designed?
Statement of reasons	Reasons include facts, evidence, and reasons for the decision	What counts as a statement of reasons for an algorithm?
Multiple avenues of review	Internal merits review, external merits review, judicial review	Who performs review, and what escalation path exists?
Scope and remedies	Merits review examines the right decision; judicial review examines lawfulness	What can be challenged, and what redress can be provided?

Within this structure, standing determines whether a person is sufficiently “affected” to contest. Fair hearing introduces participation before an adverse outcome. Notice requires that decision communications include both the result and the path to review. Multiple avenues of review prevent the original decision-maker from being the only arbiter of correctness. Scope and remedies determine whether contestation reaches only the output or extends to inputs, training data, decision rules, or the process for deriving the model (Lyons et al., 2021).

The same literature emphasizes that contestation processes must be usable in practice. Studies of social media content moderation report lack of clear instructions on how to appeal, no reply or resolution, and limited or no human intervention; many users who could appeal chose not to because they did not know how or did not expect a response. This suggests that a merely formal appeal right is insufficient unless accompanied by clear communication, responsive review, and visible outcomes (Lyons et al., 2021).

4. Justification, explanation, and evidentiary standards

Justification is the informational counterpart of contestation. In the administrative template, a “statement of reasons” should contain “findings on material questions of fact,” “evidence or other material the findings were based on,” and “the reasons for the decision.” In algorithmic settings, this raises the question: “What does a statement of reasons look like for an algorithm?” (Lyons et al., 2021)

This question is sharpened by black-box opacity. The literature states that a major challenge of black-box algorithmic decision-making systems is that their opacity obscures the decision-making process, and it may be unlikely to be easy to produce the full information required by a traditional statement of reasons. At the same time, full internal disclosure is not always necessary for contestation: depending on what a person wishes to challenge, counterfactual explanations can provide valuable information without opening the “black box” (Lyons et al., 2021). The Houston teachers case involving EVAAS is frequently cited because teachers could not access the source code or methodology, and a court held that their constitutional right to due process was violated because they were unable to “meaningfully challenge” the termination of their contracts due to “lack of sufficient information” (Lyons et al., 2021).

More recent work formalizes the insufficiency of standard XAI for contestation. One proposal distinguishes normative and epistemic contestability by comparing the actual decision $\hat{d}(\tilde{x}^{(i)})$ to two benchmarks: a normatively correct decision $d^*(i)$ and an epistemically correct decision $d(x^{(i)})$. A decision is normatively contestable if $\hat{d}(\tilde{x}^{(i)}) \neq d^*(i)$ and epistemically contestable if $\hat{d}(\tilde{x}^{(i)}) \neq d(x^{(i)})$ (Freiesleben et al., 15 May 2026). On this account, counterfactuals, LIME, and Anchors, even when combined with intuitions about continuity, monotonicity, or reasons, reveal only errors in the neighborhood of the individual and provide insufficient grounds for overturning the decision at hand. Instead, three forms of evidence are identified as warranting reversal according to the decision-maker’s own standards: predictive multiplicity, incorrect feature values, and neglected overruling evidence (Freiesleben et al., 15 May 2026).

A plausible implication is that “Contest & Justify” requires an evidentiary architecture, not only an explanatory one. Explanations may help a person identify where to challenge, but contestation succeeds when reasons can be tested against facts, data provenance, alternative models, and additional evidence.

5. Human-centred and computational implementations

A parallel strand of research argues that contestability must be human-centred. This means beginning from how people conceptualize contestability in AI, focusing on the needs of decision subjects and the expectations of the community, and recognizing that AI systems exist within socio-technical contexts shaped by legal frameworks, political systems, and social norms (Lyons et al., 2021). On this view, contestation processes should be clear and easy to access, aligned with human rights, equality, accessibility, and compensation, and calibrated to context rather than imposed as a one-size-fits-all mechanism (Lyons et al., 2021).

At the computational level, one influential position is that contestable AI requires dynamic explainability and dynamic decision-making processes. Rather than a fixed model $M : I \to O$ with a one-shot explanation, the system must be able to interact with humans or other machines, progressively explain its outputs and reasoning, assess grounds for contestation, and revise decisions when those grounds succeed (Leofante et al., 2024). Computational argumentation is proposed as a suitable substrate. In the canonical abstract formalism, an argumentation framework is $AF = (A, R)$, where $A$ is a set of arguments and $R \subseteq A \times A$ is an attack relation; contestation then becomes a matter of introducing new arguments, attacks, and revisions to determine whether the original decision remains justified (Leofante et al., 2024).

Quantitative argumentation extends this program. In an Edge-Weighted Quantitative Bipolar Argumentation Framework, an instance is written as $$\mathcal{Q} = \langle \mathcal{A}, \mathcal{R}^{-}, \mathcal{R}^{+}, \tau, w \rangle$$, with arguments, attack and support relations, base scores, and edge weights. The contestability problem asks how to modify edge weights to achieve a desired strength for a topic argument. Gradient-based relation attribution explanations quantify the sensitivity of the topic argument’s strength to changes in individual edge weights and thereby provide interpretable guidance for adjustment; an iterative algorithm then progressively updates those weights toward the desired strength (Yin et al., 15 Jul 2025). Experimental evaluation on synthetic EW-QBAFs simulating the structural characteristics of personalised recommender systems and multi-layer perceptrons shows that the approach can solve the problem effectively (Yin et al., 15 Jul 2025).

These computational lines do not replace procedural design. They indicate how a technical system might expose reasons, accept objections, and implement redress, but the institutional questions of standing, authority, remedy, and accountability remain.

6. Limitations, controversies, and future directions

The literature is explicit that importing administrative-law ideas into algorithmic systems is difficult. Administrative law is “heavily rule-based and quite complex,” and algorithmic contestation systems may not need to be as complicated, although high-stakes decisions might require comparable depth (Lyons et al., 2021). Black-box opacity, trade secret claims, and the scale of automated decision-making all complicate the production of reasons, the exercise of review, and the design of remedies (Lyons et al., 2021).

One recurring controversy concerns whether contestability should be treated as a sufficient safeguard. Stakeholder submissions emphasize that contestability protects individuals and resembles contestability in relation to human decision-making, but they also warn that burdening individuals with the duty to initiate challenge may be unrealistic for vulnerable, less informed, or resource-constrained people (Lyons et al., 2021). This suggests that contestability must be complemented by ex ante protections, oversight, auditing, and legal enforceability. Otherwise, there is a risk of tokenistic contestability: nominal appeal processes that are opaque, slow, or lacking power to reverse decisions or provide remedies (Lyons et al., 2021).

Another major issue is the relation between individual redress and systemic accountability. Administrative-law-inspired contestation systems support individual rights and provide individual redress, but they may not reveal systemic problems embedded in an algorithm, such as biases that result in discrimination. Complaints and disputes therefore need to feed into audits, systemic error detection, and broader oversight (Lyons et al., 2021). The same concern appears in human-centred contestability research, which stresses that contestability should work alongside impact assessments, fairness checks, and human oversight rather than substitute for them (Lyons et al., 2021).

The research agenda remains open. One strand seeks refinement of contestation frameworks through thematic coding, content expert interviews, and analysis of other contestability systems (Lyons et al., 2021). Another seeks empirical understanding of user requirements, especially what information people need in order to contest an algorithmic decision meaningfully (Lyons et al., 2021). A further strand develops formal and computational mechanisms for dynamic redress, argumentation-based revision, and evidentiary grounds for reversal (Leofante et al., 2024, Freiesleben et al., 15 May 2026). Across these lines, the central claim remains stable: contestation is meaningful only when decisions are justified in forms that can be scrutinized, opposed, reviewed, and, when warranted, changed.