Can a public decision really be challenged if its decisive fact cannot be corrected?
By Karina Carlos, Founder of Self-Conquest
Consider a hypothetical public-benefits decision. An applicant receives a notice explaining that the household exceeds an income threshold. The notice names the program, states the result, and provides an appeal route. Yet the applicant cannot determine which income record was used, whether it covers the relevant period, or how to correct it. A caseworker can see the same result but cannot reconstruct its evidentiary basis. The decision has become visible. The path to testing its decisive claim remains uncertain.
This example is invented. Its purpose is to locate a governance question that deserves attention as public institutions automate decisions: what must a person be able to do with an explanation for that explanation to support a meaningful challenge?
Finding the point of disagreement
A public decision can be disputed at several levels. A record may be inaccurate. A valid record may have been matched to the wrong person. An inference may exceed what the evidence supports. A rule may have been interpreted incorrectly, or implemented in a way that excludes a relevant circumstance. These possibilities require different forms of review. Asking someone to explain why they disagree can place an unreasonable investigative burden on them when the institution has not made the decisive basis intelligible.
Esra Gules-Guctas examined 71 US federal and state court dockets concerning algorithm-based public-benefits determinations. The qualitative analysis identified failure mechanisms involving data, design, and system limitations. The cases illuminate how decisions become contested; they do not estimate how frequently public systems fail, and claims in litigation should not all be treated as adjudicated facts. The article appeared online in 2025 and in a 2026 issue of Public Administration Review.1
For governance, a useful question follows: can the institution locate the point at which evidence became a consequential determination? A readable notice may still leave that transition obscure. A score, category, or eligibility label cannot by itself show whether the underlying record, the governing criterion, and the conclusion fit together.
The work required to contest a decision
Even when a disputed fact can be identified, pursuing a correction takes work. Someone must know which document matters, obtain it, submit it through an accepted channel, and establish whether anyone has acted on it. A process can demand persistence, time, digital access, and familiarity with administrative language. Institutions should examine who bears those costs and whose challenge is likely to disappear before substantive review.
Karusala and colleagues studied contestation around housing-related public services in the Chennai and Boston regions. Their qualitative work describes sustained assistance from intermediaries as people navigate requirements and adverse decisions. Crucially, the settings studied did not use algorithmic decision systems. The paper informs questions about future design; it does not demonstrate that an AI appeal interface improves access or outcomes.2
The implication I draw is practical. An institution evaluating contestability should observe whether a person can advance a relevant challenge through the process. Availability of an explanation is one condition. So are a usable correction channel, appropriate assistance, a response, and clarity about what happens while reconsideration is pending. The applicable law determines particular rights and procedures; institutional design determines much of their practical usability.
A record that can support review
Cobbe, Lee, and Singh propose reviewability across the wider automated decision-making process, including its organizational elements. Their conceptual framework draws on English administrative law and extends from commissioning through consequences and investigation. It is not a universal statement of legal duties. Its value here is the question it makes possible: which records would enable the relevant person or body to examine this process?3
That question does not require exposing every technical detail to every applicant. Indiscriminate disclosure can create confusion and privacy risks. A useful account should identify the material basis for the decision, the relevant period and criterion, and a feasible route for challenging a suspected mistake. Access for an applicant, an internal reviewer, and an external oversight body may properly differ.
Nor should correction be reduced to changing a field. Suppose the hypothetical applicant supplies reliable evidence that the income period was wrong. The institution must still determine whether that correction changes the decision. Other grounds may remain. A defensible process would record what was accepted, what remained disputed, and why the resulting determination followed. Correction should neither guarantee a favorable outcome nor vanish into a database without reconsideration of its consequences.
Human Authority in the correction process
In the published Human Authority architecture, communicative legibility concerns whether judgment can be understood, challenged, and used. Institutional function concerns the arrangements that make responsible action possible. Applied here, I propose examining a connected sequence: an intelligible decisive basis, a workable route to correction or challenge, and reasoned reconsideration. This is an operational proposal informed by the literature, not a validated assessment instrument.4
The sequence also makes the limits of individual discretion important. A caseworker should not need to improvise an unauthorized exception to obtain review. The organization should specify who can inspect the record, request additional evidence, correct an error, and reconsider the determination. Where a legal or policy interpretation is disputed, the route may require a different authority from the person who maintains the data.
Institutions could test this sequence with appropriately protected sample cases before extending automation. Can the affected person identify the material issue? Can the reviewer reconstruct it? Does submitted evidence reach someone able to assess it? Can the organization trace a correction through any dependent decision? These are proposed design questions. Their effectiveness, workload, accessibility, and unintended consequences would need evaluation in the actual service.
A public institution has reasons to seek speed and consistency. Those goals should be considered alongside its capacity to discover and repair error. My proposition is that a challenge becomes meaningful when a relevant disagreement can enter the decision process and receive a reasoned response. The measure is what the institution can examine and change when the basis of its decision is brought into question.
References
1. Gules-Guctas, E. (2026; first published online September 25, 2025). How Do Algorithmic Decision-Making Systems Used in Public Benefits Determinations Fail? Insights From Legal Challenges. Public Administration Review, 86(2), 371–382. Source
2. Karusala, N., Upadhyay, S., Veeraraghavan, R., and Gajos, K. Z. (2024). Understanding Contestability on the Margins: Implications for the Design of Algorithmic Decision-making in Public Services. CHI 2024. Source
3. Cobbe, J., Lee, M. S. A., and Singh, J. (2021). Reviewable Automated Decision-Making: A Framework for Accountable Algorithmic Systems. FAccT 2021. Source
4. Carlos, K. (2026). Human Authority in the Age of AI. Source
Comments