0.1The question and the claim
A wrong decision about a person is wrong for a while. Between the moment a mistaken ranking, flag or denial is made and the moment it is put right, someone waits in the wrong place in a queue, keeps a badge that bars them from work, or goes home later than the timetable should have let them. How long that interval lasts, and who decides how long, is part of what it means for a decision to be correctable.
The setting for this question is invented. ASI 2030, one of the RISCON family's studios, writes scenarios in which machines move toward what it names artificial superintelligence by 2030, and it borrows a device from accounting to watch them: after each machine judgment about someone's case, a human auditor signs off with an unqualified, qualified or adverse opinion, or declines to give one. The studio groups its records under eight axes, and one of them is correctability. The scenarios are the studio's inventions. Nothing in them reports what has happened, and nothing in them is a forecast that this paper makes.
Of the correctability records set in 2029 and 2030, where the studio's story forks into two branches, this paper asks a question about timing. When a judgment about a person goes wrong, whose schedule governs its repair, how long does the person stay in the wrong place, and on what event would the case be looked at again? The series to which the paper belongs asks, in every domain, what must be decided now and what would make the decision change. Here the decision is whether a machine's judgment about a person is put right, and when.
The answer this paper argues for is that correction, in these records, is a question of who owns the clock. In the first branch, which the studio builds around audit law and cross-checking, the clock is one an auditor can read. Lateness is measured in the welfare records, and in two of them an earlier repair is invoked as a standard for a later case. In the second branch, which the studio builds around emergency delegation and sealed systems, the clock belongs to the provider. Reviews come when the supplier says or are left to its discretion, and the auditor declines to give an opinion in every case; one record assumes harm holds steady while no correction is scheduled. The records do not decide whether any judgment was right. They decide whether anyone outside the system can tell how long it will stay wrong.
The contribution is threefold. To the debate on contestability, the paper adds time: a reading of how a designed record set makes the pace of correction, and the identity of whoever sets it, the hinge of whether a decision can be contested at all. To the work on discretion passing from officials to systems, it offers a set of scenes in which the officials remain, but as the people who carry a question to a system on someone else's behalf. To the debate on transparency, it shows a record set in which disclosure matters less as a window onto the system than as the condition for starting a clock.
0.2Contesting a machine's decision
The first debate is about what it takes to contest an automated decision. Tad Hirsch and others, Designing Contestability: Interaction Design, Machine Learning, and Mental Health (2017), worked on an automated tool for assessing and training psychotherapists and named contestability as a new principle for designing systems that evaluate human behaviour, in settings where life, livelihood and wellbeing are at stake. Marco Almada, Human Intervention in Automated Decision-Making: Toward the Construction of Contestable Systems (2019), starts from the right to human intervention that data protection laws have established and finds two kinds of trouble with it: uncertainty about which decisions the right covers, and practical limits, since the person affected may lack the information needed to exercise it, and the human intervention may itself bring in biases. Almada's answer is that the means of contesting a decision must be a requirement at each stage of a system's lifecycle rather than an afterthought. Kars Alfrink and others, Contestable AI by Design: Towards a Framework (2022), take the same direction and gather from the literature the sociotechnical features and practices that make a system contestable into one design framework.
What contestability is remains in dispute. Henrietta Lyons and others, Conceptualising Contestability: Perspectives on Contesting Algorithmic Decisions (2021), analysed the submissions made on Australia's proposed AI ethics framework, the first of its kind to make contestability a core principle, and found its nature disputed, though commonly understood as a safeguard for individuals. Clément Henin and Daniel Le Métayer, Beyond Explainability: Justifiability and Contestability of Algorithmic Decision Systems (2021), argue that explanation is useful but not enough for the legitimacy of high-stakes systems, and that justifiability and contestability should be the key requirements; the paper draws a conceptual difference between the two.
A third position concerns what a person needs to act. Sandra Wachter and others, Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR (2017), treat explanation as a means to help a person act, and set it three aims: understanding why a decision was reached, grounds for contesting it, and a view of what would bring a different result next time. They propose counterfactual explanations, which describe the smallest change to the world that would produce a desired outcome, without explaining the system's internal logic.
The works read here place contestability in the design of a system, in the information given to the person, and in the standard of justification. This paper reads them, beyond what their notes say, as leaving one dimension in the background: time. A contest that can be lodged but has no date by which it must be heard, or that is heard by the system it contests, is contestable on paper. The ASI 2030 records put that dimension in front.
0.3From officials to systems, and the people left between
The second debate is about who exercises discretion once systems decide. Mark Bovens and Stavros Zouridis, From Street-Level to System-Level Bureaucracies (2002), follow large executive agencies as information technology takes over the work their front-line officials once did with wide discretion, until system analysts and software designers become the key actors and, in some realms, the officials disappear. The law is then implemented almost perfectly, but new questions arise: what of the discretion of those who design the systems, and how are due process and fairness guaranteed in difficult cases? Peter André Busch and Helle Zinner Henriksen, Digital Discretion: A Systematic Literature Review of ICT and Street-Level Discretion (2018), review 44 peer-reviewed articles on digital discretion published from 1998 to January 2017 and conclude that the scope of street-level work is shrinking as more agencies become digital bureaucracies, with officials operating computers instead of meeting clients face to face; such digital discretion, they suggest, can strengthen ethical and democratic values while weakening professional and relational ones.
The ASI 2030 records do not show the street-level official vanishing. They show a welfare officer, a nurse, a caseworker, an outreach worker and a counsellor each standing between a system and a person, putting the question the person cannot put. This paper reads those scenes as a third arrangement beside the two that Bovens and Zouridis describe: the official remains, but as the one who asks, not the one who decides.
0.4Seeing without starting the clock
The third debate is about transparency. Mike Ananny and Kate Crawford, Seeing Without Knowing: Limitations of the Transparency Ideal and Its Application to Algorithmic Accountability (2016), question the long habit of equating the ability to see a system with the ability to know how it works and to govern it; they set out ten limitations of transparency and sketch a view of accountability that starts from those limits. Jenna Burrell, How the Machine 'Thinks': Understanding Opacity in Machine Learning Algorithms (2016), writes about classification systems such as credit scoring and loan qualification and separates three sources of opacity: secrecy chosen by a firm or a state, a lack of technical literacy in the people looking, and the nature and scale of machine learning itself. Knowing which form is at play, she argues, decides which remedies could help.
Both works bear on the second branch of the ASI 2030 records, where in four of the six records the evidence entry itself says the evidence is classified or undisclosed. Read against the background Ananny and Crawford give, the records show a narrower use for disclosure than seeing inside a system: knowing when it must look again.
0.5What the records can bear
The records bear on these debates in one way only: they show what a studio makes correctability mean when it imagines it under two sets of institutions. They cannot show how real agencies correct decisions, how long real people wait, or whether an artificial superintelligence will arrive by 2030 or at all. Every figure in them, from a median delay in days to a confidence level, is synthetic and is read here as a feature of the scenario.