0.1The question and the claim
An audit opinion is a short sentence that has to carry a great deal. In financial auditing it tells a reader, in one of four set forms, how far the auditor will stand behind what a set of accounts says. ASI 2030, a studio in the RISCON family, borrows that grammar for a different object. Its records are synthetic scenarios of the years 2026 to 2030 in which AI systems become what the studio calls an artificial superintelligence, and in which every judgment such a system makes is followed by a human auditor's opinion: unqualified, qualified, adverse, or a disclaimer. The scenarios are the studio's inventions. Nothing in them is a report of what has happened or a forecast that this paper makes.
This paper asks a narrow question of those records. When the judgment under audit is one that claims to have been verified, how does the record set use the auditor's opinion to frame it, and what does each record say would change the opinion? 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 the auditor's: whether to stand behind a machine's judgment, and how far.
The answer the paper argues for is that the four opinions, in these records, are sorted by what the auditor could rerun. Where a judgment could be reproduced only in part, on one date or for standard cases, the opinion is qualified. Where a rerun was made, by a second model, a cross-audit or an outside team, and it contradicted the judgment, the opinion is adverse. Where no rerun could be made at all, because a scoring log, a method or the real operating conditions were out of reach, the auditor disclaims. Whether the judgment was right is a question the records mostly leave open; whether it could be rerun is the question they answer. A second finding concerns what would change an opinion. The first of the studio's two branches for 2029 and 2030 mostly gives conditions that are tests with thresholds. The second mostly gives conditions that are already met or cannot be run, and says no procedure exists by which the opinion could act on them.
The contribution is threefold. To the debate on algorithmic auditing, the paper offers a reading of how a designed record set makes access to a rerun the hinge of an audit, which sharpens the difference between auditing from inside and from outside. To the literature on audit opinions, it shows a record set using the disclaimer as an informative verdict rather than as a failure to conclude. To the literature on scenarios, it reads a pair of branches as an instrument that varies the conditions of audit while holding the questions steady, and says what such a pairing can and cannot show.
0.2Three debates
The first debate is about who audits an algorithm and with what access. Danaë Metaxa and others, Auditing Algorithms: Understanding Algorithmic Systems from the Outside In (2021), describe the algorithm audit as a method of querying a system again and again and observing its outputs in order to draw conclusions about workings that cannot be seen. They trace the method to audit studies in the social sciences and note the obstacle that frames it: algorithmic systems change constantly, often silently, and leave no record of what users were shown. Inioluwa Deborah Raji and others, Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing (2020), start from the same observation, that most audits of deployed systems have come from investigators outside the organisations that run them, and propose a different remedy: an internal audit applied through the whole development life-cycle, in which each stage yields documents that together form an audit report, to help close the accountability gap by embedding a process that protects audit integrity. They also note that once a system is deployed, an emerging problem can be hard or impossible to trace back to its source. The two positions do not exclude each other, but they place the auditor differently. One works with what a system does when asked; the other works with what the organisation kept.
Two further works define what an audit can be asked to deliver. Jakob Mökander and others, Ethics-Based Auditing of Automated Decision-Making Systems: Nature, Scope, and Limitations (2021), treat the audit as a structured process that assesses an entity's present or past behaviour against principles or norms, a way for an organisation to validate claims about its systems; they set out the constraints such auditing meets and conclude that it can be only one part of a wider approach. Miles Brundage and others, Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims (2020), put the weight on the claim itself: developers must make claims that can be verified, and people outside the organisation need effective means of scrutinising them. Maranke Wieringa, What to account for when accounting for algorithms: a systematic literature review on algorithmic accountability (2020), organises the field by Bovens's five points of accountability, among them the forum to which an account is given and the consequences that may follow from it. And Inioluwa Deborah Raji and Joy Buolamwini, Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products (2019), show one route by which an audit can have consequences: after a published audit of facial analysis systems, the companies named released new versions within months and narrowed the disparities the audit had found.
Underneath both positions lies reproduction. Odd Erik Gundersen and Sigbjørn Kjensmo, State of the Art: Reproducibility in Artificial Intelligence (2018), surveyed 400 empirical papers against the variables a result needs documented if anyone else is to reproduce it, grouped as experiment, data and method; none documented all of them, and between 20% and 30% of the variables for each factor were documented. This paper reads the survey, beyond what it reports, as a reason to expect that a judgment is hard to rerun when little of it was recorded, and that an auditor who must rerun one depends on what was recorded.
The second debate is about what an audit opinion communicates. Bryan K. Church and others, The Auditor's Reporting Model: A Literature Overview and Research Synthesis (2008), trace how the auditor's report developed and review research on the decision an auditor makes in choosing a report and on what the report contains. Glen L. Gray and others, Perceptions and Misperceptions Regarding the Unqualified Auditor's Report by Financial Statement Preparers, Users, and Auditors (2011), report from focus groups that users value the audit but do not read the whole report, and that preparers, users and auditors alike are unsure what the report is meant to say or how much assurance it gives: the audit expectation gap. Michael Power, The Audit Society: Rituals of Verification (1997), takes a step back, to the spread of audit itself, the means by which audits are made to work (samples, specialists, systems), the work of making an auditee's performance auditable, and the relation of audit to trust and risk; its subtitle calls audit a ritual of verification. This paper reads Power's subtitle as a warning that the form of a verification can outrun its substance; that reading is this paper's, and one of the eleven records stages exactly that gap.
The third debate is about what scenarios are for. Ron Bradfield and others, The origins and evolution of scenario techniques in long range business planning (2005), describe a field rich in techniques and contradictory definitions, which they sort into three main schools. Rafael Ramírez and Cynthia Selin, Plausibility and probability in scenario planning (2014), argue that the choice between treating scenarios as plausible stories and treating them as probable outcomes has divided those schools, and propose to move past the either/or. The ASI 2030 records are scenarios in this sense. Reading them as predictions would misread them; reading them as instruments that hold a question steady while the world around it changes is what this paper attempts.
0.3Where the records bear and where they cannot
The records bear on these debates in one way only: they show how a designed set of cases distributes judgments, opinions and conditions, and so what an audit grammar borrowed from accounting is made to carry when its object is a machine's judgment. They cannot show how real auditors would behave, how real AI systems perform, or whether an artificial superintelligence will arrive by 2030 or at all. The reproduction rates, agreement figures and confidence levels inside them are synthetic and are read here as features of the scenario. The paper sets out its records and method first, then reads the qualified opinions, the adverse opinions and disclaimers, and the conditions for change, before turning in its conclusion to what the reading adds and where it stops.