When two judgments agree, the agreement is usually taken as a reason for confidence. A second lab reproduces a result; two rating agencies give the same grade. The reasoning holds only if the two judges arrived at their verdicts separately. If both read from the same source, the second verdict adds nothing to the first. This paper is about that condition, as one synthetic record set stages it in an imagined world where a few providers come to supply most of the models that judge.
The record set belongs to ASI 2030, a studio in the RISCON family. Its short synthetic records, set between late 2026 and the end of 2030, each hold a question, a judgment by an AI system or another party, and a human auditor's opinion in the four grades of financial auditing: unqualified, qualified, adverse, or a disclaimer. The studio's imagined years run as one shared past that forks at the end of 2028 into two futures, and its records sit on eight axes, one of which it calls Concentration. All fifteen records on that axis are read here. Their companies, model families, blocs, accords and figures are invented; this paper studies how the records frame a decision and takes no fact or forecast from them.
The question and the contribution
The research question is this: in the records on the Concentration axis, what does the auditor treat as an independent check on a judgment, and what becomes of that check as the scenario's providers concentrate? Under it sit the series' own questions, what must be decided now and what would change the decision, here asked of judgments that a few providers make for many users.
The argument has three parts. First, across the whole thread, agreement offered as reassurance is turned back by the auditor's opinion: a low disagreement rate between two models, two vendors' similar errors, two platforms' matched complaint statistics and three platforms' consistent scores are each read as possibly one source seen twice. Second, the auditors keep asking who stands behind a source: whether the party that gains from a judgment also supplies its data, sells the tool that recommends it, or reports on its performance. Third, in one future rival providers remain and the audit's question becomes who receives audit capacity, and through which path; in the other a single provider is left, and the question every audit leans on, whether another model would give the same answer, has nothing to be put to.
The contribution lies between two bodies of work that seldom meet. One studies what happens to the people judged when many decision-makers use the same model. The other studies concentration in the markets, compute and infrastructure on which AI systems run. The records, read closely, stage a consequence of concentration that neither puts at the centre: it removes the comparison by which a judgment is checked. When the second source disappears, agreement loses its value as evidence, reproduction becomes impossible, and an auditor's opinion has to wait not on new data but on a new provider.
Monoculture and the people it picks out
The first debate concerns algorithmic monoculture. Jon Kleinberg and Manish Raghavan, Algorithmic monoculture and social welfare (2021), raise the concern in high-stakes screening such as employment and lending: if many firms use the same algorithm, even one more accurate than the alternatives, the result may be susceptible to correlated failures, as a monocultural system is in biology. Their model of selection goes further. Even without shocks or correlated failures, under normal operations, decisions may get worse when several firms use the same algorithm, so that a more accurate algorithm can lower social welfare, a kind of Braess' paradox for algorithmic decision-making.
Rishi Bommasani, Drew A. Hudson, Ehsan Adeli and others, On the Opportunities and Risks of Foundation Models (2021), names foundation models, trained on broad data and adaptable to many tasks, and argues that their effectiveness invites homogenization: powerful leverage, but the defects of the foundation model are inherited by all the adapted models downstream. Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky and Percy Liang, Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization? (2022), formalized one of the risks. Outcome homogenization is the extent to which particular individuals or groups receive negative outcomes from every decision-maker. Their component-sharing hypothesis holds that decision-makers who share training data or specific models produce more homogeneous outcomes; sharing training data reliably did so on fairness benchmarks, while for shared foundation models the results were mixed. Connor Toups, Rishi Bommasani, Kathleen A. Creel, Sarah H. Bana, Dan Jurafsky and Percy Liang, Ecosystem-level Analysis of Deployed Machine Learning Reveals Homogeneous Outcomes (2023), moved the analysis from one model to all the models deployed in a context and found some users misclassified by every available model, with improvements in single models rarely reducing such failure.
The normative side of this debate is set out by Kathleen Creel and Deborah Hellman, The Algorithmic Leviathan: Arbitrariness, Fairness, and Opportunity in Algorithmic Decision-Making Systems (2022). Arbitrariness in one algorithmic decision is not, on their account, of moral concern except in special circumstances; but when the same algorithm, or algorithms built on the same data, are used in many contexts, a person may be arbitrarily excluded from a broad range of opportunities, and that systemic exclusion is a moral concern. In finance, Jón Daníelsson, Robert Macrae and Andreas Uthemann, Artificial intelligence and systemic risk (2022), argue that AI taking over core functions can destabilize the system, creating new tail risks and amplifying old ones through procyclicality, unknown-unknowns, the need for trust and optimization against the system.
Concentration and the infrastructure of judgment
The second debate concerns concentration itself. Jai Vipra and Anton Korinek, Market Concentration Implications of Foundation Models (2023), observe that the most capable models tend toward natural monopoly with potentially vast markets, and call for antitrust authorities to keep the market contestable, in particular to stop monopolies from propagating vertically into downstream uses, and for regulators to hold the most capable models to quality standards, among them reliability and interoperability, since market discipline is weakened. Girish Sastry, Lennart Heim, Haydn Belfield and others, Computing Power and the Governance of Artificial Intelligence (2024), treat compute as a particularly effective point of intervention, being detectable, excludable and quantifiable and made through an extremely concentrated supply chain; governments already control its flow to competing countries, and it could serve regulators' visibility, allocation and enforcement, though naive compute governance risks centralizing power.
David Gray Widder, Meredith Whittaker and Sarah Myers West, Why 'open' AI systems are actually closed, and why this matters (2024), add that claims of openness often elide the concentration of large-scale AI development and deployment. Openness affords transparency, reusability and extensibility, and so some oversight and experimentation on top of existing models, but it does not by itself perturb the concentration of power, and open systems remain highly dependent on the resources of a few large actors. Jean-Christophe Plantin, Carl Lagoze, Paul N. Edwards and Christian Sandvig, Infrastructure studies meet platform studies in the age of Google and Facebook (2016), describe how platform-based services acquire the characteristics of infrastructure, the shared and widely accessible systems often provided or regulated in the public interest, while new and existing infrastructures are reorganized on the logic of platforms. This paper borrows that vocabulary for what follows.
Where the records sit
Between the two, this paper reads a question as left in the shade (a reading beyond what the notes above say of the works). The monoculture works study the people on the receiving end of shared models; the concentration works study the provider, as market power, a lever or an infrastructure. Neither, on this reading, centres the practice that sits between a judgment and its acceptance: checking it. A check needs something to compare with. Concentration acts on that practice in a specific manner. It does not only correlate outcomes or gather market power; it can make the second judgment a copy of the first, or remove it.
The ASI 2030 records on concentration are, this paper argues, largely about that practice. Their auditors keep testing whether two agreeing sources are in fact two, whether a rival system would reach the same result, and whether the beneficiary stands behind the source. The records do not test any claim in the literature, since they are synthetic. They do show, in an invented setting built for the purpose, what the studio's auditors take an independent check to require, and how that requirement fares in two imagined futures. That is the sense in which this paper adds to both debates: it reads concentration as a problem of verification.
One chapter sets out the records and the method, then reads agreement offered and refused as corroboration, the auditors' questions about who stands behind a source, and the two futures for the second source. The conclusion returns to the debates and names the limits of a reading built on synthetic records. An appendix gives the fifteen records word for word.