The Next Five Years of Computational Antitrust

Review

Thibault Schrepel argues that computational antitrust has crossed from research agenda into an institutional field. His evidence is cumulative rather than experimental: a growing literature, recurring OECD and agency forums, cross-border exchange of code and methods, and five years of agency reports documenting tools deployed in investigations. The examples make the institutional shift concrete. Spain's BRAVA screens procurement for bid-rigging, Brazil's Cérebro has contributed to investigations that reached judicial scrutiny, Canada is building an in-house generative AI environment for sensitive case material, and several authorities now use computational systems for merger surveillance, document analysis, market monitoring, and detection. The paper's central governance contribution is therefore not the claim that agencies use more software. It is that computation is changing what enforcement institutions can perceive, what they treat as evidence, and consequently which exercises of public authority become possible.

The forward-looking argument identifies five questions: the legal framework for computational evidence, legitimacy and reproducibility of instruments, the economic assumptions antitrust should use, the cognitive effects of digital markets, and the boundary of automation. The legitimacy discussion is especially consequential. Schrepel ties an agency's authority to the contestability of its decisions: respondents must be able to understand and challenge the basis of an enforcement action, and courts must be able to inspect it. This reframes explainability from a model-quality attribute into a condition of lawful institutional power. It also exposes why the familiar assurance that a human remains "in the loop" is inadequate. When a system detects patterns across material no person can independently reperform, the human may validate a machine-produced finding without actually verifying it.

The paper is a field-building essay, not a systematic empirical evaluation of agency performance. Its evidence establishes diffusion, institutionalization, and examples of operational use, but it does not establish comparative accuracy, error rates, enforcement outcomes, or whether computational adoption improves procedural justice across agencies. The cross-agency reports are also participation-based, and the paper notes that the European Commission does not participate. Claims about the maturity of the field therefore rest partly on a network that the Stanford project itself helped convene. That does not negate the institutional evidence, but it limits what can be inferred about the entire enforcement landscape.

A deeper governance gap follows from the paper's own account. Contestability requires more than intelligibility at the point of appeal. Computational enforcement needs durable provenance for datasets, models, transformations, prompts, thresholds, software versions, and human interventions; rules for disclosure and protected access; mechanisms for challenging erroneous data and model assumptions; and clarity about when a computational finding may trigger coercive investigative powers. Cross-agency code sharing adds another layer: common tools can improve state capacity while also propagating hidden assumptions or correlated failure across jurisdictions. Reproducibility and benchmarking address scientific legitimacy, but public legitimacy additionally requires authority, procedure, redress, and responsibility for failure.

The automation question is therefore correctly framed as an institutional choice rather than a technical optimization problem. The durable issue is where decision rights migrate when machines determine what becomes visible to an agency. Computational systems may never formally issue the final decision yet still exercise substantial agenda-setting power by selecting cases, evidence, anomalies, and investigative priorities. The next phase of computational antitrust should consequently measure not only whether tools work, but where discretion has moved, what evidence survives challenge, who can reconstruct the computational chain, and which actor remains answerable when machine-mediated enforcement is wrong.

Key Insight

Computational enforcement redistributes public authority before doctrine catches up, so legitimacy depends on making machine-produced evidence contestable rather than merely keeping a human nominally in the loop.