Competition Agencies Compute. The Law Still Assumes They Read.

All 27 EU national competition agencies, DG Competition, and most agencies worldwide now rely on computational tools to enforce competition law. Spain screens public procurement for bid rigging with BRAVA, a supervised machine learning system that maps relationships between firms, bids, and individuals. Brazil runs procurement documents through Cerebro to surface signs of collusion. France queries its case database in natural language through a retrieval augmented generation system built on large language models. Greece analyzes email metadata seized in dawn raids to establish which companies communicated with each other. Chile monitors more than 80,000 products for price anomalies. Poland investigates dark patterns with eye tracking and other neuromarketing methods. They all use AI. The field has a name, computational antitrust. It has moved from experiment to routine in under a decade.

The legal framework has not moved with it. The rules that govern how agencies collect and process information were written for a world in which investigators requested documents, read them, and drew inferences a court could retrace. Once an agency’s inferences are produced by models trained on data collected at scale, that assumption breaks in two places.

The data problem

The first break concerns collection. Computational tools are only as good as the data that feeds them, and the appetite of these tools exceeds what the current framework was designed to deliver. Regulation 1/2003 lets the Commission request information that is necessary for an investigation, and the Court of Justice has policed that standard through a line of cases on requests for information and inspections. The ECN+ Directive extended investigative powers to national agencies. Yet none of these instruments was drafted with continuous, large-scale, machine-readable data flows in mind.

The strain shows on both sides. Agencies that want representative datasets run into proportionality limits, business secrecy, data protection law, and the privilege against self-incrimination. Firms that receive sweeping data requests face the mirror question of where necessity ends. The result is a framework that neither equips agencies to collect what their tools require nor tells firms with precision what they can refuse. Whether Regulation 1/2003 and the ECN+ Directive need updating to resolve that tension, and how to do it without sacrificing the rights of the defense, is an open question. It is also, remarkably, an almost unstudied one.

The fairness problem

The second break concerns processing. AI systems inherit the biases of their training data, and the computer science literature on this point is now vast. Competition enforcement has barely engaged with it. When a screening model flags one market rather than another, the selection reflects choices about data and design that no one outside the agency can examine. When machine-generated analysis informs a decision, the duty to give reasons meets an explainability problem the case law on that duty never anticipated. The AI Act adds a layer of obligations, but its application to enforcement agencies using AI against private parties raises questions the text does not settle.

None of this counsels against computational antitrust. The tools make enforcement faster and, used well, more accurate. The point is narrower. An enforcement system that relies on computational tools without a legal framework for their accuracy and fairness is exposed. Exposed to annulment when courts start probing how the evidence was produced, and exposed to a legitimacy problem before that. Agencies know this. Several have said so publicly. What is missing is the doctrinal work.

A project, and three positions

That work is the purpose of ATLANTIS, a five-year project funded by an ERC Consolidator Grant and hosted at Vrije Universiteit Amsterdam. The project runs through three strands, on accuracy, fairness, and the institutional arrangements that sustain both. The specific research questions and methods will become public through the project publications, at the pace of the work.

I am recruiting three fully funded PhD candidates to build it with me, starting 1 December 2026. Two positions follow a legal track, one on the data problem and one on the fairness problem sketched above. The third is a computational-track position working across both, providing the empirical backbone of the team, with publications in law venues. A law degree is not required for that position. A willingness to write for a legal audience is.

The conditions are the standard Dutch ones, which is to say good. PhD candidates in the Netherlands are salaried staff under the Collective Labour Agreement of Dutch Universities, at €3,059 to €3,881 gross per month over four years, with pension and holiday allowance. The project adds what the grant makes possible. A personal travel budget. Fieldwork at competition agencies through the Stanford Computational Antitrust network, which gathers more than 80 agencies worldwide. Research stays with co-supervisors at institutions including Oxford, MIT, Stanford, and Harvard. Co-authored papers from year one.

For the legal track, the central requirement is a master’s degree in law with excellent marks and demonstrated expertise in competition law, as broad and as deep as possible. Empirical skills are welcome but will be taught. For the computational track, a master’s degree in computer science, data science, AI, or a closely related field, with documented experience in machine learning. And the ability to demonstrate the willingness to work on legal issues. Lawyers with strong computational credentials may apply to that position too.

Applications close on 15 September 2026 through the VU Amsterdam vacancies portal, under vacancy 5706. Project updates appear at teamatlantis.eu. One last argument, offered as a bonus. The campus sits at the edge of the Amsterdamse Bos, one of the largest urban forests in Europe, and some of our colleagues, myself included, train there. If you run, you will find us on the trails, where you will be invited to defend your publications at a conversational pace.

ATLANTIS is funded by the European Research Council under the European Union’s Horizon Europe research and innovation programme (grant agreement no. 101228709). Views and opinions expressed are those of the author only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them.