AI and Corruption: Who Exposes Whom First?

An auditor reviews a contract and finds no obvious violation. The price is defensible, the approvals are complete, and the supplier meets the requirements. But what if the same picture appears across dozens of contracts? Names recur, certain firms take turns winning, and separate tenders converge on a narrow circle of beneficiaries. The question then concerns the relationships between files, and what the soundness of each file conceals when it is read in isolation.

Here lies one of AI's most significant opportunities in the fight against corruption: expanding the field of vision. Its value extends beyond accelerating document review to detecting patterns that are hard to notice amid scattered transactions, multiple agencies, and years of data. A single exception may seem understandable, while its repetition in favor of the same parties raises a question worth investigating.

Recent research offers a striking direction. In a preliminary paper made available in late 2025, researchers Martí Medina-Hernández, Janos Kertész, and Mihály Fazekas examined public procurement data and records of sanctioned suppliers in Mexico. Combining machine learning with contracting-network analysis, they found that the position of a supplier and a contract within the network ranked among the model's strongest predictors. The implication is significant: relationships among counterparties can add to risk assessment what contract attributes alone cannot reveal. The finding remains tied to the study's context and does not make every repeated buyer-supplier relationship grounds for suspicion.

The study also points to a deeper problem: how do we teach a machine to distinguish corruption from integrity when we do not know all cases of corruption to begin with? A supplier who has not been sanctioned has not necessarily conducted all transactions soundly. If we treat everything undetected as clean, we may train the system to trust violations that have evaded oversight. The researchers therefore used Positive-Unlabeled Learning, an approach that learns from known cases and unlabeled data without assuming that every unknown case is sound.

This is as much a management question as a technical one. An institution that feeds its decision history into an intelligent model must first hold that history to account. What if some unjustified exceptions have persisted for years until they became routine practice? What will the system learn if recorded outcomes reflect the strength of relationships more than the quality of performance? In such a scenario, old bias may gain new speed and an objective-looking interface.

That is why I consider the most dangerous sentence to enter the decision room to be: "The system recommended it." It can end the discussion at the very moment it should begin. Who defined the evaluation criteria? What information was missing? Can the result be explained and reviewed? Handing a recommendation to an algorithm does not hand it responsibility for the decision, nor does it grant its outputs immunity from scrutiny.

AI is also not a single technology performing identical functions. Supplier-network analysis differs from unusual-transaction detection, and both differ from a language model that drafts a persuasive summary. In oversight work, we should demand of every tool what can be verified: a pointer to the source document, the precise location of a contradiction, and the ability to re-examine the result. A confident text, however eloquent, is no substitute for evidence.

Conversely, we must consider that generative tools could be used to craft more convincing justifications or misleading content in support of a suspect transaction. The U.S. National Institute of Standards and Technology (NIST), in its 2024 risk-management profile for generative AI, stresses building trustworthiness considerations into the design, use, and evaluation of these systems. In the anti-corruption context, I believe this means subjecting the oversight tool itself to scrutiny, controlling who may modify its data and models, and maintaining logs that allow changes to be traced and reviewed.

For this reason, I measure managerial advantage by an institution's ability to convert a signal into a responsible inquiry. When a system raises the risk level of a transaction, the alert must reach a qualified reviewer, a clear verification pathway, and protection of the parties' rights. A flood of alerts can create oversight congestion. Value lies in the quality of what is uncovered, the speed of remediation, and the harm prevented before it occurs.

Competitive advantage emerges when this capability safeguards fairness of selection and quality of spending, giving the most capable supplier a fair chance and preventing an unjustified exception from becoming a permanent cost. AI then becomes part of management quality and of stakeholder confidence in it.

AI can show us what deserves a question. The integrity of an institution lies in its willingness to ask it, whatever the name of the beneficiary.