Brazil: twice as many findings
The paper by Elliott Ash, Sergio Galletta and Tommaso Giommoni was published in the American Economic Journal: Economic Policy in May 2025. The authors trained a classifier based on gradient-boosted decision trees on the budget data of Brazilian municipalities and predicted where corruption would be found.
The authors tested the quality of this corruption measure separately: they repeated and extended earlier corruption studies, relying on the model’s forecast.
The main result comes from a simulation: if municipalities are chosen for audit by the model’s hint, the same number of audits finds almost twice as many municipalities with corruption as random choice.
For an audit body, this is an argument for risk scoring. Staff limits the number of audits, and the model helps send them where a finding is more likely. How such scoring works and how it is tuned is covered on the page on risk scoring.
The limit follows from the design itself: the model learns from past audits and inherits their blind spots. If auditors never found a certain type of violation before, the model will not point to it either.
Spain: a three-year forecast
Felix Lopez-Iturriaga and Ivan Pastor Sanz of the University of Valladolid built a neural network on corruption cases in Spain from 2000–2012. The paper appeared in the journal Social Indicators Research in 2017.
The model estimates in which provinces new cases are more likely to appear in 1–3 years. The features are economic and political: property tax, sharp growth in housing prices, the opening of bank branches, the creation of new companies and how long one party stays in power.
According to the model, the longer one party stays in power, the higher the probability of cases, whether or not it rules with a majority. For audit work, this is an example of a feature that is not linked to a single purchase but changes audit priorities in a region.
The authors stress the novelty of the approach: the model learns from real cases, not from corruption perception indices. They did not disclose the names of the high-probability provinces, to avoid disputes.
One of the authors gives his own warning: a high probability does not mean corruption will actually happen. The forecast works for planning audits, not for conclusions about specific people or agencies.
OECD: generative AI in audit bodies
In March 2024, the OECD released a report on generative AI for anti-corruption agencies, supreme audit institutions and internal audit services. 59 organizations from 39 countries took part in the survey.
About half (30) do not use generative AI yet and are only studying scenarios; another 24% (14) are at the development stage. Supreme audit institutions have gone furthest: 47% of them (8) are at least at the development stage.
76% of organizations (45 of 59) see the main benefit in work efficiency and the analysis of unstructured data, and 37% (22) put gathering evidence and checking documents first. In Brazil, language models search large data sets for signs of fraud; in Finland, France, Greece and the United Kingdom they prepare drafts, analyze tables and write summaries.
An example from the report: Brazil’s Federal Court of Accounts launched ChatTCU in February 2023, and by December 2023 the tool had more than 1,400 users. Auditors get summaries of case materials and answers to questions about the court’s decisions.
The main obstacles are a lack of skills and IT limits, and data protection is named just as often. Among the risks, the OECD names hallucinations, opaque conclusions and bias, and advises starting with low-risk processes: document search, summaries and answers to questions.
In March 2026, the Anti-Corruption and Integrity Outlook 2026 came out, covering 62 countries for the first time: 37 OECD countries and 25 partner countries. It gives data from a 2024 survey: 29 of 38 OECD countries use AI in tax administration, 76% of them to detect tax evasion and fraud and 69% to assess risks.
The World Bank and ProACT
In June 2021, the World Bank and the Government Transparency Institute presented the ProACT platform (Procurement Anticorruption and Transparency). It gathered open public procurement data from 110 countries for 2006–2020.
The platform scores contracts by risk indicators: single-bidder tenders, the length of the bidding period and the decision period, Benford’s law for amounts, the supplier’s share of the buyer’s purchases, the supplier’s registration offshore, delays in delivery and growth in the contract price.
Transparency indicators are counted separately: are the buyer, the supplier, the subject, the contract price, the procurement method, the number of bids and the decision date stated? Empty fields are a signal in themselves, because without a price or the number of bids, the risk of a purchase cannot be calculated.
The platform was launched as a prototype to collect feedback, and its named users were procurement staff, civil society groups, corruption researchers and international organizations. For a Russian agency, it is a ready list of indicators worth testing on its own data from the national procurement system.
Open Contracting: 73 flags
In 2024, the Open Contracting Partnership published a guide with 73 red flags for all stages of procurement: planning, tendering, award and contract performance. It is not a study with a model but a reference of signs on which rules and models are built.
The guide’s main caveat: a flag means risk, not proof of a violation. Single-bidder tenders happen in a narrow market, and an urgent purchase happens after an accident.
In practice, flags work as the first layer. Rules mark purchases, a model ranks them by risk, and the auditor reviews the top of the list. Which flags to calculate first is covered on the page on red flags.
The reference is also useful because it links each flag to data fields. If a field in the procurement system is empty, the flag cannot be calculated from it, and this is visible before the project starts.
Limits of the research
These works share common limits. Models learn from cases that people have already found, so they see familiar schemes with more confidence than new ones.
The OECD survey itself warns that its findings cannot be applied to all audit bodies. 59 organizations answered, and only one of them, from Brazil, reported an advanced stage of work with generative AI; many, by the authors’ estimate, did not answer because they had not worked with AI at all.
A model’s forecast applies to a group, not to a person. A high probability for a province or a municipality tells you where to audit, but says nothing about a specific official or company.
Savings figures that agencies publish themselves rarely pass an independent check. Research is more reliable in this sense, but it works with past data and does not account for how market players change their behavior after controls start.
Hence the rule for rollout: the model chooses what to check first, and the conclusion about a violation is made by an auditor, an investigator or a court. How this principle goes into the rules is covered on the page on risks and errors, and working systems from different countries are collected in the review of global experience.