AI and Machine Learning in Anti-Corruption

Artificial intelligence and machine learning applications in anti-corruption work use statistical models trained on procurement, financial, and administrative data to flag transactions, vendors, or officials presenting an elevated risk of corrupt conduct, supporting both the Prevention and Detection stages of the anti-corruption value chain. The APEC Policy Support Unit's Issues Paper No. 15 treats AI and ML as the most consequential of the emerging technology tier, citing operational systems in Brazil, Thailand, Spain, and India, and a World Bank pilot in Brazil, alongside a sustained set of unresolved risks around data quality, transparency, and manipulation. ^[1]

1 Prevention applications

At the prevention stage, AI systems are used to risk-score vendors and bids before contracts are awarded. Brazil's Office of the Comptroller General (Controladoria-Geral da União, CGU) developed Alice, formally the "Analisador de Licitações, Contratos e Editais" (Analyzer of Bids, Contracts, and Tenders), an AI bot that screens public procurement documentation for anomalies. The paper reports that Alice has helped suspend more than BRL 9.7 billion, approximately USD 1.6 billion, in potentially problematic government purchases, a result the paper highlights as notable because it was achieved despite a 28 percent reduction in CGU's workforce between 2012 and 2022, suggesting the tool substituted effectively for reduced human audit capacity. ^[2]

Thailand's Anti-Corruption Organization of Thailand has deployed ACT Ai, a web-based application that links public procurement data with company registration records and declarations of politically exposed persons, allowing users to identify potential conflicts of interest between bidding companies and government officials before a contract is finalized. ^[3]

2 Detection applications

At the detection stage, Spain's Intervención General de la Administración del Estado (IGAE) has, with support from the OECD and the European Commission, built a data-driven fraud risk assessment system for public grants. The system applies supervised machine learning across 17 merged administrative datasets and was used to predict 42,152 high-risk grants out of a total pool of 1,050,470 grants assessed, allowing auditors to concentrate limited review capacity on the subset of disbursements statistically most likely to involve fraud. ^[4]

India's Centre for Data Management and Analytics (CDMA), established under the Comptroller and Auditor General, applies similar risk-based analytics to prioritize government audits, directing audit resources toward the entities and transactions data analysis identifies as highest risk rather than relying solely on rotational or complaint-driven audit selection. ^[5]

3 The World Bank's Governance Risk Assessment System

The World Bank's Governance Risk Assessment System (GRAS) was piloted in Brazil to screen public procurement data for indicators of collusion and undisclosed conflicts of interest. The pilot identified more than 850 suppliers showing strong indications of collusive behavior, approximately 450 suppliers likely registered under strawmen (nominee owners concealing the true beneficial owner), around 500 cases presenting conflicts of interest, and roughly 4,500 companies with detectable connections to political campaign financing. ^[6] The scale of these findings, drawn from a single pilot deployment, illustrates both the potential of large-scale data-driven screening and the scale of undisclosed ownership and conflict structures that conventional audit methods had not previously surfaced; the underlying ownership opacity these tools expose is treated further in Beneficial Ownership Registers.

4 Private sector compliance tools

The paper also documents private-sector AI compliance tools supplied to corporations for internal anti-corruption use. It cites Microsoft's development of AI-driven risk detection systems and an AI-powered compliance assistant, describing commentary from Barbara Tsai, Microsoft's Asia Head of Compliance, on how such tools are intended to help multinational firms identify bribery and corruption risk within their own operations and supply chains before regulators or law enforcement intervene. ^[7] This category of tool complements the internal governance mechanisms described in Internal Controls Against Corruption.

5 Challenges

The paper identifies four recurring challenges that temper enthusiasm for AI and ML deployment in anti-corruption work. First, data quality and bias: models trained on incomplete, inconsistent, or historically skewed administrative data can reproduce or amplify existing blind spots and inequities rather than correcting for them. Second, the "black box" problem: many machine learning models, particularly more complex ones, do not produce outputs that are readily explainable to auditors, judges, or the public, complicating both due process for flagged individuals or firms and public trust in the system's conclusions. Third, vulnerability to manipulation: sophisticated actors aware of a model's inputs and logic can structure transactions specifically to avoid triggering its risk flags, a concern paralleling the system-rigidity limitation identified for e-procurement platforms in E-Procurement Systems. Fourth, over-reliance on automation: the paper cautions against treating AI risk scores as a substitute for, rather than an input to, human investigative judgment, particularly given the mens rea and predication challenges documented elsewhere in corruption casework. ^[1]

See also: Anti-Corruption Value Chain, E-Procurement Systems, Internal Controls Against Corruption.

References

1. APEC Policy Support Unit, Technologies for Preventing, Detecting, and Combating Corruption, Issues Paper No. 15, June 2025, APEC#225-SE-01.10, section on artificial intelligence and machine learning.
2. Ibid., section on Brazil's Alice (Analisador de Licitações, Contratos e Editais), Office of the Comptroller General (CGU).
3. Ibid., section on Thailand's ACT Ai application.
4. Ibid., section on Spain's IGAE data-driven fraud risk assessment for public grants, developed with OECD and European Commission support.
5. Ibid., section on India's Centre for Data Management and Analytics (CDMA).
6. Ibid., section on the World Bank's Governance Risk Assessment System (GRAS) pilot in Brazil.
7. Barbara Tsai, Microsoft Asia Head of Compliance, as cited in APEC Issues Paper No. 15.