Governing with Artificial Intelligence

AI in fighting corruption and promoting public integrity

Integrity actors — such as anti-corruption agencies (ACAs), supreme audit institutions (SAIs), internal audit and other oversight bodies — are reaping the benefits of adopting AI to enhance their operations, audits and investigations. As data availability intensifies and digital complexity increases, these institutions are forced to adapt to remain relevant. Faced with the increased use of technology by all actors, including those meaning to evade scrutiny and oversight, it is a matter of legitimacy for integrity institutions to keep up with the digital evolution.

AI has been a particular focus for integrity actors in recent years. This is due to rapid technical improvements making these technologies more available and affordable, as well as to the possibilities afforded by AI to process all types of data efficiently, uncover complex patterns and relationships, enhance analytical accuracy and precision, and generally augment current integrity activities (OECD, 2024[93]). As in other policy areas, AI is mostly valuable for the analysis of vast and complex sets of data, where computers can be more accurate or faster than humans.

Current state of play

Across integrity actors, AI has numerous potential applications, which are currently being used, tested or are anticipated to emerge in the coming years. The primary activities have been categorised into three categories based on their potential impact: detecting fraudulent activity; achieving efficiency gains in knowledge management and synthesis; and enhancing predictive analytics and forecasting.

Detecting fraudulent activities

One of the earliest and still most profound applications of AI by integrity actors lies in detecting fraudulent activities, transforming traditional analytical methods into sophisticated, data-driven approaches. AI algorithms excel at applying statistical techniques to identify outliers, patterns, transactions and behaviours that deviate from established norms that warrant further human investigation (OECD, 2024[93]; Taşdöken, 2024[94]). These tools are able to analyse numerical data across entire populations, a significant shift from traditional sample-based analytics, which enables the identification of misallocations or misreporting of funds, among other irregularities (Köbis, Starke and Rahwan, 2021[95]) (Box 5.27). In other examples:

• The OECD recently worked with the Spanish General Comptroller of the State Administration (IGAE) on an initiative proof of concept that employed advanced analytics and ML to detect anomalies and patterns indicative of potential corruption or fraud risks (OECD, 2021[96]).

• The European Court of Auditors (ECA), as part of an audit on EU lobbying activities, has used ML algorithms to check the transparency register of the European Commission and identify outliers with the objective of isolating a more interesting, risk-based sample for analysis by auditors (ECA, 2024[97]).

• In Brazil, the Alice tool analyses daily purchasing and procurement processes to uncover risk areas and inconsistencies. Unusual patterns trigger an alert that suspends the purchase, which is flagged for further inquiry (Box 5.25).

• In the United Kingdom, the Department of Work and Pensions is using AI to identify patterns in claims that could suggest fraud or error so that these claims can be reviewed by relevant teams within the department (Box 5.28).

Box 5.27. Europe’s DATACROS tool for detecting corporate ownership anomalies

In the European Union, Project DATACROS has developed an innovative AI-powered tool to detect anomalies in corporate ownership structures that may indicate risks of corruption, money laundering, and other financial crimes. Funded by the EU Internal Security Fund - Police and coordinated by Transcrime at Università Cattolica del Sacro Cuore, the project addresses the critical need for advanced analytical tools in public oversight and integrity efforts.

The DATACROS prototype tool consists of two main components: the Restricted Area for authorised users such as law enforcement and anti-corruption agencies; and the Public Area, accessible to all citizens. The Restricted Area employs AI algorithms to analyse the ownership structures of over 70 million companies across 44 European countries, identifying red flags and hidden patterns that may indicate illicit activities. It incorporates cross-border data and leverages extensive knowledge of criminal schemes, while adhering to privacy and data protection regulations.

Testing activity of the DATACROS tool in 2021 has shown promising results. The predictive systems correctly identified 83% of companies targeted by sanctions and 88% of companies with sanctioned owners. Partner agencies reported high satisfaction (4.3/5) and likelihood of future adoption (4.3/5). By providing both investigative capabilities for authorities and transparent aggregate data for public oversight, DATACROS represents a notable example of how using AI can help enhance public integrity and combat financial crime.

Source: https://www.transcrime.it/wp-content/uploads/2021/10/Datacros_report.pdf.

Box 5.28. Using AI to tackle benefit fraud in the United Kingdom

The United Kingdom’s Department for Work & Pensions (DWP) has been using AI systems since 2021-2022 to analyse historical data and identify patterns linked to fraudulent Universal Credit claims. It is investing GBP 70 million in advanced analytics between 2022 and 2025 to reduce fraud and error in benefit claims, with projected savings of GBP 1.6 billion by 2030. The systems adopted assess risk based on inconsistencies in reported income, unusual claim frequencies and mismatches with external records. High-risk claims are flagged for manual review, where payments may be paused until further verification is completed. Key areas of focus include self-employment claims, housing benefit fraud and undeclared capital, where fraud detection has historically been more challenging.

While these systems improve efficiency, DWP recognises that ensuring fairness and transparency remains a challenge. To mitigate bias, it conducts fairness analysis before deployment and continuously monitors flagged claims. However, gaps in demographic data make it difficult to fully assess whether certain groups are disproportionately affected. Caseworkers retain full decision-making authority and are not informed of why a claim was flagged to prevent confirmation bias. Random claim selection is also used to test the accuracy of the system and avoid overreliance on automation.

To strengthen public confidence, DWP has committed to publishing annual reports on the impact of AI in fraud detection, particularly concerning vulnerable groups. Efforts are underway to refine fairness analysis by improving data collection and adjusting risk-scoring methods where needed.

Source: (UK National Audit Office, 2023[98]).

Network analysis is another valuable technique for detecting potential fraud or similarly risky scenarios. It involves examining the interactions among entities within a system – such as public officials, contractors and suppliers – to unearth and visually represent potentially hidden relationships. This method is especially useful in identifying conflicts of interest, as well as in detecting corruption activities that were previously hard or impossible to detect (due to the volume and complexity of the data), providing crucial insights that help maintain the integrity of public operations. For instance, in procurement, pattern recognition can help identify sequences of transactions that may indicate collusion or kickbacks.1 Anti-corruption in this area is further discussed above, in the section on “AI in public procurement”.

Improving knowledge management and synthesis

For integrity actors, who dedicate a large portion of their workday to reading, processing and extracting meaning from text-based documentation, AI can assist knowledge management and text analysis. This can include, for instance, extracting meaning and automating fact retrieval from documents.

Within the realm of NLP, named entity recognition allows the rapid detection and classification of essential information in texts and other unstructured data formats — such as names of individuals, places, dates and organisations — at a speed and scale that would have been impossible to retrieve manually. Named entity recognition is used in chatbots, sentiment analysis, information extraction and fraud detection. Brazil, for example, integrated it into the Federal Court of Accounts (TCU) (Odilla, 2023[99]). Machine translation is also facilitating access to better textual information more quickly, thereby enhancing the effectiveness of audits and investigations.

The rapid development of GenAI is also being used to improve efficiency. LLMs can be of great assistance by allowing the reading, interpretation, summarisation and categorisation of vast quantities of textual information with a high degree of accuracy and in a fraction of the time of a human. Integrity actors are expanding their use of LLMs for internal and public-facing applications, making not only their work more efficient but also the public more engaged and aware of their efforts. Brazil’s creation of ChatTCU is one example (Box 5.29). Others include:

• Integrity actors in Finland, France, Greece, Lithuania, the United Kingdom and at EU institutions are using LLMs to support in the drafting of documents, analysing spreadsheets, summarising texts, processing of corruption reports, analysing case law or jurisprudence, and for discovery in performance audit (Ugale and Hall, 2024[89]).

• The Netherland’s Court of Audit is piloting a GenAI system to allow citizens and other stakeholders to roam through public reports and to find answers and sources to their questions on public audit work.2

• The ECA is piloting a project with three other SAIs (Netherlands, Germany and Sweden) to enhance interinstitutional relations and knowledge sharing by using LLMs to parse through existing national audit reports (across all European languages), automate translation and create English-language factsheets and guided summarisation.3

Box 5.29. Brazil enhances audit processes through generative AI with ChatTCU

The Brazilian Federal Court of Accounts (TCU) has pioneered the use of generative AI in audit and oversight activities through ChatTCU, an AI-powered assistant launched in February 2023. Developed as an institutional initiative, ChatTCU integrates with TCU’s internal systems, providing auditors with real-time access to case summaries, regulatory guidance and administrative support. As of December 2023, the tool had over 1 400 users, demonstrating its rapid adoption across the institution.

ChatTCU, currently operating on GPT-4 32k, enhances efficiency by allowing auditors to retrieve and synthesise information quickly, reducing manual workload while maintaining accuracy and security. Hosted on Microsoft Azure, the platform ensures data protection and prevents unauthorised sharing of sensitive information with external AI providers. Future developments include deeper system integration and workflow automation, further embedding AI into audit processes.

Beyond its internal application, ChatTCU is now part of an international technology transfer effort. In October 2024, TCU signed an agreement with the Supreme Audit Institution (SAI) of Honduras, making it the first international recipient of ChatTCU’s source code. This partnership, established through OLACEFS, reflects a broader strategy to strengthen AI adoption in audit institutions across Latin America and the Caribbean. Since its launch, 33 Brazilian institutions have also begun adapting ChatTCU for their specific needs, highlighting its scalability and adaptability.

Source: (Ugale and Hall, 2024[89]).

With GenAI just entering popular discourse in late 2022, integrity actors have had little time to comprehend the opportunity generative AI presents for their work, let alone to fully integrate it into activities. When the OECD (2024[89]) surveyed integrity actors last year, respondents generally described their organisations as being in the early stages of maturity concerning the use of generative AI and LLMs, if at all, though it is becoming a primary area of focus.

Preventive and predictive analytics, forecasting, foresight

Predictive AI systems can help anticipate potential corruption and fraud risks and other issues, enabling the prioritisation of cases for further human examination. While many integrity bodies lack the mandate for ex ante actions, identifying rising risks early can help mitigate them more effectively. Although ex ante actions are often out of the mandate of many integrity bodies, anticipating a rise in risks associated with corruption and fraud before they manifest can help mitigate these risks. For example, in Colombia, VigIA was developed by the Universidad del Rosario to analyse contracts from Bogotá’s Mayor’s Office and flag those with a high risk of corruption, anticipating potential illegal conduct. The tool is now adopted by the City of Bogotá (Salazar, Pérez and Gallego, 2024[100]).

Other researchers demonstrated how AI and data science techniques can be adopted to predict corruption. For instance, focusing on Brazil, researchers demonstrated that, in comparison to random audits, an AI system using budget data as predictors could detect almost twice as many corrupt municipalities for the same audit rate (Ash, Galletta and Giommoni, 2020[101]). Focusing on Italy, researchers used police archives and applied AI algorithms to predict white collar crimes demonstrating how algorithmic forecasts could support anti-corruption policy targeting (de Blasio, D’Ignazio and Letta, 2022[102]). In Spain, a research team also developed an AI system that serves as an early warning system to predict public corruption based on economic and political factors across Spanish provinces (López-Iturriaga and Sanz, 2017[73]). The transfer of these approaches from research to governments is still limited but growing. For instance, in Lithuania, the Special Investigation Service (STT) is developing an AI-based tool to assist corruption prevention officers with their legislative analyses (Box 5.30).

Box 5.30. Corruption prevention through the analysis of legal drafts in Lithuania

The Special Investigation Service of the Republic of Lithuania (STT), the country’s main anti-corruption body, is currently developing and testing a tool to help corruption prevention officers evaluate both enacted and draft legal acts for anti-corruption purposes.

The tool uses LLMs specifically trained on corruption risk data and risk assessment methodologies. Through the identification of potential legal corruption risk factors (e.g. loopholes, insufficient or weak procedures and measures), it helps officers rapidly and efficiently analyse large volumes of legislation in order to identify potential shortcomings or vulnerabilities with the legal acts.

By using this tool, STT seeks to strengthen legislative safeguards and ensure that when a draft legal act is adopted, the possible consequences of its implementation are carefully considered.

Source: Ongoing work with the OECD and STT (report forthcoming).

Evidence of impact

AI holds much promise for integrity actors, yet its empirical impact is still difficult to measure at this stage. Anecdotal evidence shows that AI enhances operational efficiency by saving time, allowing smart automations and increasing the accuracy of analyses. Use cases in which NLP, including LLMs, summarise, translate or extract information from text-based sources are among those where AI’s added value for integrity-related activities is the easiest to see and assess. However, quantifying the broader impact of AI in the field is complex. First, maturity in the field is generally low; many AI initiatives by integrity bodies are still at the exploratory phase (proof of concepts, pilot projects) and have not yet reached scale.

Additionally, assessing the current state of AI in this field is hindered by a lack of transparency and sharing of information and lessons learned between institutions and across borders. Information remains highly fragmented, and the sharing of use cases, best practices and knowledge about AI tools in integrity remains challenging. The challenges stem from a combination of factors, including trust issues around confidentiality and sensitivity of oversight activities and investigations, and cultural differences as well as varying levels of institutional maturity in adopting AI. Additionally, the rapid pace of technological evolution, coupled with limited resources in terms of both time and expertise, further complicates efforts to share information and learn from diverse use cases. This lack of openness hinders the ability to evaluate and understand the effectiveness of AI applications and learn from different institutions and contexts.

Managing risks and challenges

Associated risks

• Inadequate or skewed data in AI systems

• Lack of transparency and explainability

• Misuse or questionable use of AI

When deploying AI in public integrity and anti-corruption efforts, it is crucial to address the potential errors that can be embedded in AI systems, such as those caused by inadequate or skewed data (Köbis, Starke and Rahwan, 2022[103]). This is particularly critical in areas like fraud detection and conflict of interest analysis, where fairness and accuracy are paramount. Efforts continue to identify objective indicators that could serve as proxies for corruption, a phenomenon that – by definition – is hidden. While the most practical, the reliance on historical data (past cases or decisions) to train algorithms can perpetuate errors. This could result in skewed insights and limit the potential to uncover new and evolving dynamics of corrupt behaviour, potentially leading to unfair targeting or oversight practices, undermining the credibility of integrity efforts. To counter bias, AI systems should be developed with robust checks and balances, including rigorous testing, continuous monitoring and feedback mechanisms that allow for the identification and rectification of unintended errors and resulting harms.

The opaque nature of many advanced AI systems also poses significant challenges for integrity actors. If the public perceives government actions and decisions as opaque and perceive that they are relying on “black box” algorithms — where AI-generated outcomes lack explainability and transparency — citizens are likely to resist or distrust the use of AI (OECD, 2024[93]). In the audit profession in particular, its use to generate insights to be used as audit evidence is still a contended issue (Mpofu, 2023[104]). It raises question about their reasonableness, reliability and compliance with professional and regulatory standards (such as International Standards on Auditing, or ISAs) that have not yet been updated to account for recent technological changes.

The integration of AI technologies in integrity bodies raises significant concerns about responsibility, ethics and data protection, especially regarding personal data. The potential misuse of data, including inappropriate data merging or public disclosure of confidential information, poses serious ethical and legal challenges. Ensuring the protection of sensitive information, such as interest declarations, is paramount. Institutions should therefore navigate data protection, and ethical standards and regulations while implementing AI. Moreover, integrity bodies should adopt robust data governance frameworks and implement comprehensive data protection measures to mitigate privacy risks, including conducting regular audits, employing advanced encryption techniques, and fostering a culture of ethical data usage within their organisations (see Chapter 4, section on “Creating a strong data foundation”).

Implementation challenges

• Skills gaps

• Lack of high-quality data and the ability to share it

• Lack of comprehensive guidance and practical frameworks

• Risk aversion

• High costs of AI adoption and scaling

OECD (2024[89]) work has found that a shortage of relevant skills and experience is one of the biggest challenges in the adoption of AI in public integrity and oversight. This encompasses technical skills (both advanced digital expertise and the upskilling of general staff), as well as change management, data governance and leadership skills.

Access to comprehensive, high-quality data is a pre-requisite for the effective use of AI by any institution, yet integrity actors often lack data autonomy. While some integrity actors “own” the data they use (for example, interest declaration databases), many – including ACAs and SAIs – often rely on externally generated data over which they have little or no control. They depend on government rules and regulations, historical practices and memorandums of understanding (MoUs) with other institutions to acquire the data necessary for their activities. This means that integrity bodies often face delays in accessing the necessary data, hindering real-time analysis and timely decision-making.

Inconsistent and low-quality data can also severely impede the development and effectiveness of AI tools, making it challenging to create reliable and robust systems. Integrity bodies often rely on data from external sources, making the availability and accuracy of this data critical for successful implementation. The quality of externally sourced data is frequently unknown or uncontrolled until it is received, potentially leading to inaccuracies and inconsistencies that can limit or complicate AI development and system training. As quality data are foundational the effective development and use of AI, regulatory frameworks need to facilitate data sharing and collaboration between government agencies and across sectors, as appropriate and consistent with data protection and privacy rules, and strategies and systems should place to ensure data quality and completeness.

Data interoperability is also a barrier when it comes to sharing or combining data. The ability to cross-reference and analyse relationships within data is central to uncovering issues such as conflicts of interests. Yet currently, the maturity of most governments’ data platforms means that cross checking information is extremely difficult, if not impossible. Regulatory constraints – for security or protection reasons – may further impede the ability to integrate different data sources, thus limiting the potential for effective and comprehensive oversight.

Perhaps because of the nature of the work, integrity institutions are often risk averse. Many integrity actors want to integrate advanced technologies but are unsure where to begin or how to proceed. This apprehension is made stronger by the fear of making mistakes in the adoption process. Government guidelines frequently focus on cautionary advice, emphasising what not to do, rather than providing concrete, actionable steps for effective AI implementation. This lack of practical, implementable advice leaves many integrity bodies without a clear path forward, stalling progress and innovation. Comprehensive guidance and practical frameworks on the trustworthy use of AI for integrity actors, which are often lacking, actors could help.

Finally, the cost of implementing and maintaining AI tools poses a significant challenge, particularly for institutions with limited budgets and technical expertise. The complexity of ensuring robust cyber risk mitigation measures further complicates this landscape. In response, collaboration across institutions becomes essential. By using centralised resources, such as government-led AI sandboxes or shared tools, integrity bodies can reduce costs while accessing advanced technologies.

Untapped potential and way forward

Looking ahead, adaptive learning systems could continually refine their predictions based on new data they can play a significant role in updating risk assessments dynamically, which will allow integrity bodies to stay ahead of potential evolving threats. AI — particularly its ability to analyse and recognise patterns in non-numeric data — can also be the source of future improvements in anti-corruption. By combining it with sensor and image technology (satellites, drones, aircrafts capturing images or other data such as thermal readings or radar signals), AI creates new opportunities to monitor and analyse patterns on a large-scale to understand corruption-related activities and dynamics (Zinnbauer, 2025[105]).

Further down the line, agent-based modelling (ABM) by modelling entities (officials, businesses, etc.) and their interactions, could simulate how corruption and influence evolve in different conditions (U4, 2025[106]). Doing so could help test the impact of anti-corruption or integrity policies, laws and strategies before implementation.

AI-driven network analysis could also help more efficiently and comprehensively map the relationships between lobbyist, politicians, corporations and other actors, revealing influence patterns, conflicts of interest and revolving-doors dynamics. While integrating AI technologies into public integrity and anti-corruption efforts offers great potential, ensuring their effective and ethical use requires careful attention to key policy considerations that promote trustworthy adoption. Recent OECD (2024[89]) work provided three pieces of advice for integrity actors in exploring the use of AI:

• Start by incorporating generative AI into low-risk areas and processes. This approach can help build capacity in areas where mistakes are less costly — either financially or from a compliance perspective — before they scale to riskier and more analytical tasks, including those that require more financial resources.

• Consider the IT requirements for both piloting and scaling. Consider what computational and storage resources, data storage and data management capabilities may be needed and make sure early decisions do not overly complicate future plans.

• Consider internally generated or open data to demonstrate value and establish quick wins. This approach is low-cost and helps demonstrate use cases faster.

As government continue their exploration of AI, they should establish robust transparency and accountability mechanisms to ensure AI-driven decisions are not only accurate but also appropriately interpreted and documented. For instance, while AI systems can provide valuable insights by processing vast amounts of data, and future systems could be increasingly autonomous, the ultimate authority over sensitive decisions need to remain with human operators to preserve fairness, professional judgement (or scepticism) and ultimately maintain public trust. Public integrity bodies should seek to ensure that: the reasoning behind decisions is understandable both to experts and lay audiences; there are recourse and challenge mechanisms in place; and, to the extent possible, AI processes are explainable. AI-generated evidence needs to meet documentation requirements to provide an appropriate audit trail, including complying with current standards to demonstrate how evidence was obtained and assessed.

Knowledge exchange is also crucial for the successful implementation and scaling of AI technologies in public integrity roles. Institutions should participate in and create platforms that enable the sharing of best practices, strategies, lessons learned and even models, code and data. Sharing will help build a more comprehensive understanding of AI's effectiveness in integrity roles and drive the development of best practices that can be widely adopted. Initiatives like the OECD Anti-Corruption in Government division’s Tech and Analytics Community of Practice or the Tech Connect for Integrity initiative serve as examples.4 These forums bring together public and private sector practitioners to discuss challenges, explore innovative solutions, and develop collective strategies to enhance the adoption of AI and other technologies. This can allow for better flow of information, pooling of resources, rapid learning and faster development and implementation of methodologies and solutions.

Building on existing guidance on the ethical and responsible use of AI, governments should go beyond broad directives and provide more detailed, actionable guidance and frameworks for implementing AI technologies. This may include comprehensive step-by-step instructions, case studies and templates that institutions can tailor to their specific needs. Such guidance could also provide advice on identifying and mitigating bias in data, utilising explainable AI techniques, building transparency into decision-making processes and how to incorporate redress mechanisms.

A unified and interoperable data foundation is essential for the effective use of AI. Governments need to streamline their data operations to ensure that databases and systems can seamlessly communicate and integrate. This involves creating standardised data formats, establishing robust data governance frameworks and investing in infrastructure that supports data interoperability. By promoting data standardisation and integration, as well as policies for safe data sharing and crossing, governments can enhance the ability of AI systems to analyse data comprehensively, identify patterns and provide actionable insights to support public integrity.

Notes

← 1.  See also the section in this chapter on the use of “AI in public procurement”.

← 2.  Interview with the Dutch Court of Audit.

← 3.  Information provided by ECA officials.

← 4.  See https://www.oecd.org/en/about/projects/tech-connect-for-integrity.html.