The research report by Aleksej Heinze, Alfredo Jimenez, Julien Hanoteau and Virginie Vial (KEDGE Business School), with contributions from Joras Ferwerda (Utrecht University) sets out a clear picture of how digital transformation, and especially artificial intelligence, is being discussed in anti-corruption research. It shows an area showing a real dynamic, which still depends on stronger data, stronger institutional design and stronger implementation conditions.
AI can support anti-corruption action, yet its value depends on context, infrastructure, data quality, governance and human oversight.
The report was produced to provide BridgeGap a stronger evidence base for the next stages of work on artificial intelligence and corruption. It builds on earlier work on public procurement and widens the perspective to capture a broader range of anti-corruption applications. The aim is to understand where digital transformation already offers meaningful options, where research is still fragmented, and which knowledge gaps require direct attention. The review also is meant to support future work on adequacy, equity, explainability and effectiveness, so the project can move from broad technological promise to more grounded judgement about what works, in which settings, and under which conditions.
The disciplined conceptual frame separates digitisation, digitalisation and digital transformation as different levels of digital maturity, then connects each level to different forms of corruption and anti-corruption action. In this logic, digitisation is linked to petty and local corruption, digitalisation to more systemic and political patterns, and digital transformation to cross-border and grand corruption. This matters because it shifts the discussion away from generic claims about technology and towards a more precise reading of how digital maturity changes both risk and response.
Digitisation is the most basic level. An analogue process is replicated through digital means, without any redesign. A paper ledger becomes a database. A cash bribe becomes a bank transfer. The underlying behaviour is identical; only the channel has changed. At this level, digitisation primarily affects petty corruption — small bribes for routine services — because the activities involved are local, routine, and involve low-level actors.
Digitalisation goes further by optimising the process, not merely replicating it. Centralised digital procurement systems that reduce the discretion of individual officials are a digitalised anti-corruption mechanism. Encrypted digital wallets that automate bribe payments are a digitalised form of corruption. The scope extends to systemic and political corruption, affecting mid- and high-level actors.
Digital transformation is the deepest level: a wholesale reimagining of processes, culture, and governance. Cryptocurrency routed through the dark web represents digital transformation of grand corruption. Automated speed cameras that issue penalties without any human officer discretion represent digital transformation of anti-corruption enforcement. This level engages cross-border and grand corruption, involving political elites and national or international actors.
The review then adopts six analytical domains – people, data, software, hardware, processes and communication – to treat anti-corruption as an organisational system rather than a narrow technical problem.
Methodologically, the report is a structured literature review of 38 studies. It combines conventional database searches with supervised use of the Elicit AI tool to identify, classify and synthesise relevant material. The resulting picture is useful and definitely uneven. The strongest attention in the literature goes to software, especially tools linked to anomaly detection, transparency support and data-driven monitoring. At the same time, the review shows that the evidence base remains immature. Many studies stay at conceptual, exploratory or pilot level. Attribution is often uncertain. The literature points to potential gains, yet it also shows that anti-corruption outcomes cannot be inferred from technical deployment alone. The review therefore positions AI as a promising support layer, while keeping a firm focus on evidence quality and real-world constraints.
The most interesting feature of the review is its reflexivity. The team used AI tools to assist the literature review while simultaneously examining how AI tools are used in anti-corruption practice. That self-conscious quality — AI studying AI — is genuinely unusual and adds credibility to the methodological transparency claims. The AI-assisted outputs were manually verified throughout; the report notes that Elicit produced errors in reference metadata that required human correction, a finding that is itself informative about the current limits of AI-assisted research.
The Methodological Composition combines empirical case studies, surveys, systematic reviews, conceptual papers, and design science contributions. The critical observation is that conceptual and pilot-stage work predominates. Most studies describe what an AI tool could theoretically do in anti-corruption contexts, or what a prototype accomplished in a controlled setting. Very few evaluate fully deployed, operational systems over a meaningful time horizon. This is not a minor caveat — it is the central limitation of the field, and it means that conclusions about effectiveness must be held lightly.
The domain review brings the imbalance into full view. In the people domain, the literature gives more attention to adoption and perceived usefulness than to training, organisational resistance, public officials’ perceptions, citizen perspectives and trust. In the data domain, the report identifies a central bottleneck: fragmented datasets, weak data quality, scarce corruption-labelled ground truth and excessive reliance on perception indicators. In software, the field is dynamic, yet the report highlights unresolved issues around transparency, bias mitigation, explainability and fairness. Hardware is the most neglected domain, even though connectivity, energy supply and computing capacity are decisive in many contexts. The processes domain exposes limited knowledge on workflow integration, human-AI interaction and procedural embedding. Communication remains underdeveloped as well, with weak evidence on how systems should explain their outputs, support accountability and operate across agencies in a coordinated way.
Research on the human dimensions of AI anti-corruption tools is more voluminous than it is deep. The dominant framework is the Technology Acceptance Model (TAM), which measures perceived usefulness and ease of use. TAM is a defensible tool for understanding initial adoption decisions. It is a poor tool for understanding why institutionally embedded tools succeed or fail over time, because it does not engage with organisational politics, incentive structures, or the dynamics of bureaucratic resistance.
Low digital literacy appears across multiple contexts as a barrier but almost no studies provide tested, evidence-based guidance on training interventions that durably raise literacy and translate into changed working practices.
The data domain concentrates the most fundamental obstacles to effective AI anti-corruption work, and the review documents them carefully.
The “ground truth” problem is structural. Machine learning models require confirmed, labelled examples of the phenomenon they are trained to detect. In most domains, this data exists. Corruption is hidden by design — its perpetrators have strong incentives to ensure no reliable labelled dataset exists. This means that AI models are overwhelmingly trained on proxies. Training on proxies creates a compounding problem. Models trained on perception data reproduce and scale existing assumptions about where corruption occurs and who perpetrates it.
Software receives more research attention than any other domain, and the review finds that this attention has been genuinely productive: AI and machine learning tools have demonstrably improved anomaly detection in financial transactions, procurement systems, and tax records. These gains are real and should not be minimised.
The hardware domain represents the starkest structural gap in the literature, suggesting the implicit assumption that connectivity, computing capacity, and energy supply are background conditions rather than research subjects. This assumption is defensible in high-income, digitally advanced settings. It is not as a basis for research that claims global relevance.
Process-focused research occupies a sensible middle position between enthusiasm about automation’s potential and scepticism about its practical embedding in bureaucratic settings. The potential is documented the embedding is not. Whether automated workflows integrate effectively with existing administrative procedures, legal requirements, and professional norms is a question that most studies avoid or address only in passing. The implication is that legal and regulatory reform may be a prerequisite for effective AI deployment, not a downstream consideration.
The report’s core conclusion is well-supported: AI-driven anti-corruption tools are not plug-and-play solutions. Their effectiveness depends on the alignment of technology with local infrastructure, data quality, institutional capacity, and governance frameworks. The report argues that technical performance, legal and institutional settings, evidentiary strength and contextual conditions all shape results. False positives, false negatives, weak accountability rules, limited standards for bias control and thin long-term evaluation all remain material concerns. The strongest message is that effective anti-corruption technology depends on alignment. Systems need usable data, reliable infrastructure, embedded processes, capable people and governance arrangements that make outputs explainable and actionable. This gives BridgeGap a clear direction. The next phase is less about celebrating AI in the abstract and more about identifying the conditions under which it can genuinely strengthen transparency, detection and trust.
Here you can find the full deliverable for a more extensive insight: https://corruptiondata.eu/wp-content/uploads/2026/02/BridgeGap_D6.5_Review-of-digital-transformation-opportunities-and-challenges-for-corruption_V1.0.pdf