Artificial intelligence and anti-corruption – time to check-in on new technology

Artificial intelligence (AI) has been developing in leaps and bounds over the last two decades.e8c88a8807d2 This rapid progress has also increasingly captured the imagination of the anti-corruption community, who have long sought to harness the potential of digital technologies for anti-corruption work. These technologies have had an evolving set of thematic labels from ‘e-government’ and ‘smart cities’d642eab0ad0f to ‘big data’,b641917ce86a ‘crowd-sourced transparency’,8486cf34f482‘algorithmic accountability’,2da305d42e94 or ‘integrity via blockchains’.f1767197b86f Around 2020, a flurry of reports – including by U4 – began to more systematically map the implications of contemporary AI systems for anti-corruption work.4be730192b40 These important contributions had to rely mainly on early-stage experiments or hypothetical cases, as actual implementations, impact assessments and empirical evidence, were still some way off.d4738211cf89

These early accounts also predate the latest transformational breakthrough in AI: the arrival of ChatGPT and other so-called generative, general purpose AI. These AI systems are capable of interacting with non-expert users and producing human-like outputs in a wide range of application areas. Generative AI has experienced the fastest adoption rate of any new technology ever and drives a fresh wave of AI investments, experimentation, and related policy conversations to new heights – and hypes. This U4 Brief provides an update on some important developments in use of AI in anti-corruption. It focuses on important emerging lessons, and points to two overlooked application areas that are particularly fertile. It also emphasises policy implications and potential engagement opportunities for donors at this stage of AI in anti-corruption deployment.0268cebce294

The emergent learning landscape – solid progress but no magic solution yet

The expanding evidence base on AI in anti-corruption yields a number of emerging insights that often reaffirm experiences with prior generations of digital technologies. At the ‘big picture’ level these include:

• A proof of concept is not a failsafe ticket to sustainability or diffusion: several early examples of AI applications in anti-corruption have been discontinued or shown limited scalability. Limiting factors such as resource constraints, poor data quality, dwindling citizen engagement, rapid technical obsolescence, ill-suited regulation, or poor integration of AI solutions into existing anti-corruption infrastructures are well-known problems for many types of digitisation projects.0fa13d505f97 It is not surprising that they also pose obstacles for AI supported applications.3a1221fed693
• A political economy and institutional lens merit more attention, when assessing technological opportunities and implementation strategies. No anti-corruption technology will succeed if the very actors tasked with deploying it are those who also benefit from corruption.2affb4a1ba30 Teasing out the implications of AI for specific configurations of interests and power is essential for drawing up effective implementation pathways. It may be necessary to build on an expanding mix of approaches to analyse political economy contexts or institutional opportunity structures.
• Digital divide issues are becoming even more critical. Women, ethnic minorities, and other marginalised communities are underrepresented in AI research. These groups are also often mischaracterised in historical training data: they are confronted with ill-tailored applications at best, and biased, discriminatory AI outcomes at worst.6484290cbc1a

Several application areas are prolific hives of experimentation and learning. Two of these are profiled below: ‘classic’ anti-corruption areas; and reducing corruptible discretion in decision-making.

Making good progress: AI in ‘classic’ anti-corruption areas

AI is making good progress in some ‘classic’ anti-corruption areas: procurement, compliance, collusion, and anti-money laundering (AML). And AI continues to expand capabilities in areas where digital technologies already make a sizeable impact. AI is making significant contributions in a number of application areas97fa951f9f1c – particularly with its ability to process and connect unstructured information at scale, and its flexible approach to unsupervised learning.

For integrity in public procurement, the use of AI can build on a long track record of digitising procurement data and processes. It also helps to advance capabilities in three main areas:

• AI makes it possible (in principle) to identify new and flexibly adjust existing red-flag indicators. This makes it more difficult for criminals to game the system. It is easier for law enforcement to catch up with new fraudulent tactics, and so possibly also reduce undue risk aversion in procurement.16e725420e23
• AI enables a new generation of analysis and monitoring initiatives to consider a much broader set of data inputs – from insolvency to political connections, open ownership to open contracting vanguards, and asset disclosures to political sponsorship information – to detect more complex patterns of collusion746476a1b5a7and conflicts of interests.3d53e5ee50d9
• AI paves the way for procurement analysis at unprecedented scale and scope.

These emerging efforts can take advantage of a number of initiatives that are assembling big data repositories and large-scale red flag assessments of procurement data – for example, by compiling 17 million tenders across the European Union (EU) (DIGIWHIST)961000a910a0 or by collating data from no less than 42 national procurement systems (ProACT).ad8db1796cdd

In the broader areas of compliance, fraud detection, and anti-money laundering, AI helps to turn episodic spot-checks and select audits into more efficient, comprehensive and real-time monitoring efforts.3d1cc3fb9e67 For instance, a global beverage conglomerate consolidated more than a dozen internal enterprise resource management systems with a number of external data streams. The result was a consolidated, AI-supported supplier vetting function that reduced costs by more than 90%.ad20694ea674

In Peru, investigators use AI to screen a growing volume of reported suspicious financial transactions. The process is much more efficient, doubling the rate of cases that could be referred to prosecutorial authorities.861a7a42ba6a Similarly, a global bank has cut the incidence of false positives in fraud reporting by 75%.ea415716df75 Another bank doubled its detection rate of confirmed bad transactions and cut transaction processing time from more than a month to just a few days.cf926aab926a

Mixed success: AI in administering social benefits or reducing corruption-prone discretion in other areas

It is clear that AI can help administer standard cases and free up human resources for more complex issues. However, high hopes that AI could play a major role in cutting corruption at public service level by reducing corruptible discretion in decision-making have so far remained unrealised for several reasons, including:

• Eliminating discretion remains an ill-conceived ambition: retaining some discretion is a desirable principle in administrative decision-making to arrive at fair outcomes that are tailored to individual circumstances.
• AI systems were found to produce many incorrect decisions in several high-profile implementations – from denying unemployment benefits to thousands of rightful claimants in Michigan, causing dire economic distress,18743e145586to wrongfully withdrawing social benefits in Serbia,9a520f2a344b or child benefits allowances from Dutch parents – a scandal that ultimately led the government to resign.a3b757b0292e

These shortcomings exemplify some fundamental problems with AI in this context. AI operates largely as a ‘black box’, partly due to technical complexity, partly due to proprietary ownership of AI models and data. Because AI decisions cannot be fully explained, this violates a basic principle of administrative justice. Biases in training data and the persistent, significant tendency to make things up further undermine accuracy and fairness. (‘Hallucination’ is the technical term used to describe an AI-generated response that contains misleading information presented as fact.) Even AI systems that specialise in legal issues are found to ‘hallucinate’ in up to 30% of cases – for example, referencing legal clauses that do not exist.7061219a37a9

All these issues suggest that AI can provide decision support, but that a ‘human in the loop’ as ultimate decision-maker is a prerequisite for achieving just and accurate outcomes – and also to establish clear lines of accountability when things go wrong. As much as discretion might invite corruption, an over-reliance on AI at this stage is not desirable and produces materially bad outcomes. Also, AI systems can potentially be tweaked to produce particularistic outcomes (those that favour a particular population or group).

These shortcomings are not insurmountable, and many efforts are underway to mitigate them – from devising assessment frameworks for model transparency,c3cdf206c643 to ethics benchmarks,6d8732e56647 and novel approaches for explainability.ead90da8e094 However, given the rapid development of AI, many of these initiatives are in catch-up mode, and issues continue to persist or reappear in different forms. Also, in the quest for proprietary advantage, many cutting-edge AI models do not include sufficient public disclosures.07663e386a45

New anti-corruption spaces for AI

So far, much attention in policy and practice has focused on ‘classic’ anti-corruption applications – from compliance to social accountability. But there are also a number of secondary application areas where emergent evidence shows a growing impact and future high potential for use of AI in anti-corruption. Two areas are profiled below: remote sensing; and inclusive participation.

Remote sensing with anti-corruption momentum

Over the last ten years, AI systems have developed their ability to extract information from images and recognise complex patterns. In parallel, we have seen a rapid expansion of available satellite imagery and earth observation data. The number of satellites in space has tripled within the last five years to reach 10,000.161260ebe93d These satellites now cover Global North and Global South countries and produce reams of up-to-date data at an unprecedented level of geographic detail. This opens up novel opportunities to harness AI for making progress on tackling a number of illicit activities that are particularly consequential, deeply interwoven with corruption, and particularly prone to impunity.

For example, satellite-based monitoring of illicit logging and deforestation has improved significantly, enabling sub-weekly updates at a spatial detail level of 10 to 50 metres.364750947a4f Related monitoring initiatives are now able to map deforestation in much more detail, and also track specific activities that drive it, including illegal gold mining in the Amazon,adab03e4c5b1 land grabs in Ghana,2aa6ac20c960 or illegal coal mining in China.336eca7ced69 A new cohort of alert systems (currently in experimental stages) also use the predictive power of AI to identify early warning signs of deforestation – for example, road buildings or adjacent infrastructures such as mills or silos8c9cfd59519d – and can cut response times significantly.

These initiatives are already bearing fruit. Forest Foresight – a collaboration between academia and a non-governmental organisation – claims to be able to predict illicit deforestation months before it occurs. In a trial implementation in Gabon, it helped park rangers carry out 34 enforcement actions and stopped an illicit gold mine.7ab7048a3046

Combined with other datasets, some of the underlying corruption dynamics can be much more effectively documented. Deforestation in Brazilian communities was found to spike by another 8% to 10% in election years when an incumbent mayor runs for re-election, pointing to the use of forbearance as a political tool (where those in authority withhold certain sanctions to gain votes).a143a3c504b9 Meanwhile, enforcement action by federal-level agencies during the 2019 to 2022 presidential term was identified to be minuscule, with less than 1% of incidences receiving a response.aa344f361870

Similarly, AI-supported remote sensing analysis is being used to better understand, track, and create the prerequisites for more accountability in a wide range of corruption-prone application areas, such as :

• Tripling the productivity of rangers to detect snares in Cambodia017efa6e9bc7
• Assigning responsibility for remote oil spills in the Mediterraneanaf27a356b2ef
• Revealing fake suppliers in procurement in Brazil1791bdb3e9e3
• Illegal bitcoin miners in Iran14b6dd957148
• Illicit fishing activities around the worlda9fda8a75d54
• Ethnic patronage in Kenyan informal settlements24f3dc716b27
• Bureaucratic incentives for harmful crop-burning in India.3323c95b3b4c

AI has also helped to detect and locate (at very detailed level): misreporting of economic activity in China0bb223e4d6a4 and of methane emissions in the USA;1088c84b6633 and uneven, politically motivated government response to disasters in Brazil, Indonesia, Mexico, and South Africa.9028ef3a9efd

Concerns remain about the availability and affordability of particularly high-quality satellite data. However, at least three characteristics make this type of AI-enabled remote-sensing particularly interesting from an anti-corruption perspective:

• It can bring transparency to areas and issues that are too remote or too tightly controlled by corrupt power brokers to be effectively monitored.
• It often makes this data available to civil society stakeholders and researchers outside the reach of repressive governments.
• It operates at scale, at near real-time frequency, and at a level of geographic detail that helps directly assign responsibility – for example, by pinpointing a specific polluting source, deviant actor, or local council.

Also, remote sensing via satellites comes with a global geographic scope that does not leave out the Global South which is typically underrepresented in many data collection efforts.

Protecting against ‘policy capture’, and promoting inclusive participation

Anti-corruption and integrity systems often seek to counter risks of policy capture by vested interests through various types of public consultation, participation, and feedback systems. However, in many contexts, such mechanisms have become almost impossible to manage. Electronic submission formats have increased the number of feedback submissions and comments. Some are mass-organised, and mechanisms are increasingly targets of attempted manipulation by special interests or propagandistic actors.e23cc741800d

For example:

• The US Consumer Financial Protection Bureau receives more than 1 million comments annually.
• The public consultation for a new Chilean constitution drew more than 280,000 individual comments that had to be hand-coded in a very short time frame.
• The EU consultation on the future of Europe attracted more than 50,000 participants and 17,000 ideas that were summarised by consulting firms into lengthy reports.

Such ‘mega-consultations’ typically face two principal problems depending on their main objectives:

• The ‘haystack challenge’ to find a small number of comments that contain very high informational value among a flood of low-quality submissions.
• The ‘forest challenge’ to identify, consolidate and numerically weigh similar concerns and opinions for an overall overview of where public sentiments are.4ec606fcb287

AI’s ability to categorise and summarise very high volumes of natural language text offers ways to address these problems and make processes meaningful and effective. For example, the Consul Democracy software platform is used by more than 300 cities and organisations around the world to help with these tasks.a6bcb442fe8f Its early iteration, pioneered in Madrid, attracted more than 26,000 proposals. Without the use of AI, time-consuming summarisation and consolidation of information meant that only two proposals reached the threshold to be considered by the city council. A new, AI-supported version of the platform allows for much better search and consolidation of similar results.b9cd4cbdde93 Overall, such AI-powered consultation summary systems are becoming better. AI systems are also demonstrating their usefulness in helping to moderate deliberative processes or negotiations that have previously been bogged down by polarisation.1e8ea0ac9251 Yet AI continues to face problems, particularly with regard to non-English language applications.dfbe892cfeb1

Nevertheless, this bodes well for AI helping in very specific applications to tackle some important challenges that affect participatory and deliberative mechanisms. AI can also help to ensure that the vital pillars for inclusive governance are not hijacked by those with special interests or overwhelmed by the scale of inputs.

Implications for anti-corruption donors and practitioners

With AI in anti-corruption moving increasingly from proof of concept and experimentation to a more mature, and pragmatic deployment stage, there are a number of implications for donors and policymakers.

A reaffirmed need to address digital divide issues

High costs and illiteracy can prevent some groups from accessing and benefiting equitably from information and communication technologies. This includes across country, gender, ethnic, or socio-economic lines. These disparities are further amplified by AI, leading to a cumulative digital divide. Unfortunately, this inequity can be easily obscured due to the appearance of neutrality that surrounds AI and its statistical approaches.b973692827a0

Equally problematic is the fact that only an estimated 22% of AI professionals are women.9ec18a6ab81f Also, marginalised communities have low representation in the digital realm, or else they can have a disproportionate high visibility in relation to specific adverse events (eg, more prevalence in crime statistics due to more police attention).

These distortions mean that less high-quality training data is available for these particular groups. This can lead to AI systems that are likely to produce more erroneous, often biased results. For example, AI-enabled decision-support systems in hiring personnel might reproduce gender disparities when relying on legacy data that is skewed towards hiring and promoting male candidates.857ccfd02896 On the supply side, AI systems are highly concentrated in and controlled by a small number of commercial players in the Global North,50e7ead53e11 making it unlikely that the most advanced models speak directly to the needs of the Global South. Therefore, targeted donor support and mobilisation of investments and partnerships to bridge critical dimensions of the digital divide should rank high on the digital policy agenda. These initiatives can be rooted in the Global Digital Compact, which the global community committed itself to in mid-2024.d5c5cd81d6ee

Curate accountability data and communities

AI can serve as an integrative tool that can help extract insights from disparate and unharmonised information repositories at scale. In Armenia, for example AI is being deployed to more effectively scrutinise asset disclosures by officials.17fab1d87f38 In Czechia, AI helps identify long chains of political connections with finance and open ownership data.8260d156b4e6

With these functionalities, AI can help to connect disparate disclosure, transparency, and open data initiatives. The technology can take the combined value of these data pools to a new level for anti-corruption analysis and monitoring. Making this possible requires cross-practice networks, which is what donors can help build. This also requires continued investments in advocacy for transparency initiatives, and maintenance of high-quality, unbiased, open government and open data repositories whose utility will be significantly amplified by AI. Where concerns about breaching regulatory provisions might delay such investments, public donors could consider helping establish experimental applications, so-called ‘sandboxes’, where such interlinked applications can be tested under proper scrutiny.833695cc84da

Invest in targeted training data, open ownership models, and broader data collection

Limited availability of unbiased training data is one of the major constraints in fully harnessing the potential of AI, including its application in anti-corruption. For instance, in remote sensing, an abundance of monitoring data contrasts with a dearth of ground-validated training data.bcc986ebd47a Targeted support to build up specific, openly accessible training datasets in close consultation with the professional anti-corruption community might help to unlock this potential.

Build resources and infrastructures for challenging unfair AI outcomes

Given the scale of AI systems’ operations, it is inevitable that, even in optimal conditions, AI will produce a large absolute number of false negatives in anti-corruption screening efforts – that is, citizens erroneously accused of fraud, debanked, denied social benefits they are entitled to, and so on. Disadvantaged groups are more likely to be affected and less likely to muster the resources for contesting such decisions. Buffering the downside of AI in these areas will mean supporting affected individuals by building practical capabilities, helping them file effective complaints and, if needed, launching court cases. This can most productively be complemented by investing in analytical tools and assessment frameworks to help identify when and how AI fails to deliver, and produces biased and otherwise erroneous outcomes.

Build capacity for AI in the broader anti-corruption community

The latest conversational AI tools come with the appearance of easy accessibility and use, in a Google-search type of way. But this is treacherous. Supporting capacity-building for AI in anti-corruption is imperative for several reasons.

• In high-risk anti-corruption environments, a prerequisite of responsible AI use is to understand its many limitations, from biases to ‘hallucinations’.
• A major challenge is how to most effectively query conversational AI systems, as results are found to vary widely with the quality of the query. This has given rise to a dedicated expert community of so-called ‘prompt engineers’.
• The next step is using existing AI systems as a platform to customise applications for anti-corruption purposes. This requires technical AI expertise in public administrations and compliance departments. On the civil society side, we need good governance and robust ‘watchdog’ non-governmental organisations.

On the horizon are opportunities to train open-source AI systems with curated anti-corruption data to build more effective mission-centred AI systems. This will require advanced development skills and pragmatic, hype-resistant awareness of AI prospects at management level. The future also calls for new collaborative efforts as resources might have to be pooled and application benefits spread beyond individual organisations.

Donors can play important roles in supporting the required technical capacity, management awareness and collaborative spirit required for entering the next phase of AI against corruption. They can help to retain a pragmatic perspective that cuts through any hyperbole to diligently examine where and how AI can be most effectively integrated into integrity systems. Donors can also identify specific applications where the costs (still) outweigh the benefits, or where simpler technological solutions are available.