Guest Post: Filling the Gap – an AI Tool for Analyzing the Integrity of Public Officials

GAB is pleased to publish this guest post examining why AI is underused in integrity analysis and outlining a future AI tool to fill this gap. The authors are Tilman Hoppe, an international expert on asset declarations and integrity vetting, and Hassan Ugail, a professor of visual computing at the University of Bradford, UK, who develops AI systems.

Law enforcement agencies are using AI for a growing number of tasks, from facial recognition, cargo screening, border monitoring, and even report writing (here). Surprisingly, for analyzing the integrity of public officials, AI remains in the shadow.

Why is AI needed?

Those who have used one of the popular chatbots such as Claude or ChatGPT know that AI can analyze a vast amount of data and identify even the most remote relation between sets of data. It can calculate finances, understand regulations, and distinguish people. AI thus seems the perfect solution for integrity bodies.

Most integrity bodies have only a few staff who are overwhelmed with a high number of public officials (and their family members) they must monitor, involving complex financial and non-financial data. For example, in Ukraine, around 700,000 declarations are submitted per year, and in Morocco 35,000. If one staff member has 220 workdays per year, and needs on average 3 days for a full audit including all steps, 30 staff members – as available in Ukraine – can audit a maximum of 2,100 declarations per year. Therefore, usually, integrity bodies manage to review only a small sample of public officials each year – and often, public officials are never checked at any time during their time in office. And even if there is only so much one person can do. For example, the standard evaluation file for judicial vetting in Moldova “contains 3,612 pages of documentation and data”, will a staff member working for days or weeks on these pages find out that one bank transaction for an electricity bill on page 1,276 may contradict the judge not declaring any real estate?

There are more advanced automated systems for verifying asset declarations, such as the ones used in Armenia or Ukraine. However, these systems only have a set of external databases they compare data with (Armenia), or at most apply a list of fixed rules (Ukraine), such as: does the declared data match the data in other databases? Or does the size of declared real estate exceed a defined threshold?

They are no more sophisticated than an Excel sheet into which one needs to enter the right data into the right field (for example, the size of an apartment in sqm). These automated systems cannot do what chatbots such as Claude or ChatGPT can do: learning from each case, being trained, concluding based on information from circumstances, detecting highly complex relations beyond fixed rules, analyzing uploaded datasets, such as bank statements, or providing directions on what to check and why.

What are the concerns?

If AI is as smart as described here, shouldn’t it be used already across the world? We have heard three main concerns – and often misunderstandings – when presenting the idea of AI at international fora, such as that IACA convened in September or on the margins of the Western Balkan Summit. They are:

  • Data protection: Analyzing the integrity of public officials involves personal data, often sensitive and protected by data privacy laws. Claude, ChatGPT, Gemini and other free online AI models use the data that users upload to them to train future versions of their models, and this disclosure to a third party has led most law enforcement agencies to bar agents from uploading sensitive data of any kind to them.
  • Ownership: How can we define rules AI must follow to check the data? If we upload the data to Claude or ChatGPT, we can provide prompts. Still, we will have no idea how the AI arrives at a certain conclusion; i.e., the algorithms are not up for real scrutiny.
  • Accountability: Who is accountable for a wrong decision on a public official if AI provides wrong conclusions in analyzing financial data of a public official?

How could a new AI tool address all these concerns?

Data protection. The solution to data protection concerns is the use of what is variously termed an off-line or open-source model. These are like Excel or Word: they run on an individual’s own computer, and any information inputted is held on that machine alone — as are any computations it makes using that data. One (important!) difference between an open-source AI model and Excel and Word is that open-source models are free. This post by Hugging Face (yes, that Hugging Face) rates open-source models.

A drawback with models that run on individual desktops or laptops is that they are less powerful than the free ones available online. Agencies that need more computing power can still keep all data and computations in-house by purchasing a more powerful computer to run powerful off-line models. These cost a few thousand dollars at most.

The use of an off-line or open-source AI model would presumably solve the data protection problems. The data would not go anywhere but remain within the agency’s control. The automated processing of data within the agency might perhaps require an additional legal basis depending on national legislation. However, such a legal basis already exists, at least for countries using already automated systems (below the complexity of AI).

Ownership and accountability. Concerns about ownership and accountability of the use of AI miss an essential point about AI. An AI tool does not decide anything. It “only” helps sort through the data (which, in fact, is 99% of the ultimate success if a public official has several bank accounts and one goes back several years in analysis). The outcome would not (just) be: “The official has suspicious income.” If not followed by reasoning, this is a useless statement – whether from a person or from AI.

Instead, the tool will point out the data it relies on (as Claude or ChatGPT does if properly prompted). Always, an investigator or analyst would need to reflect on the evidence presented and decide if any further action should be taken: Closing the case or asking the official for a statement? Requesting or reviewing further data that is not machine readable? Consulting a lawyer or banking specialist? Thus, in terms of ownership and accountability, there is no difference between the systems already used for automated verification or in purely manual mode.

AI offers ethics and anticorruption agencies a powerful new tool for policing public officials’ conduct. At a price within their budgets. Time to take advantage of it.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.