Welcome to an illuminating discussion shedding light on the critical subject of anticorruption and its symbiotic relationship with the realm of artificial intelligence and machine learning. In this captivating interview, we have the privilege of engaging with Prof. Stéphane Gagnon, an eminent figure whose studies lie in the convergence of these two powerful forces. Prof. Gagnon’s profound insights and fascinating works have not only unraveled the intricate challenges posed by corruption but also harnessed the potential of AI/ML to devise innovative solutions.
Join us as we delve into the mind of Prof. Gagnon, exploring their visionary perspectives and transformative strategies that promise to reshape the landscape of anti-corruption efforts. Get ready to embark on a thought-provoking journey where we explore the intersection of AI/ML and anti-corruption, unraveling new possibilities to build a more transparent and just society.
Darla Tverdohleb: How can AI/ML technologies be leveraged in the field of anti-corruption intelligence to enhance the identification and detection of corrupt practices? Could you provide examples of specific AI/ML techniques that can be utilized.
Prof. Stéphane Gagnon: First, thanks for this opportunity to share my thoughts on these big questions.
A warning though please: I’m not an expert in anticorruption. I’m just a professor in IT management. However, the topic, I’ve started to look at it several years ago because of a student who was doing his doctorate with me, Dr. Yanik Harnois. And I had to provide him, you know, a topic that he liked. And of course, I had to align myself with his interests. So that’s how it started in 2016, seven years ago, and he now graduated last year.
Now, as you saw in the papers that I co-authored with him, he took parts of his thesis, and published them in international journals. One part was the theory, which is very basic literature review, as the journal asked us to cut the paper by half, so we had to provide only a short list of references. Then the other part was the empirical part, which was to gather data from the field. He had 30 interviews with very senior professionals in the area, so that the empirical part was to analyze their interview records. And the third part was a recent theory paper about proposing a change of how we address anti-corruption efforts. Instead of looking at them from mechanical perspective, look at them from a more organic perspective. They must be grown from field observations, and understood as a socio-cultural phenomenon, which is in line with leading authors who suggested that approach in the past 5 years. It’s a complex theory, but I will come back I can elaborate later if you want. So, that explains why I’ve been interested in the topic. It’s not my expertise at all, but I’m fascinated how managers can navigate and survive some hostile corrupt environment, that is, without being corrupt themselves.
D.T.: What are some challenges and considerations when it comes to using AI/ML algorithms for the analysis and interpretation of corruption-related data and evidence? How can these challenges be addressed to ensure accurate and reliable results?
Prof. S.G.: So, the next challenge is, is there a relevance of AI ML to that topic? Absolutely. 100%.
Obviously, AI ML is broad, very broad. For example, I, I’ve worked a bit on various ML apps, for example, well, I’ve worked off for both, you know, desktop-based and big cluster-based approaches, both ways. I’ve worked also on Spark, you may remember, for example, Databricks, which is the commercial version of Spark, which is the open-source version. And I’ve worked also, of course, in the R language as well. So, I’m familiar with a bit of different platforms.
I’m also familiar with different algorithms, the most interesting I’ve worked on are the old algorithms, for example, Bayesian networks, and dynamic Bayesian networks, classification problems, decision trees, random forest, Support Vector Machine (SVM), and many more. So those are very old algorithms. And I worked on them for my first thesis with another student in 2015, 2016, 2017. So since then, I’ve just worked a little bit on various other of course, Bayesian networks, again, neural networks, SVM, and all that. So I’ve worked with different algorithms, but not with the deep learning yet. I’m still studying myself and preparing myself to jump into that. It’s not an area that I find personally so fascinating, yet, as the problems I work with don’t have large training datasets. But with big data, like in cybersecurity, we have no choice, deep learning is the best way to go.
There’s a lot of potential, of course, but old algorithms are the kind I work mostly about. And of course, in the case of anti-corruption, you can ask yourself, what is the data? Because that is the key, isn’t it? AI/ML depends on the data. While the data is practically hidden in the field, we do not have access to it, or not as easily as other datasets. And that is one of the major problems with anti-corruption, which leads to answer number two: there’s no data, and there’s no evidence. So you see, AI/ML is limited in value, because as you recall, for example, big, very heavy data sets are required for deep learning. That’s something I’m sure you’re familiar with as a data analytics student.
So, the training data does not exist in anti-corruption. There’s not a great abundance that would be very well documented cases of corruption and bribery and fraud by the government agencies or professionals or even elected officials. So, because of that, it remains a huge discrepancy to use AI effectively. So practically, the only thing you can do is hint to corruption events.
The usual case of forensic accounting, I’m not sure if you’re familiar with the concept of forensic accounting. It’s accounting in cases of litigations. So, it’s like forensic, as you recall, it is if there’s a murder, forensic physicians go and do an autopsy, but also analyze the crime scene. The same thing for accountants, in the case of financial fraud and financial crime, they are forensic accountants, and they are there to analyze data, of course.
Now, what can they do? Of course, they can corroborate. For example, if an expense is bigger than what was delivered, there was a discrepancy. Ah, the obvious thing, you know, discrepancies in expenses and declared actual deliveries, accurate measurements of things, for example, oh, there were three loads of trucks. But they claimed that they paid for five, as usual. So, discrepancies, it’s very basic, it’s not very complex. However, the chain of responsibility is the more challenging one. That’s where social network analytics (SNA) comes into play. So basically, you have to discover if we have relationship between the company that delivered the trucks, truckloads of, let’s say, asphalt, if you want pavements of roads, and the company that ordered it, or the ministry. However, we discovered through social networks that the cousin of the minister of someone is actually one of the owners of the company that actually lends the tires to the truck company, who is actually the owner.
So it’s very simplistic things, mundane data, that once collated together can raise a red flag. But again, you can hint as to the cause and probable cause, yet, you still don’t have enough evidence to bring that to court. So that is the problem.
AI/ML is limited by data. If you had more data, we can more seamlessly, of course, do training and of course, find inference mechanisms. So, I think it’s still very limited today. There’s a very interesting conference that happened in October 2021, in Washington, DC, entitled “Symposium on Data Analytics and Anticorruption”, I did not attend but watched a lot of recordings. It was by the World Bank who organized it, it was a very nice event of four days and they’ve recorded everything on YouTube. So you can watch, and you will be fascinated by the kind of experience they reported about. You’ll be saying to yourself, wow, it’s impressive the kinds of solutions that are out there. But again, they’re limited by access to data. Very much.
So, you see, there is what we call open access data of that kind and has been very heavily analyzed actually a lot with SNA. But again, it’s extremely limited compared to what we need to bring a company or people to the court. You need, of course, damages that are clearly identified; you need, of course, motive; you need to justify motive; you need to justify to prove that the person or the organization benefited from it; you need purchase transaction; you need all sorts of factors, for example, witnesses that can corroborate the documentation, because a document on its own is not enough. You see all sorts of factors are at play to actually get corrupt people to pay for their crime.
So, because of the lack of data, we’re stuck. So, AI/ML has limited values too, for now, but that could change, and law enforcement agencies and other organizations need to work together for digital innovation.
So that’s what I wanted to make sure we emphasize in your first two questions.
D.T.: Can you explain the role of predictive analytics in anti-corruption efforts? How can AI/ML models help in identifying potential corruption risks and proactively preventing corrupt practices?
Prof. S.G.: Question three is about, technically, predictive analytics. As I said earlier, indeed, the analytics is mostly descriptive for now, and predictive analytics does not have yet much reliability. Actually, you can describe a network of people, but it doesn’t tell you anything particular, only the fact they are “possibly connected”. So, predictive analytics is about, of course, for example, in SNA, you may find that a person has a number of many, many, many, many connections, you know, nodes and those edges between nodes. You try to figure out which one is the most central node that connects most of them. So, the network centrality is a measurement that is very crucial. Another one, how many of the nodes formed together community or sub communities. So, it’s kind of identifying which nodes are belonging to a group or subgroup. Third, if you have a node that is missing, and you know, someone must be in between those connections. You have to make an inference. Which node is possibly in between that we do not have an identity for, but it would explain numerous connections between the nodes.
So, you try to identify potential nodes missing all sorts of analysis that can allow you to predict the possible, I would say the possible source of the problem in corruption events. So again, predictive analytics, in that case, has more value in terms of finding the best place to find evidence. Again, it’s not about proving. It’s not about giving you strong value to go to court. It’s about the investigation process. How you carry out your investigation, you see. So that’s an example in terms of the value of AI/ML, linked to SNA, but you can do the same reasoning with all other cases where data is sparse and incomplete, and there is a need to find missing links.
D.T.: Discuss the ethical considerations involved in the development and deployment of AI/ML solutions for anti-corruption purposes. How can bias and fairness issues be addressed to ensure equitable outcomes and avoid unintended consequences?
Prof. S.G.: We come to the question of solutions, deployment, that’s where I would say the ethical aspect is most central.
Your question number four, I’ll be honest with you, there’s the ethical aspect of can you use, for example, private information to monitor possible fraud? Well, organizations have a great degree of latitude in monitoring their employees and executives. So, in fact organizations in general, the law of access to information, grants organizations a great deal of freedom. So technically, there’s not much limits what the organization can do for it to monitor fraud. But as soon as you cross organizational boundaries, or as you go to, for example, if you have an agency that has authority over a certain industry, for example, the telecom sector or the banking sector, you have secrecy laws that are more specific and stringent. And therefore, the agency does not even have the power to access data, unless they can subpoena, or have a formal reporting process like in banking with FINTRAC.
So therefore, in Canada, in our case, it’s obvious that the ethical considerations are not very complex, they are explicit, at least, they’re not hidden. And also, there are basic guidelines to follow. So and also the legal framework in Canada is strong. So because it’s strong, and because companies have a tendency to follow it, and ethical concerns or risks are not so big.
But if you go to another country that has less strict rules and laws in managing information, and implementing software, that’s where the concerns arise. They arise because companies have too much freedom or they are “unruly.” They do not follow laws as they should, because there is no one watching. That is where the concerns are. But in the context of an economy like Canada’s, no, not really. So, I don’t see ethical concerns much in fact, and also regarding algorithms, since companies in Canada tend to use commercial software, and automatically vendors take steps to ensure algorithms are compliant. So Canada has lots of potential for safe AI/ML deployment.
It’s interesting because algorithms have been criticized in many ways, of course, for example, the fact that nobody can explain the outcome of most deep neural network. Because it’s very hard after 1000 layers to find out which of the layers actually had an impact on the final result. It’s almost like a butterfly effect. At some point, you may recall the butterfly effect. It’s a very wide Theory of Complexity. So more importantly, though, explainability as we call it, is reachable in many instances by smaller models. So, decision trees, for example, are highly favored in the public sector.
The more important thing, they use decision trees because they prefer to have absolute assurance, when there’s a legal problem, we can explain how the decision was performed. Because it’s used as exactly precise, very crisp, very small margin of error as well. That’s an example of how this ethical issue or concern is resolved.
There’s a whole range of things to say, but, uh, I would say confirming that because AI/ML is very little applied yet in anti-corruption litigation, the likelihood that it’s going to be a major concern ethically, is not yet emergent. I mean, it’s very dormant because it’s not yet very heavily applied. Again, I’m not in touch with the police or justice systems on this, but I can guess from the kind of studies that are published that it’s not yet an issue.
D.T.: What are some legal and regulatory implications of using AI/ML in corruption investigations and prosecutions? How can the authenticity and admissibility of AI-generated evidence be established in legal proceedings?
Prof. S.G.: Question number five was about regulatory and legal? Well, clearly, as I said, it depends on each industry. For example, the telecom sector, again, the wireless sector, they have a different set of constraints. They can do a lot of control. Indeed, they can control where you go, they can control what you send, they can control who you connect with, they can control even literally all sorts of little details, such as your device contents, and whatnot. But because of regulations, of course, their limitations are that it’s only poured into their servers, but they cannot do anything with it. So, I think legal and regulatory implications, well, technically, the companies in each industry each have their existing sets of conditions that they follow. I don’t think there is a such great uncertainty as to what’s going to be done. Companies will follow the rules in Canada because that’s what they do, and they remain focused on protecting their competitive advantage, not on messing around.
Most companies know, for example, in the retail sector, that in the retail business, you care about repeat business, you care about loyalty, customer loyalty, you care about what do what do customers want, and offer them something as a good price? So how do you keep them coming back? It’s a basic name of the game, isn’t it? So companies are going to be focused on aligning AI/ML with their strategic and competitive advantage, not on doing bad deeds, so Canadian companies are not a risk of ethical misconduct, I feel.
In the end of it, for example, in some industries, such as construction, there are other problems, and it relates most importantly on the structure of transactions and contracts. Because suppliers have control over quantities reported, and nobody is checking, that’s how small transactions can be hidden, and payments extracted.
As I said earlier, I can pay for five trucks of asphalt, you deliver three. Bingo, you and I made a deal. So, you see the point. The point is, each industry has its own characteristics of what corruption is, you see what I mean? And for that reason, it depends very heavily on, again, is the country prone to corruption? Or are these companies that we call unruly? Do they have a tendency to be unruly and not follow the rules? I think it can in many contexts, but it is very unlikely depending on the culture. It’s not a very highly generalized situation in Canada.
So I’m just confirming to you that, again, if you look at each industry, you will find that conditions exist that AI/ML would be properly regulated. Technically, it wouldn’t be a big challenge. And companies are familiar with the kind of data they have. Companies are familiar with their own industry regulations, so they’re capable to handle the implementation. Again, if vendor take the right steps to supply reliable and ethically well-designed AI/ML functionality, then Canadian companies will be safe and ethical end-users too.
D.T.: How can AI/ML technologies be utilized to facilitate the collaboration and information-sharing between anti-corruption agencies and stakeholders? What are some potential benefits and challenges in implementing such systems?
Prof. S.G.: Now the last quetion is about collaboration among agencies. But I think that’s where some of the issues will and definitely are arising. For example, there is in Canada, for example, no centralized agency for that kind of work.
For example, it took years and years to obtain that Quebec do the Charbonneau Commission. In Quebec, the province, there was a big, big parliamentary commission to investigate fraud in the construction industry. The process evolved technically more like five years, but it took a lot of time to investigate and charge people with financial crime. So, we have to remember all of these tasks, take time and also take personnel and then take money to spend on that task. But we cannot generalize from one industry in one province to the whole country. Practically, very few people falsify evidence in Canada. I don’t think it’s a frequent case. The reason? Because there’s not much money at stake in most day-to-day transactions, we don’t all have multi-million dollar deals going around. You know, if there’s a lot of 100 millions dollars at stake, well, I’m pretty sure some criminal will falsify evidence, but we’re talking about small transactions 100,000, 10,000. Who is going to go and risk his own personal life to falsify evidence just for a $10,000 transaction corruption? So you see the incentive is very small and therefore there is no incentive to do again “criminal acts.” It depends on the value of the reward for corruption. So, in some economies there are risky deals like this, and in some other economies there are not. So, I don’t consider that issue a Canadian issue yet. I don’t.
As we draw the curtain on this captivating interview with Prof. Gagnon, we are left inspired by the groundbreaking possibilities that arise at the intersection of AI/ML and anticorruption endeavors. Prof. Gagnon’s words have illuminated the path to a future where corruption’s grip is weakened by the transformative potential of technology.
The integration of AI/ML in anticorruption efforts opens new doors, offering innovative tools for detection, prevention, and enforcement. By leveraging advanced algorithms, data analytics, and automation, we can accelerate the identification of corrupt practices, enhance risk assessment, and empower anticorruption initiatives with unprecedented efficiency.
However, it is our responsibility to ensure that the deployment of AI/ML in anticorruption measures aligns with principles of fairness, transparency, and human rights. We must guard against potential biases, promote explainability, and accountability, and constantly adapt our strategies to the evolving landscape of technology and corruption.
Thank you for joining us on this enlightening journey, and may the insights gained from Prof. Gagnon inspire us all to be agents of change in the ongoing fight against corruption.
Be the first to comment