Unpacking the logic behind the magic in the use of AI for anticorruption screening (re Pastor Sanz, 2022)

‘Network of public contracts, contracting bodies, and awarded companies in Spain’ in 2020 and 2021; Pastor Sanz (2022: 7).

[Note: please don’t be put off by talk of complex algorithms. The main point is precisely that we need to look past them in this area of digital governance!].

I have read a new working paper on the use of ‘blackbox algorithms’ as anti-corruption screens for public procurement: I Pastor Sanz, ‘A New Approach to Detecting Irregular Behavior in the Network Structure of Public Contracts’. The paper aims to detect corrupt practices by exploiting network relationships among participants in public contracts. The paper implements complex algorithms to support graphical analysis to cluster public contracts with the aim of identifying those at risk of corruption. The approach in the paper would create ‘risk maps’ to eg prioritise the investigation of suspected corrupt awards. Such an approach could be seen to provide a magical* solution to the very complex issue of corruption monitoring in procurement (or more generally). In this post, I unpack what is behind that magic and critically assess whether it follows a sound logic on the workings of corruption (which it really doesn’t).

The paper is technically very complex and I have to admit to not entirely understanding the specific workings of the graphical analysis algorithms. I think most people with an interest in anti-corruption in procurement would also struggle to understand it and, even data scientists (and even the author of the paper) would be unable to fully understand the reasons why any given contract award is flagged as potentially corrupt by the model, or to provide an adequate explanation. In itself, this lack of explainability would be a major obstacle to the deployment of the solution ‘in the real world’ [for discussion, see A Sanchez-Graells, ‘Procurement Corruption and Artificial Intelligence: Between the Potential of Enabling Data Architectures and the Constraints of Due Process Requirements’]. However, perhaps more interestingly, the difficulty in understanding the model creates a significant additional governance risk in itself: intellectual debt.

Intellectual debt as a fast-growing governance risk

Indeed, this type of very complex algorithmic approach creates a significant risk of intellectual debt. As clearly put by Zittrain,

‘Machine learning at its best gives us answers as succinct and impenetrable as those of a Magic 8-Ball – except they appear to be consistently right. When we accept those answers without independently trying to ascertain the theories that might animate them, we accrue intellectual debt’ (J Zittrain, ‘Intellectual Debt. With Great Power Comes Great Ignorance’, 178).

The point here is that, before relying on AI, we need to understand its workings and, more importantly, the underlying theories. In the case of AI for anti-corruption purposes, we should pay particular attention to the way corruption is conceptualised and embedded in the model.

Feeding the machine a corruption logic

In the paper, the model is developed and trained to translate ‘all the public contracts awarded in Spain in the years 2020 and 2021 into a bi-dimensional map with five different groups. These groups summarize the position of a contract in the network and their interactions with their awarded companies and public contracting bodies’ (at 14). Then, the crucial point from the perspective of a corruption logic comes in:

‘To determine the different profiles of the created groups in terms of corruption risk, news about bad practices or corruption scandals in public procurements in the same period (years 2020 and 2021) has been used as a reference. The news collection process has been manual and the 10 most important general information newspapers in Spain in terms of readership have been analyzed. Collected news about irregularities in public procurements identifies suspicions or ongoing investigations about one public contracting body and an awarded company. In these cases, all the contracts granted by the Public Administration to this company have been identified in the sample and flagged as “doubtful” contracts. The rest of the contracts, which means contracts without apparent irregularities or not uncovered yet, have been flagged as “normal” contracts. A total of 765 contracts are categorized as “doubtful”, representing 0.36% of total contracts … contracts belong to only 25 different companies, where only one company collects 508 granted contracts classified as “doubtful”’ (at 14-15, references omitted and emphasis added).

A sound logic?

This reflects a rather cavalier attitude to the absence of reliable corruption data and to difficulties in labelling datasets for that purpose [for discussion, again, see A Sanchez-Graells, ‘Procurement Corruption and Artificial Intelligence: Between the Potential of Enabling Data Architectures and the Constraints of Due Process Requirements’].

Beyond the data issue, this approach also reflects a questionable understanding of the mechanics of corruption. Even without getting into much detail, or trying to be exhaustive, it seems that this is a rather peculiar approach, perhaps rooted in a rather simplistic intuition of how tenderer-led corruption (such as bribery) could work. It seems to me to have some rather obvious shortcomings.

First, it treats companies as either entirely corrupt or not at all corrupt, whereas it seems plausible that corrupt companies will not necessarily always engage in corruption for every contract. Second, it treats the public buyer as a passive agent that ‘suffers’ the corruption and never seeks, or facilitates it. There does not seem to be any consideration to the idea that a public buyer that has been embroiled in a scandal with a given tenderer may also be suspicious of corruption more generally, and worth looking into. Third, in both cases, it treats institutions as monolithic. This is particularly problematic when it comes to treating the ‘public administration’ as a single entity, specially in an institutional context of multi-level territorial governance such as the Spanish one—with eg potentially very different propensities to corruption in different regions and in relation to different (local/not) players. Fourth, the approach is also monolithic in failing to incorporate the fact that there can be corrupt individuals within organisations and that the participation of different decision-makers in different procedures can be relevant. This can be particularly important in big, diversified companies, where a corrupt branch may have no influence on the behaviour of other branches (or even keep its corruption secret from other branches for rather obvious reasons).

If AI had been used to establish this approach to the identification of potentially corrupt procurement awards, the discussion would need to go on to scrutinise how a model was constructed to generate this hypothesis or insight (or the related dataset). However, in the paper, this approach to ‘conceptualising’ or ‘labelling corruption’ is not supported by machine learning at all, but rather depends on the manual analysis and categorisation of news pieces that are unavoidably unreliable in terms of establishing the existence of corruption, as eg the generation of the ‘scandals’ and the related news reporting is itself affected by a myriad of issues. At best, the approach would be suitable to identify the types of contracts or procurement agents most likely to attract corruption allegations and to have those reported in the media. And perhaps not even that. Of course, the labelling of ‘normal’ for contracts not having attracted such media attention is also problematic.

Final thoughts

All of this shows that we need to scrutinise ‘new approaches’ to the algorithmic detection of corruption (or any other function in procurement governance and more generally) rather carefully. This not only relates to the algorithms and the related assumptions of how socio-technical processes work, but also about the broader institutional and information setting in which they are developed (for related discussion, see here). Of course, this is in part a call for more collaboration between ‘technologists’ (such as data scientist or machine learning engineers) and domain experts. But it is also a call for all scholars and policy-makers to engage in critical assessment of logic or assumptions that can be buried in technical analysis or explanations and, as such, difficult to access. Only robust scrutiny of these issues can avoid incurring massive intellectual debt and, perhaps what could be worse, pinning our hopes of improved digital procurement governance on faulty tools.

_____________

* The reference to magic in the title and the introduction relates to Zittrain’s Magic-8 ball metaphor, but also his reference to the earlier observation by Arthur C. Clarke that any sufficiently advanced technology is indistinguishable from magic.

Interesting paper on use of theory in public procurement research (Flynn & Davis, 2014)

I have just read A Flynn and P Davis, "Theory in public procurement research" (2014) 14 (21) Journal of Public Procurement 139-180, which had been on the "to read" pile for a bit too long. Their paper focuses on procurement/supply-chain management scholarship, which will make it seem "peripheral" for legal academics. However, I think that their findings on the need to beef up the use of theory in public procurement research will translate well to legal scholarship, particularly that of an interdisciplinary nature. Well worth reading.