Fighting Fire with Fire: A Technological Solution Proposal to the Data Issue of Bid-Rigging Detection through Artificial Intelligence

Artificial Intelligence (hereafter, ‘AI’) systems are widely adopted by public administrations. Public procurement does not escape the rule. This is unsurprising, given AI systems promise to increase the probability of bid-rigging detection by screening structural and behavioural indicators. This paper does not discard the benefits of AI-driven screening, but argues it is not a silver bullet. This algorithmic solution faces, amongst other issues, a twofold data challenge.
First, as any data-dependent system, AI-driven screening is impacted by problems in the availability of the data it relies on. “No data, no prediction” is the prevailing idiom. Assuming there is no (or not enough) example(s) of collusive and non-collusive behaviours to train the algorithm, this paper explores the possibility of using a training dataset from a market that is comparable to the targeted market and assesses the condition under which the comparability will produce reliable predictions.
Second, the data gathered must be quality data. If not, then the AI system might be prone to type I and type II errors. The idiom hence become, “without quality data, bad prediction.” Again, this paper argues this does not constitute a dead-end. An embryonic solution is given by the European Commission’s Proposal for an AI Act. Albeit its applicability to public procurement is far from certain, this paper draws inspiration from the data quality requirements set out in Article 10. This provision constitutes a good but steep starting point. Some improvements are needed (e.g. to define what is 'appropriate' data governance). Against that background, this paper proposes a concrete solution based on semi-supervised learning.