"AI is incorruptible"

The Guardian published an article yesterday claiming that Albania created an AI 'minister' that would be responsible for public procurement in the country. The logic behind this idea is to make Albania "a country where public tenders are 100% free of corruption" in the words of Edi Rama, Albania's prime-minister. The objective being to remove the decision-making power for awarding contracts from ministers (and officials) and instead trusting that to a piece of software. I am at a loss for words.

Diella, the new minister, seems to be an evolution of an AI system introduced in January to help users when using the e-Albania with voice inputs, working as a virtual assistant of sorts. Unfortunately the reports in English are not really technical enough to disclose exactly how this AI system is operating. Is it a LLM? Is it simply a decision-tree translated into code? Are these smart contracts masquerading as AI systems?

Beyond the obvious jokes about attendance to cabinet meetings and if the Albanian government will undertake a reshuffle every time a future update to Diella is made available, there something important to discuss here.

Albania may be the first but surely won't be the last country falling to the siren call that humans when handling procurement are fallible and as such the decision-making should be passed on to a machine instead. Beyond the usual pitfalls that one can point here, for example LLMs are probabilistic systems while procurement decisions are deterministic, it is the fundamental idea that is problematic.

An IT system - 'AI' or otherwise - will be as good, secure and aligned as programmed with a leeway for bugs and omissions. It is perfectly conceivable that the programming can be weaponised and hijacked to achieve certain purposes, be them corruption or mere influence peddling. When using an 'AI system' to make decisions in procurement the corrupt intent becomes a lot harder to find out since, well, it was the machine that decided, not a corrupt human. This is more so for AI systems operating like black boxes. And even without any foul play, software always has bugs so any aggrieved bidder would need to attack a decision not on its merits but on the software that led to such decision. Even assuming all relevant code is made available for inspection, good luck in explaining that to a judge.

So, yes AI is definitely corruptible be it by being harnessed for corrupt purposes or bugs in the code/training 'corrupting' the decision-making itself.