Initially, Diella (meaning “sun/ray of sunlight” in Albanian) was launched in the form of a virtual assistant on the electronic platform e-Albania. It was an advanced chat bot replying to the questions of citizens on public services in simple language with an accompanying video – Diella appeared in front of a user as an avatar of a woman in a traditional Albanian attire of Zadrima region. It is stressed that it was developed in an incubator of the AKSHI artificial intelligence based on the Azure OpenAI innovative technologies and the latest models of artificial intelligence by Microsoft.
Edi Rama, re-elected Albanian Prime Minister, who had said he had considered AI as a potentially effective anti-corruption instrument, now appoints Diella a member of a new Cabinet of Ministers. The neural network will oversee and define the winners in all public tenders, where the government concludes contracts with private companies: the process of delegating functions from the public bodies to the new digital minister is planned to be implemented step-by-step. Mr. Rama said he was confident that Diella would make it possible to ensure that public procurement in Albania is 100% free from corruption.
At the same time, some media highlight that the government has not provided any details with regard to any human oversight over Diella’s activities and has not addressed the risks of manipulation of AI.
At the same time, these risks give rise to serious concerns about the actual effectiveness of the neural network in combating corruption.
In the first place, the quality of data on which it learns is critical for the neural network. The lack of information about who and how collects this data and understanding of their content raises legitimate concerns about probable use of the data “favourable” to certain persons, hidden editing of directories, engineering attacks etc. throughout Diella’s development or its additional learning. This, in its turn, creates new corruption hotbeds rather than mitigates corruption risks.
Additionally, the “rules” used in learning can affect the outcome of the operation of the model: if they — accidentally or deliberately — are formulated incorrectly, the neural network can start to systematically churn out incorrect conclusions and/or “hallucinate” (see, for example, a recent study by OpenAI, where it is concluded that the reason for frequent “hallucinations” of large language models is due to the assessment criteria applied to the responses throughout learning on which it is based), which is almost impossible to detect in a timely manner in the absence of external control mechanisms.
In the second place, as we know, the algorithm of neural networks’ learning (or, more precisely, how exactly the weights are distributed within a model) remains a “black box” for AI-engineers. For such a high-risk area as the determination of the winner of a tender, lack of transparency and clarity of the process is inadmissible: the participants of procurement are entitled to take a motivated decision, while the “black box” complicates the appeal and audit. Moreover, the Artificial Intelligence Act adopted recently in the EU, whose member Albania aspires to become, stipulates that there must be documented risk management, transparency and mandatory human control in the high-risk areas.
In the third place, we can assume that in order to detect corruption, the model learned on standard indicators of undue practices in procurement. However, such “evident indicators” do not always reflect real corruption. Models normally “catch” simple patterns (for example, in case of procurement – short deadlines, sole-source procurement and the like), but do not properly “see” latent linkages – beneficial entanglements, cartels, fraud competition schemes – everything that requires cross-cutting network analysis, examination of different databases, public registries, analysis of procurement “in the dynamics”. As a result, if Diella learned on standard corruption offences in procurement, it will overlook less evident corruption practices.