During IGF 2021, hosted by Poland, we've had several interesting discussions. One of them, organized by the Polish Office of Competition and Consumer Protection (UOKiK) sparked our particular interest. Its topic was "Digitalization through the use of AI in public administration". What was discussed was an AI-based system to support UOKiK's work at detecting abusive clauses in legal agreements.
Tomasz Chróstny, UOKiK's president, stressed what a great challenge consumer protection is, given the speed of technological advancement.
During the work on the aforementioned abusive clause system a few important conclusions emerged.
First of all, it was noted that decreased efficiency of public administration results from focus on legal matters instead of technological ones, due to the nature of the work at public institutions.
It was emphasised that an important facet of AI system development and implementation is understanding their inner logic and building a knowledge base of AI law infringement remediation. Cybersecurity knowledge and analytical skills are essential here, too.
The need for regulations on software vendor selection was noted; the key criterion should be quality and clear vision of the outcome. What's more, previously conducted pilot AI programs should be used to create a strategy of building AI systems based on best practices.
Our Foundation thinks it necessary to work out good standards of AI system use by the public sector. We'll soon publish a report, prepared jointly with public procurement law experts, entitled "Artificial Intelligence and Automatic Decision Making in Public Procurement - Guidelines for the Public Sector".
The discussion surfaced the point that AI development should involve the end user of the system, who will later train it. Understanding why a given system is created and how a given organization works, as well as learning to read and use public registry data is key to success. Because of this, an essential yet challenging aspect of AI implementation is finding people with the right knowledge and experience.
The panelists pointed out to the moment of when the ready system starts its operation at the target institution as one where many unpredicted situations occur. Such situations may happen despite a well prepared plan and well executed prior work. That's why constant monitoring of AI systems and their outcomes is always necessary.