Only 9 Percent of Polish Public Offices Admit Using AI

The Sieć Obywatelska Watchdog Polska (Citizens Network Watchdog Poland) checked how Poland's public administration uses artificial intelligence, and it turned out that most offices either don't know they're using it or would rather not say. Of 5,480 institutions surveyed, only 9 percent admitted to using AI, even though many of them went on to describe tools that qualify as artificial intelligence under the EU's AI Act definition.

What the study found

In the summer of 2025, Watchdog Polska sent public information requests to 5,480 institutions carrying out public tasks in Poland, asking directly whether they use artificial intelligence systems, what they use them for, and whether they have any internal rules governing their use by employees. A total of 4,042 institutions responded, or 74 percent of those asked, a result the organization itself considers good compared with other public information access monitoring efforts.

Of those responses, only 9 percent of institutions openly admitted to using AI-based tools. The rest either denied it or failed to answer the question unambiguously. The problem is that some of the same offices that denied using AI went on, later in their responses, to describe using speech transcription, chatbots serving residents, image analysis, predictive models, medical diagnostic systems, automatic meeting summarization, or tools supporting administrative decision-making.

A gap between types of offices

The scale of declared AI use varies sharply by type of institution. In cities with county rights, the largest urban centers with more developed administrations, 34 percent of institutions admitted to using AI. In county administrations (starostwa) the share dropped to 14 percent, and in ordinary rural and urban-rural municipalities to 8 percent on average. Central government bodies declared AI use in 20 percent of cases. The gap shows that access to AI tools and the competence needed to implement them are concentrated in large cities, while smaller local governments lag behind.

No rules in place

The most worrying element of the report isn't AI use itself but the lack of any oversight of the process. Of the more than two thousand municipalities that responded to Watchdog Polska's request, only 20 reported having written rules for employee AI use. Translated to ordinary municipalities, that means fewer than 1 percent of institutions have any regulation, training, or data security policy tied to the use of these tools.

That means that in the vast majority of offices, employees can use chatbots or generative text models without any guidelines on protecting residents' personal data, verifying results, or accountability for errors made by the algorithm.

A dispute over the definition of AI

Michał Zemełka, the author of the study, points out that part of the gap between declared and actual AI use stems from officials' narrow understanding of the term, who equate artificial intelligence almost exclusively with generative text tools like ChatGPT.

Yet under the EU's AI Act, AI systems can also include predictive models or tools that support administrative decisions. One example is the model used by ZUS (Poland's Social Insurance Institution) to analyze sick leave, which supports decisions on inspections but isn't treated as AI. This makes it harder to assess risks and exercise public oversight - Michał Zemełka, report author, Sieć Obywatelska Watchdog Polska

This definitional ambiguity has concrete legal consequences. If an office doesn't classify a given tool as an AI system, it doesn't apply the obligations stemming from the EU's AI regulation, such as risk assessment, documenting how the system works, or informing citizens that their case was processed with the help of an algorithm.

What Watchdog Polska proposes

The organization is calling for a public registry of AI applications used by public administration, a requirement to inform citizens when AI is applied to their specific cases, clear criteria for classifying systems as AI, and internal data security policies in every office. The registry proposal has already been brought before the Sejm (the lower house of Poland's parliament) and Senate digitization committees, where Zemełka argued for writing the registry requirement directly into the regulations implementing the EU's AI Act in Poland.

For an average resident, the report means there is today practically no way to check whether a decision on their case, such as a benefit application or a permit, was supported by an algorithm, or whether the data they submitted to the office ended up in an external AI system. For the offices themselves, the report is a warning sign ahead of the full AI Act requirements taking effect, which impose specific documentation and disclosure obligations on public institutions for high-risk systems.

The scale of the problem, with fewer than 1 percent of municipalities having any rules at all, suggests that implementing these requirements in practice could take local governments far longer than the EU deadlines assume, unless central support arrives in the form of ready-made regulation templates or state-funded training.