Polish Public Agencies Don't Know If They Use AI as Law Deadline Nears

In just under a month, on August 2, 2026, the next round of provisions under the European Union's AI Act will take effect, and a significant share of Polish public agencies still don't know whether they even use artificial intelligence systems covered by the regulation. That's the finding of a report by Watchdog Polska, which analyzed responses from thousands of public institutions about how and why they use AI.

The report, published in April 2026, is based on inquiries sent to 5,480 local government units, ministries, courts, healthcare facilities and municipal companies. Around 1,500 of them did not respond at all. Among those that did, there is considerable interpretive confusion over what actually counts as an artificial intelligence system under the EU regulation.

A Definition That Fits No One

The AI Act defines an AI system as a solution designed to operate with a degree of autonomy, capable of adapting after deployment, and generating outputs such as predictions, recommendations or decisions that influence physical or virtual environments. That definition covers everything from chatbots built on large language models to scoring systems, predictive models and the classic machine-learning tools administration has used for years.

The problem is that agencies interpret this definition very loosely, bending it to suit their own needs. Some institutions consider only content-generation tools like ChatGPT to be artificial intelligence, while others treat ordinary OCR software for scanning documents as AI. Still others refuse to disclose information about advanced analytical models, claiming their tools don't meet the regulation's definition.

ZUS, STIR and ABW Say No

The report cites specific examples of this definitional gymnastics. Poland's Social Insurance Institution (ZUS) uses a predictive model that analyzes sick leave and assigns policyholders a risk score, yet the institution claims this tool is not an artificial intelligence system, which lets it sidestep the obligations imposed by the AI Act.

Similarly, STIR, the risk-indicator system run by the Head of Poland's National Revenue Administration and used to block taxpayers' bank accounts, relies on analytical models, but the technical details remain classified under the guise of confidentiality. The Internal Security Agency (ABW) went even further, refusing to answer at all, arguing that disclosing even general information could reveal the security services' operational methods.

Every entity must independently assess whether its system meets the AI Act's definition, and broadly interpreting narrow definitions may reflect attempts to escape rigorous procedures - Paweł Dymek, lawyer

Disputes over definitions often serve as a pretext for a lack of transparency - Michał Zemełka, author of the Watchdog Polska report

What Watchdog Polska Proposes

The report's authors propose several concrete solutions. The first is introducing a single, operational definition of artificial intelligence for the entire public administration, so agencies can't arbitrarily narrow or broaden their interpretation depending on what's convenient. The second is creating a central, publicly accessible registry of AI applications used by state institutions.

Watchdog Polska also calls for an independent verification mechanism for cases where an agency refuses to disclose public information by invoking exceptions to the definition. Crucially, systems should be classified based on their actual functionality rather than on institutions' own declarations, since, as the ZUS and STIR examples show, those institutions have an incentive to avoid classifying their tools as AI.

What This Means for Citizens and Businesses

For the average citizen, this confusion means it's currently hard to determine which automated decisions affecting their life, from risk assessments at ZUS to account blocks under the STIR system, are even subject to oversight and the right to an explanation. The AI Act grants citizens the right to be informed when a decision was made or supported by an artificial intelligence system, but that right is meaningless if an institution simply denies using such a system in the first place.

For businesses working with the administration, it creates uncertainty about their own obligations. If agencies themselves can't consistently classify their systems, it's hard to expect coherent guidance for technology vendors and subcontractors supplying analytical or scoring tools to the public sector.

Little time remains before the next stage of the AI Act takes effect, and as the Watchdog Polska report shows, much of the administration still treats the issue as a distant legal matter rather than a real obligation of transparency. The mere passing of the deadline is unlikely to resolve the definitional disputes that have blocked effective enforcement for months.

Sources: AI Act Is Coming Soon, and Public Agencies Still Don't Know If They Use AI (prawo.pl)