"The Unholy Trinity of AI: Article 14, Transformers and Responsibility - The Epistemically Responsible Street-Level Bureaucrat"

This paper suggests that the currently ensuing situation with article 14 of the EU AI Act asks for dedicated human intervention and options to stop processes if we have an escalation in the processing of generative artificial intelligence systems that are running in organisations the EU, for this reason, has a distinction on three different risk layers, and each application is measured against its criticality and the actual operations of an AI in the public sector can be considered with good reasons as being on the high risk level (Voigt & Hullen, 2024). Per se, we have as well, embedded in today's predominant models. It is the so called transformer model, by which against the input token of a user, based on probabilistic calculation, an output is generated that is responding to that what the AI thinks is the next best word that it would fill in to close and reply to the user's request (Vaswani et al. 2017). We have as well coming from article 14, the actual quest for responsibility, the responsibility, however, as well, is part of the actual job of the street level bureaucrat who is there to enact and execute the actual law at an operational level. In the way how today, AI is implemented in the public sector, we can see that professional values and corresponding elements of holding each other accountable among colleagues are dimished or dried out in the light of the implementation of (Gen)AIs (Vredenburgh, 2025; Busch, 2025). However, if we want to have on the one hand side, actual verification on the outputs of an artificial intelligence (Krone, 2026 forthcoming), and we want to be compliant to the actual article 14 stipulation that outputs must be verified and have to be held in balance to mitigate and prevent dangerous behaviour we have a lack of resources because we cannot make the AI itself its own supervisor (Biswas & Talukdar, 2025). An agentic system would only raise the escalation levels so that we are back and need to consider to which extent the only trustee that we have left in the loop is the actual street level bureaucrat, whom we have to support to deliver, verify and engage with the social world around them, to define and assess to which extent the token outputs are true in respect to the social norms of the occupational group of street level bureaucrats, to which extent those outputs matter in a correct and good way for the actual client (Vredenburgh, 2025; Krone, 2026 forthcoming). And as well, we need to consider to which extent society in its own right, even if having a critical eye to the actual discretionary levels that we have, that possibly it is better to have a critical eye to the actual street level bureaucrat discretion (Vredenburg, 2025), instead of relying on an unreliable probabilistic output of an generative artificial intelligence. This paper, for this reason, suggests that we need to equip street level bureaucrats with a knowledge integration square knowledge as well as with dedicated tools that gives them easier access to verify when an artificial intelligence is talking nonsense that looks right (Krone, 2026 forthcoming).