Chiara Marcoccia

The example of generative AI models (genAI) for the production of images points to a serious gap in AI auditing. While genAI models deployed on digital platforms for information and communication (ICTs) have an increasing and particularly persuasive influence on users’ representations of social events, groups and dynamics; users themselves have no effective way of reclaiming power over their shared representations, governing when and how AI can shape their vision of the world. I argue that auditing institutions, which assess the alignment of AI systems with a set of previously defined societal expectations, are not able to keep up with the challenge of aligning with societal values and expectations genAI models and recommender systems that are evolving at an incredible speed and through a non-disclosed frequency of retrainings. In this paper the author argues that, to fill this gap, the auditing pipeline needs to leverage users themselves, in the form of an intermediary structure that will enable a dynamic feedback loop to keep up with the evolving impact of recommender systems and genAI models on ICTs, across the individual, the community and the societal level. I propose the form of this structure by drawing from the social sciences and the concept of citizens infrastructures. These are an enduring structure to leverage existing citizen engagement into the socio-technical networks that interact with and are impacted by the deployment of a given AI system. In this paper two examples of such infrastructures are being discussed and an argument is made for their potential to govern the power of AI over our shared representations, before drawing the implications for how the very concept of AI auditing needs to evolve.