Social Representations and Genai: Representations of Children through Synthetic Imagery

Generative Artificial Intelligence (GenAI) took the world by surprise in 2022, heralding a new AI boom. The success of the platforms of large-language-models such as ChatGPT, Claude and Gemini and Diffusion models such as Dall-E and Midjourney can be attributed to its relatively reliable results which largely surpass previous models. Behind this technology, however, ethicists call attention to the practices necessary to power them up. Bender et al (2021), some of the most notable critics and former insiders, noted that Generative AI is trained on (by scraping) internet sized data. We note that the functioning of these platforms is based on the creation of token probability, and that in its internal architecture the models assemble together concepts in order of importance/dominance according to the query (see Wendler et al, 2024). In this paper, I present the idea that this architecture sounds very similar to the concept of social representations by Moscovici (1988) and notably its interpretation by Abric of the core and peripheral representations (1994). Bringing to fore the internet-sized of their training data, I investigate the GenAI representations of a group of subjects who are particularly prone to be reduced to certain icons and representations - children (James & Prout, 1998), and I employ a novel, exploratory multimedia elicitation of meanings and senses from Generative AI models through Imagery. I ask the question: what kind of representations ChatGPT and others hold of children, and does it reflect the classic works about children imagery?