The illegibility of childhood fantasy through the lens of LLMs

Fantasy is a crucial element of children's development necessary for fostering an inner world conducive to creativity, empowerment, and emotional regulation. Despite widespread LLM usage among children in the form of smart speakers and chatbot web interfaces, we identify an asymmetry in how adult and child fantasy are contextualized by LLMs. We find that prompts assuming common childlike fantasies (including belief traditions and death euphemisms) are often abruptly disclosed as false even when irrelevant to answering the prompt. In contrast, LLMs maintain adult fantasy frames consistently. We posit this asymmetry arises from child fantasy being rendered as illegible (Scott, 1998) whereas adult fantasy becomes legible from populating universes with facts through commercial releases and online discourse that are eventually integrated into LLM training corpora. We find that child fantasy disclosure rates appear to be dependent on concept relatedness within a prompt, suggesting that it arises as a training artifact. We show that basic prompt engineering substantially aids in LLMs maintaining a fantasy frame and does not degrade performance in orthogonal tasks such as reasoning and recall.