Child-focused suitability investigation of current Information Access Systems

Abstract

Information Access Systems (IASs), such as Search Engines (SEs) and Large Language Model-based conversational agents (LLM-CAs), have become central tools through which users access information online. Their widespread use extends to children, who increasingly rely on these systems for learning and entertainment. However, children cannot be treated like adults as they differ in their suitability requirements.

Current IASs are primarily designed and optimized around general users, despite children having distinct requirements. At the same time, recent developments in generative AI have added more ways in which information is accessed, with LLM-CAs providing a conversational search approach and SEs increasingly integrating AI-generated summaries into presented results. Existing research has investigated children’s interactions with online information environments, yet it remains under-explored whether modern IASs textually portray content in ways that fit children’s needs across both retrieval-based and generative forms of information access. In addition, IASs operate across languages that differ in linguistic structure, content availability, and system support, making findings in one language not necessarily generalisable to others. As children’s information needs differ from those of adults, and most existing IASs are primarily designed for general audiences, questions arise regarding the suitability of these systems for child users. Therefore this work investigates the suitability of portrayed content for children across IASs and languages, focusing on two complementary dimensions: comprehension and harm exposure. Comprehension is operationalized through readability, ambiguity, and concreteness measures, while harm exposure is investigated through the detection of potentially harmful content categories, including harassment, hate speech, sexual content, and violent content. We conduct an empirical investigation using child-written queries in English and Dutch, collecting responses from traditional SEs and LLM-CAs. The response types we obtain from these two types of IASs consist of SERP results, AI-generated search summaries, retrieved webpages, and LLM-CA responses. To investigate suitability, the responses were investigated using a collection of automated indicators associated with child comprehension and harm exposure. In addition to comparing different systems and response types, the study examined differences between child-written and general-user queries as well as differences between English and Dutch responses.

The results suggest that currently adopted IASs display varying degrees of child suitability. Responses to child-written queries consistently achieved higher readability scores than responses to general user queries. However, improvements in readability did not consistently extend to improvements in ambiguity and concreteness. In contrast, harm-exposure scores remained low across the investigated response types, suggesting positively limited exposure of users to overtly harmful content. Furthermore, the investigation revealed that suitability characteristics do not always generalise across different languages. while readability-related findings remained relatively stable between English and Dutch, ambiguity and concreteness measures showed several language-connected differences.