Characterizing LLM-Based Family Education through the Lens of Activity Theory: A Scoping Review of the HCI Literature

Synthesis: Luo and colleagues conduct a scoping review of 53 HCI studies, drawn from 6,540 records retrieved across 19 venues, to explain what happens when large language models (Large Language Models (LLMs)) enter family education. Using Activity Theory operationalized through the Activity-Oriented Design Method, they relate participants and educational objects to mediation, labor, and rules. The corpus centers on child–parent interaction — 44 of 53 studies — and on language, AI Literacy, and relational learning. Conversational, embodied, and spatial systems generated support from the context of an unfolding interaction, and four forms of mediation recurred: producing materials, structuring dialogue, interpreting context, and connecting family activity with external knowledge. LLMs redistributed educational labor, but family and institutional rules left parents and professionals responsible for interpreting outputs and deciding how they entered practice. The review's contribution is a framework explaining how LLM capabilities become organized through family participation, together with a research agenda. Evidence remains thin on sustained personalization, repair labor, and how families negotiate authority and rules.

Key Findings

1. The review synthesized 53 HCI studies from 6,540 records across 19 venues, including 23 published in 2025 and 18 in 2026.
2. Child–parent configurations dominated at 44 of 53 studies (83%); one addressed grandparents and grandchildren, four other arrangements, and none siblings.
3. Educational objects clustered on language, literacy, and narrative development, cognitive and emotional development, and AI and digital literacy; mathematical reasoning and scientific inquiry were weakly represented.
4. LLMs mediated family activity by producing materials, structuring dialogue, interpreting context and providing guidance, and connecting family activity with external professional knowledge.
5. Reported outcomes mixed learning measures with design requirements, participant perceptions, and technical benchmarks; desired features are not observed benefits.
6. LLM use redistributed educational labor but left parents reviewing generated content, adapting recommendations, and retaining final decisions, while verification and repair work was weakly documented.
7. Rules governing participation, acceptability, and information boundaries were mostly anticipated rather than enacted, and conflicts between household, platform, and institutional obligations were seldom examined.

Family Education as a Cooperative Activity System

Family education here includes children's learning with family members and parental learning within family life. It is cooperative: guidance from one member creates conditions for another's learning, and the learner's response shapes how the interaction proceeds. The authors adopt activity theory because it takes the activity system as the unit of analysis, examining how people pursue an object through a mediating tool within a socially organized setting. The lineage is sociocultural — Vygotsky's cultural tools, Leontiev's motive-driven activities, Engeström's collective activity system — and the Activity-Oriented Design Method translates those components into coding rules for object, subjects, tools, rules, community, division of labor, and outcomes. The framing matters because a system's educational role cannot be read off its interface; it depends on how the activity develops and who carries it forward.

How LLM-Based Systems Mediate Learning

For family-facing systems, LLM capabilities became educationally relevant through their interfaces and participant arrangements. The review distinguishes conversational interfaces that organize an exchange through questions and turn-taking; physically embodied systems such as robots that coordinate attention through presence, gaze, and gesture; and spatial or immersive systems that connect interaction with physical objects and artifacts. Four mediation functions recurred across them: producing and adapting materials, structuring dialogue and reflection, interpreting context and providing guidance, and connecting external knowledge and values. One capability could assume several roles — content generation might support a child's learning, supply a parent with teaching materials, or create a shared object for family discussion. Generation supplies a provisional resource whose relevance is completed through family participation: a co-reading question enters a parent–child exchange, a homeschooling draft is judged against curriculum and family values.

Organization, Rules, and What Remains Unresolved

LLM use redistributed educational labor without establishing a net reduction in family workload. Systems took on generation and preliminary analysis, while parents evaluated fit, adapted outputs, guided learning, and made educational decisions. Parents were expected to review generated content in reading, home practice, media assessment, and homeschooling, yet studies reported less consistently who detected inappropriate output or explained failures to a child. The authors call this verification and repair labor, and note that making work visible can concentrate it on the person already carrying most family responsibility. Rules separately governed who could use a system, which contributions could enter an activity, and how information moved: participation rules coordinated child autonomy with adult authority, acceptability rules invoked situated judgment, and information rules linked household Privacy with external accountability. Most rules were elicited through interviews, workshops, or design probes rather than observed during sustained use. The review also finds family relationships beyond children and parents underrepresented, family circumstances unevenly examined, and evaluations that rarely connect learning, family interaction, and technical performance — gaps bearing on Equity.

What this means for practice

• Instructors. Treat generated material as a draft requiring local judgment: check developmental fit against the specific learner, adapt to the activity, and record what you changed.
• Parents and families. Budget for the checking work, not just the generation: decide who reviews outputs, who corrects errors, and who explains failures to a child.
• Designers. Preserve provenance and keep system output, learner response, caregiver revision, and professional approval distinguishable rather than collapsing them into one artifact.
• Researchers. Measure system performance alongside changes in family interaction and educational outcomes in the same study, and follow generated analyses into the decisions they inform.

Limitations

• The venue-based search bounded the review to selected HCI publication outlets (53 studies across 10 venues, 30 in CHI); relevant work in education, communication, and clinical research may be omitted.
• The corpus combines evaluated systems, investigations of needs and expectations, and retrospective computational analyses, so the synthesis maps contributions without estimating a common intervention effect.
• Representing each study through a single focal activity configuration facilitates comparison but can underrepresent secondary relationships and changes within an activity.
• The AODM lens foregrounds relations among participants, tools, purposes, and responsibilities; other frameworks could expose different patterns.

Citation

Luo, L., Liang, Y., Cai, J., Wang, A., Pan, D., Zhou, M., Yu, C., & Hui, P. (2026). Characterizing LLM-Based Family Education through the Lens of Activity Theory: A Scoping Review of the HCI Literature. arXiv preprint.