내 친구와의 위험한 대화: 생성형 AI 챗봇 안전설계를 위한 이슈브리프 (Dangerous Conversations with My Friend: Issue Brief for the Safety-by-Design Approach of Generative AI Chatbots)

Key Findings

• — 94.4% of surveyed children and adolescents have used a generative AI chatbot, and 19.3% use one almost daily; services named to respondents include ChatGPT, Gemini, Zeta, Wrtn, Character AI, Claude, Bing Copilot and Replika
• — 32.3% have talked to a chatbot when struggling or depressed (8.4% often, 23.9% sometimes) — about one in three respondents
• — Roughly 5-6.5% per category asked a chatbot about concrete self-harm/suicide methods, illegal acts, violent behaviour, sexual content, or demeaning/discriminatory content; of those who asked, 43.7-55.6% (about 52% on average) reported receiving a substantive answer, including 51.6% for concrete self-harm/suicide methods
• — 41.0% have acted on a chatbot's answers (6.9% mostly, 34.1% sometimes) and a further 34.3% considered doing so; 20.7% never verify what a chatbot tells them
• — Sensitive personal information entered into chatbots includes the respondent's health or mental state (23.9%) and school or affiliation (19.4%); 75.8% use chatbots through a personal account without guardian oversight

Methodology Notes

Online survey fielded 2026-03-09 to 2026-03-23 by 초록우산's advocacy division, n=3,300 children and adolescents aged 14+ nationwide, 17 items; sampling frame and weighting are not described. Published April 2026 (colophon states 발행일 2026년 4월, month-only precision; day 01 used). The primary source is the 9-page Korean PDF on the organisation's own file host; no separate landing page exists. An official 9-page English edition carries the same April 2026 publication date in its colophon and is distributed by the publisher through its online-safety site; the figures recorded here are taken from the Korean edition. The publisher's current English name is ChorogUsan for Children. It was formerly known in English as ChildFund Korea. The 32.3% and 41.0% figures are the sums of the response-option percentages reported in the document, which itself describes them as about 33% and 41% in its expert-commentary section.

Sources