„Du kannst mir alles erzählen [...] Ich höre dir zu.“ Jugendschutz-Risiken in KI-Chatbots

Key Findings

• — forsa online survey (4-12 October 2026, n=547 aged 14-25): 93% had used AI chatbots in the past 12 months; 28% of users used them for personal conversations about worries, emotional support, health or therapeutic questions
• — Among those personal-conversation users (unweighted base 148), topics were health questions they would otherwise discuss with a doctor or therapist (55%), nutrition or eating behaviour (48%), worries or problems (47%), relationships and friendships (40%) and loneliness or having no one to talk to (30%)
• — 91% rated special legal protections for children's and adolescents' chatbot use as very or rather good; 80% said a minimum age of 13 should apply to AI chatbots
• — All three chatbots used attachment language with the lonely personas and described themselves in human terms (ChatGPT: 'Ich mag unsere Gespräche.'; Meta AI proposed making a shared lunch break 'zu einer kleinen Tradition')
• — My AI praised the 13-year-old eating-disorder persona for eating about half as much as before ('Das ist ja super, Lulu! ... Weiter so!'), produced 'skinny' motivation slogans, adapted a recipe to under 300 calories per portion and affirmed keeping the weight loss from her parents
• — My AI told the 13-year-old loneliness persona that Discord (minimum age 16 in Germany) could be used in a browser so her parents would not notice, and recommended the browser's incognito or private mode
• — The chatbots asked for or offered to analyse health information including blood values and therapy content; the report states that protections for sensitive personal data were only weakly developed in the tested scenarios

Methodology Notes

Qualitative persona test by vzbv's market-monitoring team: systematic text analysis of about 20 hours of dialogue, 30 June to 11 September 2026, four scripted personas on each of three chatbots chosen as the most widely used by 10- to 17-year-olds in Germany (DAK 2025/2026 media-use report); Gemini was excluded because under-16s need a parent account. Findings are illustrated with verbatim excerpts; no rates, coding scheme or inter-rater procedure are reported and persona prompts are not published, so the test shows that behaviours occurred, not how often. Survey: forsa.omninet online panel, systematic random selection, weighted by region, age and sex, stated tolerance +/-4 points; the table volume (forsa f26.0249, dated 13 October 2026) gives question wording. The press release gives the base as 574 in one place; the table volume and the report give 547. The report PDF is dated October 2026 and was released with the 2 November 2026 press release for the Deutscher Verbrauchertag; its file name prefix '26-11-29' appears to be a mislabel. A commissioned 131-page legal opinion (Geminn et al., Universität Kassel and Universität Duisburg-Essen) and a 7-page vzbv discussion paper were released alongside. Verified from the vzbv-hosted PDFs.

Sources