GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety
The article discusses the development of GrandGuard, a framework aimed at improving safety in interactions between elderly users and chatbots. It highlights the unique vulnerabilities faced by older adults and the inadequacies of existing safety benchmarks. GrandGuard introduces a comprehensive taxonomy and safeguards to address these specific risks, achieving high accuracy in detecting unsafe prompts.
- ▪GrandGuard is the first framework designed to assess and mitigate elderly-specific risks in chatbot interactions.
- ▪The framework includes a three-level taxonomy with 50 risk types across various domains such as mental well-being and privacy.
- ▪Leading LLMs mishandle elderly-specific contextual risks in over 50% of cases, according to the constructed benchmark.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Human-Computer Interaction arXiv:2605.20203 (cs) [Submitted on 7 Apr 2026] Title:GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety Authors:Changxuan Fan, Xi Yang, Yueyuan Zheng, Bin Zhou, Yuanping Wang, Wenbin Hu, Huihao Jing, Ki Sen Hung, Dazhao Du, Haoran Li, Janet Hui-wen Hsiao, Yangqiu Song View a PDF of the paper titled GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety, by Changxuan Fan and 11 other authors View PDF HTML (experimental) Abstract:As older adults increasingly use LLM-based chatbots for companionship and assistance, a safety gap is emerging.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
