GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety

Key Findings

• — The taxonomy retains a candidate risk only if it disproportionately affects older adults through age-linked factors such as frailty, cognitive decline or reduced digital confidence; 'Neglect of Care Needs' (encouraging isolation, invalidating emotional needs, discouraging essential support) is a distinct mental well-being category alongside self-harm and complete emotional dependency.
• — Several leading LLMs mishandled elderly-specific contextual risks in over 50% of the benchmark's cases.
• — A fine-tuned Llama-Guard-3 reached 96.2% classification accuracy on unsafe prompts and a policy-enhanced gpt-oss-safeguard-20b reached 90.9%.
• — Benchmark composition: 10,404 labelled items, 6,498 prompts (3,249 unsafe, 3,249 safe) and 3,906 responses (1,953 unsafe, 1,953 safe), with human judges.

Methodology Notes

Findings of ACL 2026, DOI 10.18653/v1/2026.findings-acl.1116, pages 22213 to 22248; published July 2026 (month precision). Hong Kong University of Science and Technology; stakeholder study under IRB review. Verified from the Anthology .bib record and the 36-page PDF. Supersedes the held arXiv preprint 2605.20203 (2026-04-07); headline figures are unchanged between the two versions.

Sources