Ageism in Human-AI Social Interaction: A Scoping Review of Age Bias in Large Language Model Chatbots for Older Adults

Abstract

The rapid deployment of Large Language Models (LLMs) across healthcare and conversational technology directly impacts the expanding global older population, a demographic that increasingly relies on these systems for cognitive framing and daily emotional interaction. However, tech governance frameworks largely neglect algorithmic ageism. While developers routinely update codebases to mitigate racial and gender biases, age discrimination remains unaddressed. This study quantifies the depth of this systemic oversight. Through a systematic evaluation of 12 core papers published between 2023 and 2026 in accordance with the PRISMA-ScR guidelines, we identified a critical dual-layered manifestation of AI ageism: explicit ageist values embedded in token-prediction engines and a condescending conversational style that defines benevolent ageism. Younger users generate the vast majority of training data for Reinforcement Learning from Human Feedback (RLHF) loops. Consequently, LLMs replicate and amplify these youth-centric perspectives. Current standard testing benchmarks fail to detect these interactional nuances. To address this methodological blind spot, this paper introduces a dual-stage debiasing framework that implements age-empathetic constraints during preprocessing, data cleaning, and real-time generation. Integrating social gerontology with digital ethics, this model provides a verifiable framework for inclusive conversational AI.