Robotic Sycophancy: A Scoping Review

Abstract

Sycophancy in social robots is an emerging threat brought on by the launch of ChatGPT and other powerful large language models (LLMs) that can speak in a near-fluent fashion. Short- and long-term findings on LLM-powered chatbots and conversational agents are raising the alarm. With work bridging communication-centred LLM use and social robots in production, the deceptive and persuasive capabilities of LLM-imbued robotic companions needs urgent and critical consideration. Notably, how social robots aided by sycophantically-inclined LLMs may overly influence decision-making and elicit overtrust needs interrogation. Using scoping review methodology that bridges robotics with AI and LLMs, we surface dimensions of sycophancy, constructs as research targets, and a suite of measures for research on robotic sycophancy. Our analysis of historical and modern studies (𝑁 = 23) sets the stage for empirical and theoretical work on the potential misuses and unexpected effects of sycophancy in human–robot interactions.