From Generic to Personalized: Lessons for Meaningful Human–AI Interaction in LLM-Supported Care Documentation

Abstract

We present an LLM-based documentation support system that extracts care-relevant information from nurse–resident interactions in residential care. While care workers generally responded positively, the inclusion of care worker–related information in reports remains a key issue. We investigate whether personalization via user profile injection can reduce such unwanted output. To this end, we introduce an LLM-based pipeline for automatically extracting care worker facts from transcripts. Results show that both profile injection and candidate-constrained prompting reduce care worker information, but lead to lower evaluation scores, indicating limited post-editing of the initial reports. We therefore argue for incorporating reflection mechanisms to mitigate automation bias, as well as feedback loops to further improve system performance and enable adaptation to individual documentation preferences.

Keywords: LLMs, Care Work, Documentation Support, Personalization, Human-AI Interaction

DOI: 10.54941/ahfe1008166