Reproduce “When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yulin Chen, Yang Zhan, Zhuoran Lu, Qiao Jin
A scoping review of 58 studies examines how AI assistants built around a single user become involved with families, coworkers, bystanders, third parties, and shared infrastructure. Across 118 risk instances, it identifies five operational pathways and five risk domains, while finding that only 27.1% of instances had an implemented or tested safeguard.
Key insights
- Single-user assistants can become multi-party systems through ordinary mechanisms such as shared accounts, persistent memory, delegated communication, common retrieval infrastructure, and ambient sensing.
- Information-boundary failures are the most frequently reported problem, with 52 instances, followed by authority and resource-control failures with 21 and collective coordination failures with 17.
- Protecting identity at input time is insufficient: person-specific context, consent, ownership, and authorization must persist through storage, retrieval, planning, summarization, and external action.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.