On replacing humans with human simulators in human-in-the-loop systems built to interact with humans

With the rise of large language models (LLMs), which can generate fluent text and follow users' instructions, many attempts have emerged to replace humans on a variety of occasions: data annotation, customer service, online tutorials, and even user testing. While it might seem that LLMs could easily take over many of these tasks—especially those that are already highly automated with minimal human involvement—we want to take a step back and consider whether these promising replacements would truly respond to the reasons we put humans in the loop of computer systems in the first place. In this chapter, we present a case study focusing on a specific area of research, namely interactive crowd-powered systems. These systems incorporate human crowd workers, who can be recruited and managed rapidly and programmatically from the internet, into the loop of computer systems to interact with users in (nearly) real-time. We first overview the history of such systems, discuss the original motivations for having human workers within these systems, and look into a series of recent attempts to replace these humans with LLMs. Realizing some gaps within these efforts, we then zero in on an interesting case study, comparing two projects in parallel: both studied users' interactions with voice-based smart speakers that can hold open-domain conversations, one powered by an LLM, and the other powered by human workers. Through this comparison, we argue that while last-mile interaction issues such as conversation cut-offs and persistent speech recognition problems across both human-powered and LLM workflows, each approach requires distinct considerations. Specifically, certain accommodations must be made when replacing human workflows with LLMs. LLMs may introduce new challenges that are rarely encountered by human workers, and vice versa.