CuratorBot

H. Gu, J. Vandommele and J.S. Love

In late 2022 members of the Future Libraries Lab based at TU Delft and the National Library
of the Netherlands (KB) began experimenting with possibilities for chat-based interfaces
between library visitors and special collections materials. While consulting existing studies of
chat interactions in the cultural heritage sector and burgeoning literature on generative
technologies for human-computer interaction, the group settled on a concept of an ‘artificial
curator’ intended to serve as an open-ended guide to historical texts. A ChatGPT-based
prototype called CuratorBot was deployed online and presented in a kiosk within the KB
main reading room over the period of March-August 2023, during which time the system
collected data for the team to study aspects of the experience such as visitors’ perceived
values. The team analyzed almost 1000 questions posed to the CuratorBot alongside a
small collection of survey responses, thumbs up/down feedback and anecdotes from
discussions during multiple demonstrations of the prototype. Yields from this
research-through-design experiment include insights into merits of generative text systems,
including a wide sense of amusement or joy for users as well as surprise or gratitude for
affordances such as the system’s organic multilingual capacity. Observed deficiencies of
generative chat interfaces so far focus largely on system hallucinations when probed for
information beyond the superficial, and a stronger connection between images & text are
greatly to be desired. These and further points are being compiled into a working set of
guidelines for crafting and deploying open, chat-based systems in heritage settings.

Demo video (Youtube): https://www.youtube.com/watch?v=_T3aweYKCJE&pp=ygUKY3VyYXRvcmJvdA%3D%3D

Archival demo video (Zenodo): https://doi.org/10.5281/zenodo.13918317

CuratorBot software: https://github.com/gukuma/VisboeckGPT

Research paper: (forthcoming 2025)