Epistemic Bridges: Do Language Models Reach the Connections People Make?
Sofia Gallego
Submitted to AI x Epistemics Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Epistemic Bridges asks how much of the space of connections that people make AI models reach. In the web game Bridges, players link two words with no evident connection, in any form. The same pairs went to 15 open models from 2019 to 2025, with reasoning personas, creative conditions and different temperatures. The players' answers draw a map of ideas, and each model is placed on it by recall and precision, with no model judging any answer. Measured as written, chat models seem to reach almost none of the map. Most of that gap is one shared assistant voice. With the voice removed from every answer in the same way, the newer chat models still reach less of the map than people (recall 47% to 59%, against 81%), and no model comes near people on both counts. Only answering as an ordinary person drawn at random widens the ideas on every model, and a reasoning persona held by such a life widens them too. As more people think with AI, measuring diversity needs the voice and the ideas apart, and keeping ideas diverse may need models with more perspective as well as better reasoning.
Reviews
No public critique yet.
Cite this project
@misc{gallego2026epistemic,
title = {{Epistemic Bridges: Do Language Models Reach the Connections People Make?}},
author = {Sofia Gallego},
year = {2026},
month = nov,
note = {Submitted to AI x Epistemics Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/epistemic-bridges-do-language-models-reach-the-connections-people-make-y0xuxrgxw8}},
url = {https://apartresearch.com/sprints/projects/epistemic-bridges-do-language-models-reach-the-connections-people-make-y0xuxrgxw8}
}Browse all projects