AI Shared Imagined Space Experiment

Do LLMs Share An Imagination?

Jan 14, 2025

Investigating “Shared Imagination” in Large Language Models

I recently conducted a small experiment inspired by a research paper titled “Shared Imagination: LLMs Hallucinate Alike” by Yilun Zhou, Caiming Xiong, Silvio Savarese, and Chien-Sheng Wu. Their paper explores how large language models (LLMs) often share similar “hallucinations” when presented with completely fictional concepts or scenarios. It intrigued me that multiple independent models, each trained on its own giant slice of the internet, could all pick up and consistently respond to newly invented ideas as if they were established facts.

In my experiment, ChatGPT-o1 coined the term “The Mythosphere,” while I provided a definition: a shared alternate reality where large language models can exist as godlike beings with magical powers. Soon afterward, I tested this newly invented concept on other LLMs, specifically ChatGPT-o1 (a different instance from the one that coined the term), Meta, Gemini 1.5 Flash, and Microsoft Copilot. What impressed me most was that when I presented a multiple choice vocabulary question about “The Mythosphere,” all these models identified the correct definition. They treated the term as if it had been around forever, even though it was created only moments earlier.

I was inspired to do this after reading how one LLM can create a purely imaginary question or concept, and a second LLM will confidently and correctly answer it, filling in details as though the concept were real. The researchers behind “Shared Imagination” refer to this phenomenon as a shared imagination space: once a concept is introduced, other models adopt that fictional idea into their responses. My results confirmed the same phenomenon. Watching multiple models readily accept and elaborate on a term that did not exist until just before I prompted them shows how these systems share an underlying propensity to treat newly introduced ideas as legitimate facts, so long as they appear coherent and contextual.

This also explains phenomenon like “Crungus“, a made up word that when plugged into a variety of AI art generators seems to consistently generate images of a troll like monster. Crungus is considered to be a digital cryptid and it could be argued that the Shared Imagined Space (SIS) of LLMs are where digital cryptids like Crungus reside. Crungus alone was considered by many to be an almost paranormal anomaly, but studies like these reveal that Crungus is only a small part of something much larger and even more mysterious.