Imagine a world where AIs share the same imaginary friends, invent the same fictional laws of physics, and tell the same made-up stories. That’s not science fiction—it’s the surprising reality revealed by new research from Salesforce AI. In a fascinating study titled "Shared Imagination: LLMs Hallucinate Alike," researchers discovered that large language models (LLMs) have a hidden talent for collaborative hallucination. They don't just make things up—they make up remarkably similar things. This discovery was made using a clever technique called "imaginary question answering" (IQA). One LLM is prompted to create a fictional scenario, like a made-up physics concept, and then devise a multiple-choice question about it. Another LLM, completely unaware of the first one's fabrication, is then asked to answer. The astounding result? These LLMs often answer each other's imaginary questions correctly, at a rate far exceeding random chance. Even more surprising, when given a paragraph of fictional context, the accuracy shoots up to a staggering 86%. This shared “imagination space,” as researchers call it, points to a deeper homogeneity among LLMs than previously thought. While their performance on standard benchmarks can vary widely, their tendency to invent similar concepts and answer fictional questions reveals a common thread in how they process and generate information. This discovery has major implications for the field of AI. For one, it could help explain the mystery of why LLMs hallucinate, offering clues on how to build more trustworthy and accurate models. It also raises fascinating questions about computational creativity. If AIs are predisposed to similar flights of fancy, how can we encourage truly unique and innovative outputs? This research also opens up new avenues for detecting hallucinations. By understanding the patterns in how LLMs invent, we might be able to better distinguish fiction from fact in their generated text. The shared imagination of LLMs offers a tantalizing glimpse into the inner workings of these powerful tools, urging further research into this uncharted territory. This discovery not only reveals the unexpected homogeneity in how LLMs hallucinate but also offers new pathways to address the challenges of hallucination, trustworthiness, and the very nature of creativity in artificial intelligence.