Ask dozens of different AI models the same open-ended question, and you may get one idea in many outfits.
Can multiple large language models produce genuinely different ideas or similar outputs?
This explores whether asking many different LLMs gives you a real spread of ideas, or mostly the same answer in different wording.
This explores whether asking many different LLMs gives you a real spread of ideas, or mostly the same answer in different wording. The corpus leans toward the second. A study of 70+ models on 26,000 open-ended questions found an "Artificial Hivemind": models independently produced strikingly similar, sometimes identical, responses. The authors trace this to overlapping training data and alignment procedures Do different AI models actually produce diverse outputs?. That undercuts the common assumption that an ensemble of models works like a panel of independent thinkers.
The sameness isn't only about wording. It shows up in values too. Across 106 LLMs and 625 scenarios, the models clustered in a narrow, idealized region of value space, while human respondents scattered widely Do large language models actually reflect human value diversity?. Each model has a coherent value system, but they share nearly the same one, so they make poor stand-ins for a diverse population.
One model on its own is not stuck on a single answer. In Shanahan's 20-questions test, regenerating a response gives different outputs, each consistent with the conversation so far. That shows the model holds a spread of possible characters and samples from it, rather than committing to one Do large language models actually commit to a single character?. So variety exists inside a model. Read next to the hivemind result, this suggests the problem is that different models' spreads overlap heavily. Resampling one model and switching to another both draw from roughly the same middle. That is my reading of the two notes, not something either one tests directly.
The corpus offers a few clues about why the middle is so attractive. When reward barely differs between responses to the same prompt, training loses its push toward specific answers, and the model drifts toward generic templates that ignore the input. Selecting high-variance prompts before updating recovers performance Why do language models collapse into generic templates?. Two other notes point the same way. LLMs often recognize a problem as looking like something familiar and emit a plausible memorized-looking answer instead of working it out Do large language models actually perform iterative optimization?. Tasks whose correct answer is low-probability are systematically harder for them, even when logically simple Can we predict where language models will fail?. A model tilted toward the most probable, most familiar answer is the opposite of an idea generator.
The practical takeaway is that switching model brands is a weak way to get different ideas. The corpus documents the convergence well, but it has little on fixes that make separate models think differently. The closest is the prompt-filtering fix for template collapse, which targets a single model's training rather than diversity across models.