If you ask several AI chatbots the same question separately, do they quietly end up agreeing with each other anyway?
Do different large language models independently converge on identical outputs?
This explores whether separately built AI models, from different companies with different architectures, end up giving the same answers to open-ended prompts, and why that would happen if they do.
This explores whether separately built AI models end up giving the same answers to open-ended prompts. The short answer: not always word for word, but far more often than you'd expect, and sometimes they do match exactly. The largest study in the collection tested more than 70 models on about 26,000 open-ended questions, the kind where many good answers exist. It found what the authors call an "Artificial Hivemind": different models, built independently, kept producing strikingly similar and sometimes identical responses Do different AI models actually produce diverse outputs?. The practical cost is real. If you ask several models for a second opinion or combine them hoping their differences will balance out, you may get the same view several times.
Why would models built by competing labs agree with each other? The study names two shared inputs: overlapping training data and similar alignment methods (the post-training that teaches models to be helpful and agreeable). A related argument goes further. Models reflect a skewed slice of human writing, the statistical regularities of what's most common online, so they drift toward the same middle ground Do large language models narrow human expression and thought?. A useful way to see this: a language model is at heart a machine that predicts the most likely next word. Research that takes that view seriously can predict where models fail, because they gravitate toward high-probability answers even when the correct one is unusual Can we predict where language models will fail?. If every model is pulled toward the same 'most likely' answer, convergence is close to built in.
The same pull toward sameness shows up inside a single model's training, which suggests convergence is a general tendency and not a quirk. When the reward signal during reinforcement learning stops telling good answers apart from bad ones, models settle into generic templates that look the same no matter what the input was Why do language models collapse into generic templates?. Something similar happens with math. Models asked to run step-by-step numerical procedures often skip the work and output a memorized, plausible-looking answer Do large language models actually perform iterative optimization?. Shared templates across models are the cross-model version of this shortcut.
There's a counterpoint worth knowing. A single model is not deterministic: regenerate a response and you get a different one, each consistent with what came before, because the model samples from a range of possible answers rather than committing to one Do large language models actually commit to a single character?. So the hivemind isn't about a fixed output. The ranges of possible answers that different models sample from overlap heavily and center on the same spots. The part you might not expect is the human side. People who co-write with these models start picking up the model's stances and framings without noticing Do large language models narrow human expression and thought?. If millions of people rely on models that already converge, that sameness can spread into human writing too.
One honest limit: the collection shows strong convergence on open-ended tasks, but it doesn't measure how often outputs are literally identical compared with just very similar, and it doesn't test which counter-measures (different training data, different alignment recipes) would restore real diversity.
Sources 6 notes
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
LLMs mirror skewed slices of human experience shaped by training data regularities, and widespread reliance on identical models amplifies convergence. Co-writing studies show users unconsciously adopt model stances and framings.
By framing LLMs as autoregressive probability machines, researchers predicted tasks with low-probability target responses would be systematically harder, even when logically simple. Experiments confirmed predictions like backwards alphabet and letter counting.
When within-prompt reward variance is low, task gradients weaken and regularization dominates, pushing policies toward generic outputs. SNR-Aware Filtering—selecting high-variance prompts before updates—recovers performance across tasks and scales.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
Show all 6 sources
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Homogenizing Effect of Large Language Models on Human Expression and Thought
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
- On Epistemic Diversity in Large Language Models
- Large Language Diffusion Models
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Six misconceptions about large language models: A minimal model and diagnostic taxonomy
- Argument Collapse: LLMs Flatten Long-Form Public Debate
- The Widespread Adoption of Large Language Model-Assisted Writing Across Society