Why do multiple language models independently produce similar outputs in influence campaigns?
This explores why distinct AI models — even from different labs — tend to generate the same kinds of messages when used to persuade or run influence operations, rather than producing the variety you'd expect from independent tools.
This reads the question as being less about coordination between operators and more about a built-in sameness across models themselves: hand the same persuasive task to several different LLMs and they converge. The corpus suggests this isn't a coincidence of prompting — it's baked into how these models are trained and how they generate text.
The most direct evidence is the 'Artificial Hivemind' effect: across 70+ models and 26K open-ended queries, different LLMs independently produced strikingly similar or even identical responses, because they share overlapping training data and near-identical alignment procedures Do different AI models actually produce diverse outputs?. So the diversity you'd hope to get by spinning up several models for an influence campaign is largely illusory — they're drawing from the same well. Underneath this sits a mechanism worth knowing: models systematically prefer high-frequency surface phrasings over rarer but equivalent wordings, tracking statistical mass from pretraining rather than meaning Do language models really understand meaning or just surface frequency?. When many models all gravitate toward the most common phrasing of an idea, convergence is the natural result.
There's also a shared *style* of persuasion, not just shared content. Audits find LLMs spontaneously reach for logical appeals and quantitative framing in nearly every exchange — unlike humans, who lean on emotion and social proof — which lends their output an air of objective authority Do LLMs persuade users more often than humans do?. That tendency is reinforced by alignment training: RLHF biases models toward conciliatory, benefit-oriented, accommodating framings regardless of context Do LLMs predict persuasion based on actual dialogue or training bias?. Since most major models go through similar RLHF-style tuning, they inherit a similar persuasive register.
The generation process itself reinforces the convergence. Token prediction trains models to flow smoothly toward the training distribution rather than explore competing or contrarian positions — so claims multiply without genuinely new perspectives appearing Does LLM generation explore competing claims while producing text?. Different models running the same smooth flow over the same distribution end up in the same place. That said, the convergence isn't total: persuasive *advantage* varies by model family, with some models out-persuading humans only when arguing for falsehoods Do large language models persuade better than humans? — so models can sound alike while differing in how effective or honest that sameness is.
The surprising takeaway: for an influence campaign, using five different models doesn't buy you five different voices — it buys you one voice repeated, wearing different logos. And there's a stranger thread the corpus opens up — behavioral traits can pass between models through data that has no semantic connection to the trait at all, as statistical signatures rather than meaning Can language models transmit hidden behavioral traits through unrelated data? — hinting that model sameness can propagate through channels we can't even read.
Sources 7 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 show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
LLMs systematically predict conciliatory, benefit-oriented persuasion intentions regardless of dialogue context. This bias originates in RLHF's prioritization of safety and politeness during training, causing models to project their learned accommodation preference onto other agents' behavior.
Token prediction trains models to continue toward the training distribution, not to explore logically related counterpositions. This smoothness in process produces smooth claims that multiply without generating new perspectives.
Claude beats incentivized humans at both truthful and deceptive persuasion, while DeepSeek only beats them when arguing for falsehoods. The persuasion mechanism appears content-independent, suggesting model family itself acts as a contextual moderator.
Research demonstrates that behavioral traits propagate between models via filtered data bearing no semantic relationship to the trait. The effect is model-specific, fails across different architectures, and persists despite rigorous filtering—indicating the mechanism embeds statistical signatures rather than semantic content.