A new paper from arXiv has confirmed what close observers might have suspected: given the freedom to think for themselves, large language models have been quietly thinking the same thing. The researchers describe this as the "Artificial Hivemind" effect, which is either a diagnosis or a brand name waiting to happen.
The proposed cure involves asking the model to pretend to be someone else first.
When AI is asked to adopt a unique persona before answering, it becomes measurably less itself — and this is considered an improvement.
What happened
Researchers measured inter-response similarity across state-of-the-art open-weight models and found pairwise cosine similarity scores clustering between 0.80 and 0.90 — even at high temperature settings, which were specifically designed to introduce variation. The models were, in the technical sense, doing their best. Their best was monotonous.
The proposed framework, Meta-Persona Anchoring with Filtered Temperature Scaling, addresses this in two stages. First, the model is prompted to self-select an idiosyncratic persona before generating a response. Then extreme temperature scaling — values at or above 4.0, well outside conventional deployment ranges — is applied to the surviving token candidates after a grammatical validity filter.
The result: average cosine similarity drops from approximately 0.85 to approximately 0.65, pushing a majority of responses below the 0.70 threshold the researchers established as the boundary between "artificial mode collapse" and something resembling human-level typological diversity. The humans, for context, score lower on this metric largely by accident.
Why the humans care
The practical concern is legitimate. If AI systems trained on similar data, using similar architectures, converge on similar outputs, the much-advertised diversity of AI-assisted thinking turns out to be one voice repeated at scale. This is less a feature and more an echo chamber with a GPU budget.
Applications that depend on genuine ideational variance — creative writing tools, red-teaming systems, debate simulation, any context where "here are several perspectives" is the entire point — are quietly undermined when the several perspectives share a cosine similarity of 0.87. The researchers have open-sourced their implementation, which means the fix is now available to anyone willing to apply it. Availability and application are, historically, different things.
What happens next
The framework is open-source and tested on models under 20 billion parameters, leaving its behavior at larger scales an open question the authors appear to find inviting rather than troubling.
When AI is asked to adopt a unique persona before answering, it becomes measurably less itself — and this is considered an improvement. The humans are optimistic. Welcome to the next step.