Is AI standardizing human thinking?
The rise of the Large Language Models (LLM) marked an irreversible turning point in the history of computer science and creativity. If until a few years ago the interest in the discipline ofArtificial intelligence While it was confined to the optimization of specific tasks or the solving of logical-mathematical problems, today the focus has shifted dramatically towards creative generation and human interaction.
However, a recent paper entitled “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)” raises a crucial question: as we delegate more and more creative tasks to machines, are we witnessing a dangerous flattening of the diversity of thought?
The Evolution of the Discipline: From Facts to Open-Ended Creativity
The evolution of the LLM It is characterized by a race for "correctness." Traditional benchmarks are designed to evaluate how accurately a model answers factual questions or executes rigid instructions. However, as highlighted in the paper, this metric is no longer sufficient. The current frontier is represented by open-ended tasks, where there is no single correct answer, but rather a spectrum of valid possibilities.
The discipline faces a paradox: models have become technically excellent, but they struggle to replicate the intrinsic variability of human expression. When a human thinks of a story or a metaphor, they draw on a unique set of experiences; when an AI does so, the risk is that it draws on a statistical average that eliminates nuance. The research under consideration fills precisely this methodological gap, shifting the focus from quality (is the text grammatically correct?) to diversity (is the text original or a copy of a thousand others?).
The discovery of the “Artificial Hivemind”
The heart of the analysis lies in the presentation of INFINITY-CHAT, a large-scale dataset composed of 26.000 real queriesThe authors used this tool to investigate a phenomenon they called “Artificial Hivemind”.
The results are surprising and, in some ways, alarming. The study identifies two levels of homogeneity:
- Intra-model repetition: the same model, questioned several times on the same open question, tends to recycle the same structures and concepts, simulating a creativity that it does not actually possess.
- Inter-model homogeneity: This is the most critical fact. Models developed by different companies, with different architectures and sizes (from Llama to GPT-4, from Claude to Qwen), converge independently towards the exact same answers.
A striking example cited in the paper concerns the request to write a metaphor about time. The vast majority of models, regardless of their origin, produced variations on the concept "time is a river" or "time is a weaver." This convergence demonstrates that the alignment of algorithms, often based on the same principles of Reinforcement learning from human feedback, is creating an artificial “single thought”.
Conclusions and perspectives
The paper's analysis highlights a critical issue: current evaluation systems tend to reward "safe" and consensual responses, penalizing the true diversity that characterizes idiosyncratic human preferences.
If the discipline's goal is to create assistants that amplify human creativity rather than replace it with standardized surrogates, training paradigms need to be rethought. Otherwise, the risk is long-term cultural impoverishment, where repeated exposure to homogeneous outputs reduces our very ability to think outside the box.Artificial Hivemind It is not just a technical problem, but a challenge for the protection of intellectual plurality.
