Researchers gave the same prompt to a human writer and five AI models. 61,608 stories later, the machines had quietly agreed on something, and the humans had not.
The study is a simple one in conception, ambitious in scale. Writers, human and artificial, were given the same creative prompts and asked to produce short fiction. The researchers then measured narrative rarity: how far each story departed from the distribution of everything else. A rare story is one that does not look like the others. A common story sits in the centre of the distribution, doing what most stories do.
The finding is striking in its clarity. GPT, Claude, Gemini, DeepSeek, Kimi, from different companies with different architectures and different training approaches, all drifted into the same narrow corner of narrative space. Their rarity distributions overlapped. Their stories were competent, fluent, structurally sound, and similarly shaped: tidy plots, themes spelled out, morals resolved. The machines had converged.
The humans had not. Human stories were rarer, stranger, more willing to leave a question open. On narrative structure alone, you could separate human from machine with 93% accuracy.
What convergence means
The convergence of AI systems on similar narrative territory is not a training failure. It is what these systems are designed to produce. Large language models are, at their core, next-token prediction engines optimised on the statistical centre of human expression. They are extraordinarily good at the middle of the distribution, at producing the most probable continuation of what has come before. That is their strength. It is also their structural condition.
The middle of the distribution is not where originality lives. It is where originality goes when it has been averaged across all the other originality that ever existed and reduced to the mean. A model trained on the full range of human creative expression, optimised for coherence and fluency, will converge toward the common, not because it is unintelligent, but because that is precisely what it is doing. It is the world's most sophisticated machine for producing the expected.
The machines had quietly agreed on something the humans had not.
The same dynamic in reasoning and decisions
The narrative rarity study measures convergence in creative writing. But the dynamic it captures, AI systems gravitating toward the statistical centre of prior human expression, is not specific to storytelling. It applies wherever AI is used to process and synthesise human thinking.
When an AI system summarises stakeholder input, it produces the most coherent account of what the inputs had in common. The outlier perspective is more likely to be smoothed away than amplified: the one that departed from the expected framing, the challenge that came from an unexpected direction, the insight rooted in lived experience rather than the documented baseline. Synthesis compresses toward the average.
This matters because the most valuable information in any deliberation is rarely in the centre. The assumption everyone shares is invisible until someone who does not share it points it out. The risk no one has named is nameless until someone on the edge of the conversation names it. The value collision underneath an apparently technical dispute is invisible until someone with a different stake in the outcome makes it visible.
The human tendency to scatter, to produce narratives that are rarer, stranger, harder to categorise, is not noise. It is signal. It is the characterisation diversity that makes deliberation valuable. And it is exactly what AI synthesis is structurally inclined to compress.
What being human-aware requires
A human-aware process does not treat human thinking as input to be processed. It treats human thinking as the thing to be preserved: in its diversity, its particularity, its genuine strangeness when strangeness is present.
This has specific architectural implications. A human-aware system cannot aggregate human contributions into a summary without losing what made them individually valuable. It needs to be able to hold the full distribution: to show where the centre is and what is at the edges, to distinguish the outlier perspective from the noise, to make the rare contribution as analytically available as the common one.
The SparkMap does this by preserving the relational structure of deliberation rather than collapsing it into summary. Each contribution retains its epistemic ground. The connections between contributions show how thinking developed. The tension between opposing perspectives is not resolved away but represented as information: here is where genuine disagreement exists, and here is why it matters. The result is a collective understanding that has not been flattened. The rare contributions are still visible.
The work that was always there
The study's finding, that on narrative structure alone you could separate human from machine with 93% accuracy, says something that has been true for longer than AI has existed. Human thinking is individual. It scatters. It departs from the expected. The most valuable ideas, the most important challenges, the insights that change how a question is being asked, do not come from the centre of the distribution. They come from people who are willing to occupy the edge.
AI writes fluently as everyone at once as stated in the existing data. Competence has stopped being the moat. What AI cannot produce is the individual perspective: the one shaped by a specific life, a specific professional experience, a specific positional stake, a specific set of values encountered in a specific order. Being human-aware is how you preserve that. Not as a gesture toward humanist values. As a structural choice about what your deliberation is designed to keep.
*Source: Narrative rarity research across 61,608 AI- and human-authored stories (arxiv.org/pdf/2604.03136). Researchers demonstrated 93% accuracy in separating human from AI-authored stories on narrative structure alone, with human stories showing significantly higher rarity percentiles than outputs from GPT, Claude, Gemini, DeepSeek, and Kimi.
Hunome's SparkMap preserves the full distribution of collective thinking, including the edges, so that the rarest and most valuable contributions remain analytically available.*