Gravity of the Generic: Why Distinctiveness is Now a Cost

Augmentation is the common description. But it is misleading.

The human is framed as being augmented by the system.

The observable dynamic points to the opposite.

The model is extended by the human.

After a thousand drafts, do you perceive that the output of a document is closer to your personal unique style, or the LLM's?

LLMs pull toward the statistically likely centre with black hole level force. Significant human intervention moves the output briefly toward the creative edge, but only temporarily. The LLM has the stronger force.

I'm calling this force the gravity of the generic.

Probability becomes plausibility

LLM output is not only usable but often credible and at times brilliant. But despite this LLMs orientate to the statistical middle.

Rather than a design flaw, it is commercially viable and also the nature of a language model built on a transformer architecture.

If you want people to use your LLM then the centre is where to aim it. Confidence is highest, output sounds most assured and the fewest people object.

Readers are getting sharper at detecting AI-shaped prose, even when the underlying content is competent. Not because the output lacks quality, but because the output lacks origin. It is quietly built from patterns that belong to everyone and therefore belong to no one. The phrasing is authoritative. The structure is considered. The reasoning is balanced.

What the research is finding

When Italy banned ChatGPT in April 2023, a team of researchers studied what happened to the content produced without it. Marketing output from restaurants in Milan became 15% more lexically diverse and 12% more syntactically diverse. Consumer engagement rose by 3.5%.

A separate study tracked what persists once ChatGPT is no longer available. The performance lift falls away quickly but the homogenising effect does not unwind in the same way. The tendency toward similarity keeps rising even months later, and the researchers describe that persistence as a "creative scar". A drift toward the model's centre line can become a new baseline that outlasts the session and outlasts the tool.

A cross-disciplinary review published in August 2025 pulls together evidence from linguistics, cognitive science, and computer science around one recurring effect: when many people lean on the same generative systems, expression converges. More outputs land nearer the centre, and fewer outliers carry local voice, local framing, or culturally specific texture. Perceived quality and similarity can rise together, which helps explain why the drift is hard to notice inside organisations.

A CHI 2025 paper tested a Western-centric writing assistant with 118 participants from India and the United States. AI suggestions shifted writing style across structure, tone, and rhetorical moves, with Indian participants' writing moving toward Western stylistic norms and cultural nuance reduced toward the model's centre line.

Why this needs naming

In August 2024, Amazon CEO Andy Jassy described the internal impact of Amazon Q, their AI coding assistant:

"The average time to upgrade an application to Java 17 plummeted from what's typically 50 developer-days to just a few hours. We estimate this has saved us the equivalent of 4,500 developer-years of work."

And with that, augmentation returns. Augmenting the machine is faster, keeps the approval chain calm, and improves this quarter's numbers. The cognitive labour involved in resisting that pull, which I examined in a companion essay 'Human Middleware', goes unmeasured.

The vocabulary for objection to generic writing does not exist in most approval processes. Errors, compliance risk, and missing data are discussed, but whether the output is indistinguishable from what a competitor would produce using the same tools is not usually a part of any review process.

Distinctive output forces someone in the review chain to take a position and defend the choice to be different. And it takes time.

Competing companies using the same models, trained on similar data, prompted in similar ways, will arrive at similar outputs. The convergence is not only internal. And the organisations most exposed to it are the ones least likely to detect it, because they are tracking efficiency and adoption, not drift toward the mean.

How many companies created a baseline for what their voice sounded like before LLMs, and are analysing how far it has moved since.

The absence of measurement does not indicate the absence of effect. It indicates the absence of attention.

And if the gravitational pull of LLMs has a permanent reshaping effect on working patterns, then the choice of where to embed these systems may carry consequences that outlast the deployment. I examined the broader cognitive labour involved in managing AI output in a companion essay, 'Human Middleware'.

Allowing the Gravity of the Generic to move an organisation to the centre does not require a decision. Only the absence of one.