Why Do All AI Models Drift Toward the Same Voice?

This is not a Claude quirk you can prompt your way out of. It is structural, and it has a name.

Models increasingly train on text that earlier models produced. Generation after generation, the rare, idiosyncratic, distinctly-human edges of the writing get sampled out, until output converges on a smooth, uniform middle.

The Oxford and Cambridge team that formalized this published in Nature in 2024: training on recursively generated data makes "tails of the original content distribution disappear" and the result collapse toward a single, averaged mode (Shumailov et al., Nature).

A Nature commentary on the study traces the trajectory: feed models too much AI-generated data and their output first flattens into sameness, then, over enough generations, degrades into gibberish (Wenger, Nature). LinkedIn sits at the early stage of that curve, the flattening, not the gibberish.

That loop is already running on LinkedIn, and you can feed it without noticing.

You stop on a post you admire, which may itself have been written by AI. You ask a model for "something similar," and that model was trained on posts already calibrated to sound like AI. You publish, and your post becomes one more example the next reader copies and the next model learns from. Each pass adds another coat of the same gloss.

That averaged middle is, in practice, what LinkedIn calls slop: polished on the surface, with no real perspective underneath. It is also why these few patterns now show up everywhere at once, the fuller story of how AI writing took over the LinkedIn feed and where the patterns came from.

It is also why heavy prompt engineering or bolt-on "skills" only move the needle so far. You can push Claude off its defaults for one draft. But the gravity always pulls back toward the mean, because the mean is what it was trained to produce.

There is a flip side worth keeping in view. The more the feed fills with averaged AI prose, the rarer a genuinely original human post becomes, and the more it stands out. In 2026, reading carefully, thinking for yourself, and writing in a voice no model could average into the middle is not just good manners. It is the edge.