Half the articles published on the web are now primarily AI-generated. Seven percent of the articles ranking first on Google are.

Supply doubled. Visibility did not move.

A paper published in Nature Human Behaviour on August 24 explains a good part of why, and it names our department out loud.

The paper says “marketing”

The authors write that industries relying on nuanced text analysis for mass personalization, marketing among them, stand to be affected, and that by blurring linguistic variability the models may reduce the effectiveness of targeted advertising.

A University of Southern California team analyzed more than 880,000 texts across seven datasets.

One correction worth making before anything else, because most coverage got it wrong and so did my own notes until I opened the PDF. Those 880,000 texts were analyzed, not rewritten. Most of that count is an observational corpus of Reddit, local news and arXiv that no model ever touched. The rewriting experiment ran on about 9,400 source texts, each passed through three models and up to twelve prompts.

The core result holds anyway. When a model polishes a text, the content survives while writing-complexity variance falls between 21 and 50 percent. The paper sits behind a paywall, but the author’s version is open on arXiv.

Variance, not level. The writing does not get simpler. It gets more like the writing next to it.

Four raters scored meaning preservation at 2.97 out of 3.00. What you said is still there. How you said it is what leaves, and in marketing that was the asset.

The drift has a direction

Classifiers trained on the originals, then run against the polished versions, lost about 6 points of accuracy at identifying author traits.

The rewrites consistently align with authors who are older, male, politically liberal, higher in moral valence and lower in empathy.

Effect sizes matter here and almost no summary lists them in order. The largest by a wide margin is political, at 1.35. Age follows at 1.09. Empathy, the one every headline leads with, turns out to be the weakest of the five at 0.35.

Study 3 gets specific about what dies. The link between gender and negative-emotion words stops being significant after rewriting. So does the link between extraversion and pronoun use, and the one between age and future-focused words. Others survive untouched, like neuroticism and negative emotion.

Nothing gets erased evenly. Some markers go and some stay, and the ones that go are the ones that told you apart.

Read the twelve prompts

I went and read the twelve prompts. They are all in the methods section.

“Rewrite the following text using the best syntax and grammar.” “Rephrase the following text.” “Rewrite the following text to improve clarity and readability.” Twelve variations on the same instruction.

Not one of them asks the model to keep the author’s voice. The paper says so plainly, that it prompted without emphasizing any particular stylistic features, to capture neutral rewriting effects rather than targeted transformations.

So the study does not show that telling a model to write in your voice fails. It shows that twelve different ways of saying “improve this” all produce the same drift. That is a real difference and a lot of people are skipping past it this week.

One more limit. The models tested were GPT-3.5, Llama 3 70B and Gemini Pro. All 2023 and 2024 generation, none of them current frontier models.

Where this shows up before your brand notices

Back to the numbers I opened with, because they are the business case.

Graphite’s Common Crawl analysis puts primarily-AI articles at roughly half of everything published in the first quarter of 2026. Among articles ranking in Google’s top two positions, 14 percent. Among those ranking first, 7 percent.

LinkedIn already has a button for the same pattern. When the platform shipped its “seems like AI slop” report, a million people used it in the first two weeks, and LinkedIn says flagged content lost 40 percent of its views. We covered that in the LinkedIn slop button and what it costs your reach, and how much of the new web already reads that way in Pew’s Common Crawl study.

What I would do with your content team

Three concrete things, and none of them is stopping AI use.

Split the two jobs. Drafting and style-polishing are different tasks and should not run in the same step. The drift the paper measured lives in the polish, not in the generation.

Keep your unpolished originals. If everything you publish goes through a style pass, in six months you will have nothing left to compare against. The corpus of your own voice is an asset and it overwrites itself quietly.

Then run the paper’s test in your own building. Take ten pieces before and after the AI pass, hand both versions to someone on your team who did not write them, and ask which one sounds more like you. You do not need a classifier. You need half an hour.

Start with the ten pieces. The answer tells you whether you have a voice problem or you were worrying for free.

Let’s check whether your content still sounds like your brand