In 2025, Hachette pulled a horror novel from shelves. Not because of bad sales or a lawsuit, readers on Reddit figured out something the publisher’s own editors missed.
*Shy Girl* by Mia Ballard had 4,900 ratings on Goodreads, a loyal BookTok following, and a deal with one of the Big Five publishers. Then a YouTuber named Frankie’s Shelf posted a nearly three-hour breakdown of the text. The evidence was damning. The word “sharp” appeared 159 times in 214 pages, used to describe everything from light to sounds to abstract thoughts. “Heavy” showed up 74 times. “Weight” 94 times. “Edge” 84 times. The book leaned on the “rule of three” pattern that language models default to, and nearly every chapter opened with the protagonist waking up, as if each section had been prompted separately. A million views later, the New York Times got involved. AI detection firm Pangram ran their analysis and Hachette cancelled the US release, making this the first time a major publisher has ever walked back a book over AI accusations.
Ballard says she didn’t use AI. She claims a freelance editor introduced it without her knowledge. Maybe that’s true but this is what stuck with me: readers caught it before the professionals did. People are developing an instinct for when something doesn’t sound like a person wrote it.
The Web Is Starting to Sound the Same
I spend most of my working hours inside content audits, and I’ve noticed something that’s hard to unsee once you spot it. Open ten articles about any SEO topic (AI content strategy, E-E-A-T, algorithm updates) and they read like the same person wrote all of them. Identical structure, recycled transitions and copy-paste vocabulary.
So I built a tool to test if that feeling was real.
I put together a content homogeneity analysis at Pantelope. I pulled 12 top-ranking competitor pages on this topic and ran each one against 12 pattern-detection rules derived from peer-reviewed linguistics research.
- Buzzword density
- Sentence length variation (what researchers call “burstiness”)
- Passive voice percentage
- First-person pronoun usage
- Hedging phrases.
This is what separates writing with a personality from writing that just… fills space.
I ran it against 12 top-ranking articles about AI content and SEO. The results were revealing but not surprising:
- “Actionable” appeared in 5 out of 12 pieces.
- “Comprehensive” showed up in 5 out of 12.
- “Leverage” in 3.
- “Landscape” in 3.
- 7 buzzwords appeared across multiple pieces, out of only 16 unique buzzwords total.
- Almost none of these writers used first-person pronouns in any meaningful way. The average first-person density was under 5 per 1,000 words; nobody was saying “I tested this” or “we found that” or “what actually happened when we tried it.”
What got me was that every one of those pieces was about authentic writing in the age of AI, and they all sounded identical.
This lines up with what the research is showing at a broader level. A Cornell study by Dhruv Agarwal found that AI writing assistants homogenize users’ output toward Western, specifically American styles, diminishing cultural nuances for writers in the Global South. Indian users in the study didn’t get the same productivity benefit because they kept having to correct culturally inappropriate suggestions.
A Princeton study published in March 2026 by Yu Xie and Yueqi Xie went further. They showed that generative AI models are inherently prone to what they call “regression toward the mean,” where the variance in output shrinks compared to real-world distributions. AI produces relentlessly average output and average at scale, drowns out everything else.
Meltwater’s data backs this up: mentions of “AI slop” across the internet increased ninefold from 2024 to 2025. Merriam-Webster named “slop” the 2025 Word of the Year. Not a niche concern but definitely a cultural shift in how people talk about what they’re reading online.
Google Answer: Experience Over Everything
Google noticed too and they’re not being subtle about it.
March 2026 hit hard. SEMrush Sensor recorded a 9.5 out of 10 volatility score, among the highest they’ve ever seen. Over 55% of monitored websites experienced ranking shifts in the first two weeks. A separate spam update landed just two days before the core update, creating what tracking tools described as the highest SERP volatility of 2026 so far.
But most coverage misses the nuance: Google is NOT anti-AI. They’re anti-garbage content.
Danny Sullivan has said repeatedly that they evaluate content based on whether it helps searchers, not based on how it was produced. The March update deployed what analysts believe is a Gemini 4.0 Semantic Filter to identify content produced at scale without meaningful human editorial oversight.
- Sites using AI as a drafting tool, where humans add real expertise and editorial judgment, performed fine.
- Sites that used AI as a content factory got hit.
How do numbers back this up? A 16-month study by Digital Applied tracked 4,200 pieces across 140 domains. Pure AI output ranked 23% lower on average than what people wrote. But AI-assisted work, where a person edited, added original data, and brought a genuine perspective, performed within 4% of fully person-written pages. That’s basically noise.
Semrush owns a study of 20,000 keywords and 42,000 blog posts, finding that position 1 results are 8x more likely to be human-written. But 87% of SEO teams still keep humans directly involved in production. The most common workflow? Human-led, AI-assisted. 64% of teams work this way.
One thing nobody mentions is backlinks. Digital Applied found AI-only pages acquired 61% fewer editorial backlinks than what people wrote on comparable topics. Nobody wants to link to something that reads like a summary of the top five search results. Links still matter and they flow toward originality.
Google’s Helpful Content system isn’t a standalone update anymore. As of 2026, it’s integrated directly into the core ranking. One weak page can drag down your entire site if the pattern is widespread enough. What Google’s evaluating now is if your pages show genuine knowledge through lived experience, or just restate what already exists in a slightly different order.
The Authorship Question Is Older Than You Think
Most people treat this like a new debate and we’ve been arguing about it for over a hundred years.
A Fortune article from March 2026 drew a line I hadn’t considered. In the late 19th century, a sculptor went to court because someone accused his assistant of completing his work. The judge ruled that an artist could accept a certain amount of mechanical assistance, but there was a threshold when that assistance became “dishonest”. He made the sculptor create a bust in real time to prove his skill.
Now imagine every content creator facing that same test. Except the judge is an algorithm.
Legally, things are getting clearer. In Thaler v. Perlmutter, the D.C. Circuit ruled unanimously that AI systems can’t be named as the author of a copyrighted work. Supreme Court denied review in March 2026. Settled. But AI-assisted works are protectable. The U.S. Copyright Office’s January 2025 report draws a line between AI-generated (no protection) and AI-assisted (protectable if you exercised meaningful creative control). Simply entering prompts is not enough. You have to edit, arrange, select, or substantially modify what comes out.
In academia, 88% of students now use generative AI for assessments. A blind test at a UK university found that 94% of fully AI-written exam answers passed undetected and actually scored higher than real student submissions. What’s emerging across institutions is a distinction between AI as an “assistive” tool versus a “substitutive” tool (replacing the writer entirely). That distinction matters for content marketing too.
Fortune ghostwriting parallel stuck with me because it reframes the whole debate. Ghostwriting has existed since 1908; editors have always shaped books, television shows get written in collaborative rooms. Nobody questioned whether assistance was acceptable. What they questioned was how much assistance you can accept while still calling the work yours, and if you disclose it.
Generative AI, as one ghostwriter told Fortune, is “the ghostwriter for the masses”. But there’s a difference. A ghostwriter brings their own skill and judgment to the collaboration, since a language model brings statistical patterns and what you get out depends entirely on what you put in and what you do with it.
What Actually Differentiates Now
ANA (the Association of National Advertisers) chose two Words of the Year for 2026: “authenticity” and “agentic AI”. That tension between the demand for genuine connection and the pressure to automate everything is where every marketing team is operating right now.
CXL published a piece in December 2025 that nailed something I’d been feeling but couldn’t articulate: they described how AI creates a feedback loop that slowly kills brand voice. Each AI draft becomes the new baseline. Writers NEED to adjust to match it. Editors approve it because it’s “good enough”. New hires see AI drafts getting approved and assume that’s the standard. Over time, nobody on the team can explain what makes the brand sound like itself anymore because the muscle memory of good writing weakens from disuse.
When we ran our homogeneity analysis, this is exactly what we saw reflected in the data. 12 pieces about the same topic, and almost none of them had a distinguishable personality. High burstiness scores (which technically read as “person-written”) but are functionally interchangeable. You could swap the bylines and no reader would notice.
And that’s what worries me. Not that AI will replace writers. The actual fear is that Al will make every writer sound the same.
Fair counter-argument worth hearing. A CMSWire piece from April 2026 argued that AI might actually be the best consistency tool available if you invest in teaching it your patterns properly. Brand identity was already diluting before AI arrived, they pointed out. Every new hire, every agency, every regional office introduces drift. When AI is properly trained on documented style guidelines, it can enforce consistency at scale in ways teams couldn’t on their own.
I think both things are true. AI amplifies whatever you feed it. If you feed it with generic prompts, you’ll get generic output. Feed it with deeply documented brand patterns with proper editorial oversight, and it can scale your identity without flattening it. But that second path requires serious work that most teams don’t take.
