The Real Problem With AI Content

Imagine the whole planet hired one single writer to produce all of its content. That's more or less where we are.

You can generate content today without a human touching it, and it can rank — in AI search, sometimes even in Google. With a proper pipeline and good skills, pure AI content gets results.

The catch: it only works while nobody else is doing the same thing. Thousands of people are doing the same thing.

The thousand-people problem

You want a piece on content marketing ideas for B2B SaaS. You type the prompt, take what comes out, publish. Thousands of people did exactly that this week and published roughly the same article.

That's why content "isn't getting crawled" and "isn't getting indexed."

Google's spam policy targets unoriginal content at scale, "regardless of how it's created." It doesn't punish AI content. It punishes sameness.

The 2024 Content Warehouse leak points the same way: the API docs contain an OriginalContentScore attribute that grades content on originality. A leaked attribute isn't proof of a live ranking signal — but the policy and the leak agree. Originality is the axis Google measures.

AI writes like one person

The best AI detector today is Pangram — near-perfect in independent studies from the University of Chicago and the University of Maryland, with essentially zero false positives.

The accuracy isn't the interesting part. The interesting part is why detection works at all.

For every human document in its training set, Pangram generates an AI mirror — matched on topic, tone, and style — so the classifier can only learn the tells of LLM writing itself. It works, which proves the tells exist: AI has one recognizable style.

Where does that style come from? Pangram's Joe Stech calls it the Annotator Consensus Dialect. Pre-training gives a model the full range of human writing; alignment crushes it. Idiosyncratic writing gets 5/5 from one human rater and 2/5 from another. A hedged, symmetrical answer gets a safe 4/5 from everyone. Optimization collapses the variance into one smoothed voice:

the conversational equivalent of hotel lobby decor

Academic research confirms it: RLHF causes mode collapse that persists across prompts, and stylistic diversity is already declining on Reddit and in scientific writing.

So when you and I both ask an LLM to write, we publish the same person's writing.

That's also why generic content can't win in AI assistants — a model has no reason to cite a page it could generate itself. In Ahrefs' analysis of 1.4 million prompts, original data and verifiable statistics were among the strongest predictors of getting cited. Models cite what they can't produce.

And skip the humanizers. In the Chicago tests, Pangram caught humanized text while other detectors fell apart. Even one that worked would only change how the text sounds — it can't add information you never put in.

What to do instead

There are only two problems to solve:

1. The writing isn't unique. By default, an LLM writes like the one writer the planet hired.
2. The research isn't unique. Ask an LLM to research a topic and it runs the same searches, reads the same sources, and returns the same summary it gives everyone who asks.

None of what follows is a formula. Do these things and you don't get identical results — analysis of 2,000+ domains after Google's March 2026 core update found the sites that gained visibility weren't the ones checking tactic boxes. They were sites that owned something structural — proprietary data, a real brand, a job the page actually let you finish. Sites that just aggregated or commented on other people's content lost ground regardless of effort. The list below is what "owning something" looks like in practice, not a checklist that guarantees it.

Fix the writing

Dictate, don't prompt. Talk through what you think, let the model clean it up after. This article started as voice notes. A model can shape your thinking — it shouldn't originate it.

Write your voice down. A brand-voice doc with real examples from your own writing, checked against every draft. Models can't invent a voice, but they can hold one you hand them.

Track your own sameness. Sentence-length variance and lexical diversity are measurable. Run them across your last twenty posts — if everything clusters around the same rhythm, you've collapsed into one voice too.

Fix the research

Map the field first. Read what already ranks and name, specifically, what your piece adds. Can't name it? Don't publish it.

Publish your own numbers. Original data is the one input a model can't generate. The GEO study measured up to 40% more visibility in generative engines from adding statistics. A number nobody else has is the cheapest real differentiation there is.

Build research agents that skip the default path. Don't ask one model for "research on X" — that's everyone's research. Send agents at the topic from several specific angles, force them into primary sources (Exa, Parallel Web), and make them report what the top-ranking pieces don't say.

Epilogue

AI didn't kill content. It repriced it: sameness now costs nothing and is worth nothing. Uniqueness is the only thing left with a price.

Using AI to write was always a means to an end. The end is knowing something nobody else does. That was always the job — AI just stopped letting us pretend otherwise.

Notes

• Pangram. AI content detector, self-reported ~1-in-10,000 false-positive rate – pangram.com
• Jabarian, Sina, and Alex Imas. "Artificial Writing and Automated Detection." Becker Friedman Institute Working Paper, 2025 – bfi.uchicago.edu
• Russell, Jenna, Marzena Karpinska, and Mohit Iyyer. "People Who Frequently Use ChatGPT for Writing Tasks Are Accurate and Robust Detectors of AI-Generated Text." ACL 2025 – arxiv.org/abs/2501.15654
• Pangram Labs. "How AI Detection Works" – pangram.com/research
• Pangram Labs. "Why Perplexity and Burstiness Fail to Detect AI" – pangram.com/blog
• Stech, Joe. "The Information Theory Behind Why AI Writing Sucks." Pangram Labs, 2026 – pangram.com/blog
• "Understanding the Effects of RLHF on LLM Generalisation and Diversity." arXiv 2310.06452 – arxiv.org/abs/2310.06452
• "The Homogenizing Effect of Large Language Models on Human Expression and Thought." arXiv 2508.01491 – arxiv.org/html/2508.01491v1
• "Homogenizing Effect of Large Language Models on Creative Diversity." ScienceDirect, 2025 – sciencedirect.com
• Google. "Spam Policies for Google Web Search" – developers.google.com
• King, Mike. "Inside the Google Algorithm Leak." iPullRank, 2024 – ipullrank.com
• Ahrefs. "Why ChatGPT Cites the Pages It Cites" – ahrefs.com/blog
• Aggarwal, Pranjal, et al. "GEO: Generative Engine Optimization." arXiv 2311.09735 – arxiv.org/abs/2311.09735
• "Google's March 2026 Core Update Shifted Visibility Away From Aggregators." Search Engine Journal – searchenginejournal.com
• Walter Writes AI – walterwrites.ai
• Exa – exa.ai
• Parallel Web – parallel.ai