By Ingrid Schneider, Founder — Train in Your Lane
My boyfriend runs a Silicon Valley AI club with roughly a billion members. Fine, not a billion. But it feels like a billion, and what they actually do in that club surprised me when I learned about it a few years ago —
They read research papers. (At first this sounded BOR-RING.)
Not newsletters about research papers. Not LinkedIn hot takes about research papers. The actual papers — the PDFs with the tiny superscript numbers and the citations that go on for nine pages. And somewhere along the way, that habit rubbed off on me. I’m now the person geeking out over arXiv links at dinner. At an Ingrid level, which means I skip the math and go straight for the part that matters to people who run businesses.
Here’s the one that got me this month.
Three researchers at USC — Sourati, Ziabari, and Dehghani — published a paper called “The Homogenizing Effect of Large Language Models on Human Expression and Thought.” It landed in Trends in Cognitive Sciences, which is not a blog. It’s peer-reviewed cognitive science.
Their finding, in plain English: AI is making us all sound the same. And worse, it’s starting to make us think the same.
What the paper actually says
The researchers pulled evidence from linguistics, psychology, cognitive science, and computer science, and the pattern shows up everywhere they looked.
When people write with AI, their writing converges. The word choices flatten. The metaphors drift toward the same defaults — and those defaults are Western, English-language, dominant-culture defaults, because that’s what the models were trained on. It’s not just style. The arguments people reach for start converging too. The way we reason.
Their phrase for what’s at risk is “cognitive diversity.” The variety in how people express themselves and work through problems. That variety isn’t decoration. It’s where creativity and collective intelligence actually come from. Twenty people who think differently will out-solve twenty people who think identically, every time.
And because we’re all leaning on the same three or four models, the convergence compounds. The models mirror the dominant patterns in their training data, we mirror the models, and the loop tightens.
Unchecked, millions of people writing and reasoning in the same voice. Theirs.
Why I can’t stop thinking about this
I run a marketing agency. I also run an AI training company. So I sit in the most uncomfortable seat in the room for this paper.
Because I see it. Every day.
I can spot an unedited AI draft from across the street now. The em dashes. The three-part lists. The “it’s not just X, it’s Y” construction. The paragraph that sounds smooth and quotable and says absolutely nothing you’d remember tomorrow. Scroll LinkedIn for four minutes and you’ll see a dozen posts written by a dozen different people in exactly one voice.
Which means the paper isn’t predicting something. It’s describing something that already happened to your feed.
And in franchising, this should scare people more than it does. Franchise development is a differentiation business. Your entire job is convincing a candidate that your brand is not the other 4,000 brands. If your discovery emails, your brochures, your posts, and your follow-ups are all generated from the same default voice as every competitor’s, you didn’t save time. You erased yourself.
The part everyone gets wrong
The lazy read of this paper is “AI bad, don’t use it.”
That’s not what it says, and it’s not what I teach.
The researchers’ warning is about unchecked use. Passive use. Prompt, paste, publish. The homogenization happens when the model’s voice replaces yours instead of serving yours.
This is exactly why the E.A.T. framework at Train in Your Lane starts with Education before Application. If you don’t understand what the tool does to your output, you can’t catch what it’s doing to your output. The teams I train who use AI well all share one habit: they treat the draft as raw material, not a deliverable. They put their own metaphors back in. Their own stories. Their own weird.
My dad owned waterbed stores. I was raised in a mall and burned my allowance on Ms. Pac-Man. No language model would ever generate that sentence for me, and that sentence is doing more work in this article than any polished paragraph could.
That’s the whole lesson, honestly.
What to do with this
Read the actual paper if you’re curious — it’s free on arXiv at arxiv.org/abs/2508.01491, and it’s more readable than you’d expect. Then go look at the last five things your brand published and ask one question: could a competitor’s name sit on top of this without anyone noticing?
If the answer is yes, you don’t have an AI problem.
You have a you problem.
And if the answer is yes and you’d like it to stop being yes, that starts with training the people writing it — not buying them another tool.
book a discovery call — free · 30 minutes
