it's not x, it's y

Today’s post is on reading content and just assuming it’s written by AI.

Even if it’s not (like this newsletter for instance)

I’m sure you’ve read LinkedIn posts, seen Instagram reels, and paused and been like this has to be AI.

A similar experience led me to write this.

Last week, one of my students sent me a LinkedIn post he had written to review about his recent fundraise. 

He said he’d written it himself, claimed no AI was involved, and just wanted a second pair of eyes.

I read it twice and it was a fine post, but also, line by line, the kind of post ChatGPT writes.

There were evident tells with the em-dashes, rule of threes, and the "it's not about X, it's about Y" patterns. We all know them by now.

He hadn't used AI, and was sure of it.

I believe him.

That is what makes this the strangest thing I have watched happen on the internet in years.

We have always done this in some form.

Anyone who has watched two seasons of Shark Tank India will notice that every pitch now sounds identical, in cadence and structure, without the contestants having ever met each other.

Picking up rhythm from whatever surrounds you in volume is one of the oldest features of how language develops and communication works.

What is different this time is that the style we are absorbing does not have an author.

For two years now, professionals everywhere have been drafting and editing in ChatGPT, asking it to clean things up.

The output of the model has been entering our heads at volume.

And it has started coming back out as our own writing.

Words like delve, realm, meticulous, underscore.

Then they went looking for those words in human speech.

They went through 360,000 academic YouTube videos and 770,000 podcast episodes, recorded between 2020 and 2024.

Some of the top ‘GPT words’ saw annual usage growth of roughly 25–50%.

One of the researchers said the project started with himself. "I realized I was using ‘delve’ more. I wanted to see if this was happening not only to me but to other people."

It was, indeed, happening to many others.

The same effect shows up in academic writing.

It’s interesting to note that some of them being AI generated don’t get caught, a classic type 1 error in academic publishing.

Languages shift, but they do so across decades, through migration and conquest and generations ( think GenZ language- I hear it's dope). 

They do not normally lose their fingerprint in eighteen months because a chatbot got popular.

The model did not coin the word "delve".

It absorbed it from a particular kind of human writing it was trained on, and then a particular kind of algorithmic reinforcement that came after.

In the absence of organic user feedback,  annotators are used to rate the model's outputs, the model learns from the ratings, and after enough cycles, it starts producing the kind of writing those annotators rated highly.

A significant share of that annotation work was outsourced to countries like Kenya and Nigeria, where delve is not an uncommon word at all.

So the voice coming out of the model is a kind of unintentional global English, stitched together from cheap labour in one part of the world and expensive product design in another.

And now that register is being exported back to American boardrooms, LinkedIn posts and founder pitches in Bangalore, where readers and listeners flag it as artificial.

When the YC founder Paul Graham tweeted in 2024 that anyone using delve in a cold email was probably using ChatGPT, Nigerian Twitter responded sharply.

The writer Elnathan John wrote back:

When those same writers ran their essays through ChatGPT to sound more native, the false-positive rate fell by half.

To be believed as human, you had to sound less like yourself.

We are already in the early phase of what researchers call model collapse.

When AI is trained on the internet, which itself is now somewhere between 35-50% AI-generated, the next generation of models is partly learning from the last.

In language terms, the future gets less weird, less specific, less regional.

The minority phrasings and local idioms and writers who only sound like themselves get smoothed first.

A Cornell study tested this with 118 writers, half Indian and half American.

AI suggestions helped the American writers more than the Indian ones, and they pushed the Indian writers' prose toward American English.

This includes the choice of words, structure, idiom, and things that made the writing recognisably Indian.

India also crossed 100 million weekly ChatGPT users recently, second only to the United States.

At exactly the moment the Indian voice on the global internet was becoming legible, we are translating it out.

Three things, in roughly the order I would do them if I made a living writing.

The first is to learn what AI sounds like.

Most people cannot hear it yet because they have not been looking.

Once you spend an hour with the patterns, you start hearing them everywhere, including in things you wrote yesterday.

I have put together a short field guide of the most common ones with rewrites.

Reply to this email, and I will send it across.

The second is to change how you use AI when you write.

Most people give it a topic, ask for a draft, then edit.

This is the worst possible workflow because the model's rhythm enters at the foundation.

Use it earlier, for thinking through an argument or finding holes in a position, then write the prose yourself.

This is a game-changer, I have been using this to strengthen my academic arguments as well.

I get AI to be my harshest critic and write irrefutable arguments. 

Find one or two pieces of specificity in everything you publish that the model could not have invented.

Like the name of a colleague who said something useful, a number from a deal you closed, or a failure with enough detail that nobody could have made it up.

The founders and creators who stay ahead from here will do it on details only they could have. 

Everything that is yours is the only moat.

Anything generic about your voice is a liability now.

Reply and tell me a word, phrase, or cadence that is uniquely yours that you have caught yourself editing out lately.

I want to know what we are all about to lose.

Nandini

PS: There is an old idea called mimesis which says that we end up sounding like what we absorb from whatever surrounds us in volume. For all of human history that was other humans. We are the first generation absorbing it from something that is not.