You Don't Sound Like You Anymore.

Sofia got married last month.

Six weeks later, she pulled me aside to tell me she used Claude to write her vows.

She had been too nervous to write them cold. She opened a document three separate nights in June, stared at it, closed it, and went to bed. On the fourth night, she opened Claude instead. She typed a few paragraphs about her fiancé, how they met, what she wanted the marriage to be. The model gave her back four paragraphs of vows. She kept about eighty percent of what it wrote. She stood in front of her family and read it.

Then, six weeks later, she reread the printed page and could not find herself anywhere in it.

The vows were lovely. Grammatically fine. Emotionally correct. Every word in the right place. And they were not hers. She had stood in front of the two hundred people who loved her most and made a promise using sentences she did not build.

The part that was breaking her was not the guilt.

It was that she could no longer remember what she would have written herself.

 

You do not sound like you anymore. And you are not going to notice until the day it matters most.

 

Where your voice actually went

I want to be careful with this letter. This is not a lecture about authenticity. Authenticity is a word people use when they want to sell you a course. What I am about to describe is more specific and more mechanical than that.

Voice is a muscle. It is trained by use. When you stop using it, it atrophies, exactly the way the deep-work muscle atrophies for a coder who has become a manager of slot machines. (Letter 89, Sofia's other confession. Same person. The theme keeps arriving at her door.)

Voice loss is not about corporate writing. It is about your capacity for first-person thought. That is the whole letter, so let me say it again, cleanly. When a model rounds the edges off enough sentences you would have written yourself, you stop being the person who could have written them. Not "you sound generic." Deeper than that. You lose access to the specific line of thinking that only you could have followed. The thought was yours. The pathway that leads to the thought is worn by repetition. The repetition is gone. In its place, a smooth model output that sounds like a thousand other model outputs and does not require you to have thought anything at all.

Here is the part that should scare anyone who writes for a living, and by "for a living" I mean anyone whose job includes an email.

This did not start with AI. AI is the last mile.

 

The seventy-year flattening

Read a business letter from 1948. Then a McKinsey memo from 1985. Then a TED talk transcript from 2012. Then the top three posts on LinkedIn this morning.

The distance between the 1948 letter and the 1985 memo is enormous. Different verbs, different rhythms, different assumptions about what a professional sentence sounds like. The distance between the 1985 memo and the 2012 TED talk is smaller. The distance between the TED talk and the LinkedIn post is nearly zero. Same three-word openings. Same soft imperatives. Same performative vulnerability. Same em-dashed reveal. Same "P.S." at the bottom.

Professional prose has been converging for seventy years. It has a history and it has villains who mostly meant well.

Ben Yagoda has documented this arc more carefully than anyone. The Chicago Manual of Style. The AP Stylebook. Strunk and White's Elements of Style. Each was, in its moment, a well-meant intervention. Prune the excess. Prefer the short word. Cut the passive. They taught a common tongue so professionals from different regions and industries could communicate at scale.

The intended effect was legibility. The side effect, compounded over decades, was flattening.

By the 1980s there was a McKinsey voice. Legible in one paragraph. By the 2000s there was a TED voice. Legible in the first eight seconds. By the 2010s there was a LinkedIn voice. Legible in the first three words. Each new medium came with a house style. Each house style rewarded a specific register with reach and status. Each writer entering the medium, consciously or otherwise, adjusted their sentences toward the register that got rewarded.

This is not a conspiracy but a market. The registers that traveled won. The registers that did not quietly died in the desk drawers of writers who never got promoted for sounding like themselves.

By 2020, most people who wrote professionally in English were operating inside four or five registers. Consulting. Journalism. Academic. Tech-founder-blog. LinkedIn. A knowledge worker could switch between them like changing shirts.

Voice, as a distinct property of a specific human, had already been demoted from a requirement to a garnish.

Then 2022 arrived.

The models were trained on the internet. Which is to say, on seventy years of the flattening. The registers that had already won the market got trained into the weights at industrial density. The models did not invent the professional voice, but they inherited it, then perfected it, then offered to write in it for free.

Steven Pinker argued in The Sense of Style that clear prose is a moral project. He was right. The models learned Pinker's rules and now enforce them at a scale he never imagined. Ask a model to write anything and it will hand you back the median of the median of every professional register it was trained on. That output is not "bad." It is technically fluent. It is the endpoint of the seventy-year flattening, delivered instantly, at zero marginal cost, to anyone who asks.

 

What the research is starting to show

The flattening is now being measured.

In 2025, Nataliya Kosmyna and her collaborators at MIT Media Lab published a study titled Your Brain on ChatGPT. They tracked brain activity, essay content, and memory in participants writing essays under three conditions: with ChatGPT, with a search engine, or with nothing but themselves. The ChatGPT group showed the lowest neural connectivity in the brain regions associated with language production. Their essays clustered on the same vocabulary. Human evaluators, blind to which group had produced which, rated the ChatGPT essays "soulless." Weeks later, the ChatGPT writers could not accurately recall what they had written. Their own sentences had passed through them without leaving a mark.

Read that last sentence again: the writers could not recall their own writing, because it was never their writing.

Other studies are finding similar shapes. Reduced lexical diversity. Sentence-length distributions collapsing toward a common mean. What linguists are starting to call stylometric convergence: distinct writers producing measurably more similar prose the longer they use the same model.

 

The wedding-vows scene as limit condition

Sofia is a senior engineer. She writes professionally. RFCs, code review comments, Slack decisions, quarterly planning docs, migration post-mortems. Her professional writing was already trained inside the engineering register, which is itself a flattened dialect optimized for scanning, clarity, and legibility to a stranger who needs to understand the decision fast.

When she sat down to write the most personal sentences of her life, the muscle she had for personal voice was already thin. It had been fed the engineering register for a decade. It had been fed the LinkedIn register for another decade. The model completed the arc. It offered her a set of vows in the median register of "wedding vows on the internet."

That register is loving, cadenced, mildly funny, structurally reassuring. It has no Sofia in it. There was never going to be a Sofia in it. The model has never met her.

She read those vows in front of two hundred people. Half of whom had known her for thirty years. Not one of them said, at the reception, "those didn't sound like you." Not because they were being polite. Because the register of "wedding vows" has drifted far enough from the register of "how Sofia actually talks" that the two do not need to overlap for the vows to sound acceptable at a wedding.

That is the horror. Nobody notices. Because everyone's ear has been trained on the same flattened dialect. You are speaking to a room that no longer expects to hear you.

 

The uncomfortable truth

Voice is not decoration. Voice is what your first-person thought sounds like.

If your voice is gone, the thought is gone with it. Not softly gone. Structurally gone. You did not think a thing you couldn't write. You thought approximately the thing the register let you write, then you convinced yourself, in the writing, that this was what you meant all along.

Every time you take a model's suggestion for a sentence you could have written yourself, you are training your brain to expect the completion next time.

Your brain stops holding the whole thought and reaches for the assist. The thought never quite lands. You send the email. You move on. The next thought is slightly less yours. The next one, slightly less again. Over a year of this, the interior monologue that used to be uniquely yours becomes a set of prompts your brain writes for a model that finishes them.

(That's me in 2026.)

At the extreme, and Sofia is describing the extreme, you lose the ability to recall what you would have said. The pathway is not just under-used. It is overgrown. The line you would have written in 2019, cold, in your own kitchen, with no model in the room, is no longer accessible to you. You cannot find it. It is not there.

Plausible and correct are not the same thing. Plausible and yours are not the same thing either.

AI is an amplifier, not an equalizer. It amplifies the writer who already knows what she sounds like. It flattens the writer who does not.

If you're not a "professional writer," you might think: I don't care.

But your voice isn't a luxury. It's how other people tell there's a real human behind the words.

So if we all end up sounding the same, and sounding exactly like the free machines, ask yourself this:

What's your value in a world where almost everything you produce (mostly writing and articulation) can be generated in a couple of seconds?

 

Keeping YOU alive

Ok, so what do we do about it?

The sharpest people I know are getting sharper with AI. Everyone else is quietly getting duller. I've been trying to figure out the difference.

The best practices I've found might sound cliché or obvious, but that's exactly the point. It's the difference between "I can get everything you just said by prompting a chat window" and "this person has a real edge."

All roads seem to converge to one thing:

One day with zero AI.

Brain gym and a weekly proof of life. Evidence that there is still someone behind the words.

I never thought I would write this one day. I've started this practice for a while and the feeling is strange (exactly like the gym). You feel struggle all day long and you feel sharper and UNSTOPPABLE by the end of the day.

 

Back to Sofia

I did not have a clever answer for Sofia that night. I stood there for a minute before I said anything.

I said the vows she read were not the vows she meant. That was true and it was the part she needed a friend to name. I said the vows she would have written were still, somewhere, in her. Under a decade of engineering docs and LinkedIn posts and Slack messages, but still in her. I said the way you find them again is not by trying to remember them. It is by writing anything, everything, for ninety days, without help. Texts. Notes. A journal by the bed. Emails you would normally polish. Send them ugly. Send them raw. Do not let the model finish your sentences. Not once. For ninety days.

I told her that if she did that, on some Tuesday around week eight, she would write a sentence to her husband and know, in her chest before her head, that no model on earth could have written it. That would be the sentence that told her the muscle was back.

She did not say anything for a while.

Then she said, "That's what the vows were supposed to be."

I nodded. It was.

She said she was going to try. She was going to write her husband a letter on paper, by hand, for their six-month anniversary. No AI in the room. She said she was terrified.

I told her that was the muscle waking up. That the terror was the entire signal.

I don't have a better system than this. One day a week with no assist. If you have one, reply and tell me.

AI did not take your voice. You gave it away one email at a time.

The good news is you can take it back the same way.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)