My Friend Sounds Like ChatGPT Now — and It's Not All Bad

There's a simple, pretty reliable way to tell a colleague has gone all-in on AI: they suddenly start writing better than you. Overnight, they've mastered a structure of thought they never once managed to demonstrate in three years of meetings. Where they used to write "so like, basically, we need to do this, let's do it like we talked about," now they send "I'd like to propose we fix three points: first, second, third," and you honestly don't know whether to be happy for them or worried.

Then on a call you ask, "Hey, so what was point two?"

The person hesitates for a second, then deflates right back down: "Yeah so, basically, the second thing is we need to, like, finish that piece, I'll write it all up properly later."

Science fiction likes to scare us with a recurring plot in which a person gains something valuable and, without noticing, loses the ability to realize they've already stopped being themselves. Writers using AI fall into the same scenario. The superpower is sudden fluency: the thought comes out coherent and persuasive right away, and you get for free what used to take hard, grinding effort. The price is the pause before the word: it’s in precisely those few seconds, when the thought is still raw and you have to grope your way toward it yourself, that a person finds out what they actually think.

Four Stages of Turning a Living Mind into an Artificial One

1. At the first step, you still carry the thought through to the end yourself. The idea, the order, and the whole structure of the text is yours. The LLM only comes in afterward: to smooth out the grammar a bit or to find a slightly more precise word, like an editor with an already-finished text. This is the safest stage. Which is not quite the same as saying nothing happens. The danger lurks in its very harmlessness.
2. The main symptom of the transition is that you start asking before you've finished thinking it through yourself — "fix this" becomes "suggest some options." Researchers call this reactive writing. Technically the choice is still yours, since you pick one of the options the model offers, but choosing from a ready-made menu written by someone else takes far less effort than hovering over a blank line.

A 2023 experiment by Jakesch and co-authors had more than a thousand people write a post about the benefits of social media alongside an AI assistant deliberately tuned to pull for or against. The goal was to find out whether such co-writing changes not just the text but the writer's own opinion, and it turned out that it does — participants paired with a biased assistant agreed with its position twice as often as those who wrote on their own. Most unexpectedly, almost none of them noticed they'd been nudged anywhere at all.

3. Here the substitution stops depending on what you're asking for. You're still using the model sparingly, you read the output, you're not skipping a single step of control, and yet the substitution still happens. It doesn't occur at the level of content you're checking, but at the level of syntax you never inspect at all. The study "Voice Under Revision" shows that even a harmless request like "just fix the grammar" statistically shifts the text toward more formal, impersonal language and reduces the share of first person — the "I" is quietly washed out of a text you still consider your own. The pause before the word still exists at this stage, but its outcome gets covertly adjusted when you've already fixed everything, moved further down the text, and stopped looking.
4. Next, what changes isn't the text anymore, but the very instrument you might have used to notice it. An MIT Media Lab study split participants into three groups — one wrote essays with ChatGPT, another used a regular search engine, and a third managed without any tools at all — and recorded their brain activity via EEG the whole time. In the group that wrote with the model, neural connectivity (how coherently and how distributedly different brain regions work together while performing a task) turned out to be noticeably weaker. These same people had worse recall and couldn't quote their own text just a few minutes after writing it.

You can only spot the pause we keep mentioning by contrasting the way you write with AI and the way you wrote before, and that context is being erased. If you believe the research on cognitive offloading (when the brain stops holding onto work that it can hand off to an external tool — like phone numbers nobody memorizes anymore), the very ability to write without mediation weakens.

The difference between picking up a skill and losing your identity hinges on a single detail — whether that pause before the word survived. A person can take the structure itself from the model and still decide what to say without prompting, or the model can decide and the person can simply agree, and both versions look almost the same. The only way to feel out the problem is to ask “...so what was point two?”

What we're afraid of when we're afraid of machines

This is just another rerun of an old show. In previous episodes we already blamed the typewriter for erasing individual handwriting and the spell-checker for depersonalizing the mistake. The anxiety follows the same path every time: a scandal over the mere fact of using the thing, the institutionalization of the tool as the norm, and finally, the argument shifting from "Is it allowed?" to "Are you obligated to disclose that you used it?"

In three years, people still haven't settled on what exactly scares them most, but the field has narrowed down to three main fears.

The most visible one is that the text itself is degrading, becoming flattened and averaged-out. This is where you'd also file the whole cult of the em dash, and the test the NYT ran in the spring of 2026, where 54% of readers given two passages on the same subject preferred the AI-generated texts over the originals by Cormac McCarthy and Ursula Le Guin. At bottom, this is anxiety about the product, not about the person who makes it.

The second is the fear that culture itself is losing its diversity of voices, and that in twenty years or so they'll all merge and dissolve into one grey, faceless mass.

The third, and quietest, belongs to an EEG at MIT: here we're afraid of losing the very capacity to think for ourselves.

Bad Machine Habits That Have Actually Done Some People Good

This sad story has a flip side. Columns just don't get written about it as often.

You've become more polite. Messages no longer start with "listen.” Now it's "great question" (even when the question is mediocre at best). Oddly enough, people you talk to like it.

You've traded filler words for structure. "So like, basically, we need to do the thing" has turned into a calm "first... second... third." Your boss is pleased. Your ex would be surprised too, hearing you now.

There are fewer fights in your messages. Your hand reaches on its own to type "I understand your point of view, but I'd like to point out" before any disagreement, and the drama usually ends there before it even starts. Your parents think you've finally grown up. Your friends think you've been swapped out for a slightly different version of yourself.

You've also become more convenient for the machines themselves. Your texts now clear the filters more easily and read more smoothly.

If someone ever asks what you actually think about all this, there's a real chance you'll answer with "great question, let's break it down point by point."

Anxiety About Money, Wrapped in Anxiety About the Soul

The loudest voices belong to those with the most to lose — columnists at major outlets, writers, editors, and generally anyone whose profession has, for decades, been built on choosing words better than everyone else. The danger gets framed in the language of loss, the claim being that it's not even the text losing its soul, but the person who used to be its author. In reality, this is anxiety about money, wrapped in anxiety about the soul because the latter sounds nobler than the former.

The whole raft of systemic solutions the professional community is proposing today comes down to a single word: disclosure. People want authors to confess to using AI, to draw a clear, visible line between "I wrote this" and "I got help writing this." We've been through this before, just with photography instead of text: World Press Photo drew this line long ago, separating a tweaked exposure from a person erased out of the frame (the latter has cost entrants prizes and sometimes even careers).

For text, that same logic is already taking shape as concrete editorial standards. The Globe and Mail requires staff and freelance writers alike not to use AI for writing or editing text — full stop.

In November 2025, The New York Times published an essay by Canadian writer Kate Gilgan in its Modern Love column. In March 2026, Lit Mag News editor Becky Tuch wrote on X that the piece read as if it had been generated by a neural network, quoting a passage with its telltale parallelism ("Not hate. Not anger. Just the flat finality..."). Opinions split: writer Ann Bauer defended her colleague, calling the parallelism simply the recognizable style of a longtime columnist, while others argued that an unsubstantiated accusation set a dangerous precedent. Gilgan did not respond to press inquiries. The Times itself limited its response to a statement that journalism at the paper "is, at its core, a human endeavor" and "that won't change," but no concrete disclosure policy followed.

There is, however, a world beyond newsrooms. It is larger. And it does not care.

A woman from Long Beach lost her eviction case where she represented herself because she hadn’t been able to explain that she'd been entitled to a payment deferral. She turned to ChatGPT and Perplexity, asking them to find procedural errors in the judge's ruling and draft a formal response to the court. "I never, ever, ever, ever could have won this appeal without AI," she later told reporters, having fought off eviction and fines totaling more than fifty thousand dollars. She didn't become a better writer in any sense that would matter to an Atlantic columnist — she had no literary voice before, and she didn't gain one now. For the first time in months of litigation, she simply had a document the court was obligated to read.

There are, in fact, a great many such documents. Debating the fate of the human voice in instructions, contracts, or business correspondence isn’t really appropriate. Many texts have always been about pure function. The norm shifts from the bottom up, not the top down, moving from forms to business correspondence to legally weighty documents, and only then do newsrooms have to codify what readers had already decided on their own long before.

Anxiety over the loss of voice requires one precondition: a voice had to exist before it became possible to lose it. Grief is available only to those who managed to acquire one in the first place — a narrow circle of people for whom text was not a function but an achievement, a source of pride. The louder someone mourns this loss, the more reason to suspect they're mourning not some universal human drama, but their own privilege.

How to Hold Out to the Bitter End Without Upsetting Your Editor

In the late seventies, the admissions committee for the creative writing (MFA) program at Johns Hopkins University looked at manuscripts typed on a word processor with suspicion: you could just tell the author hadn't sweated over the text the way one was supposed to sweat over a typewriter. Today, a handwritten manuscript would raise more eyebrows than a computer file — and no one bothers to remember the situation was ever any different.

The argument over voice will end the same way on both sides. Sooner or later, the models will stop churning out the same slick patter about tapestry and learn to adapt to a given author's style and habits, instead of handing everyone the same averaged-out text scraped from the internet. People, in turn, will stop flinching and working themselves into a state at the sight of an em dash, and might even pick up a useful habit or two from the machine — learning, say, to actually finish a thought instead of abandoning it mid-paragraph. The ending is already written; it just hasn't arrived yet.

Until it does, it's worth resisting. Catch yourself and the machine in triads and anaphora that sound weighty right up until the moment you notice your beautiful, singular thought has been swapped out for a template. Catch that polite "I understand your point of view, but," which more often covers up disagreement than expresses it. Catch the moment when it's easier to ask the model than to think it through yourself, and you ask before you've finished thinking. Guard the pause, not because it will outlast this argument, but because, while the argument is still going, it still means something — to you, and to whoever is editing your text.