Skip to content
Main Site News Console

To Ditch the "AI Vibe," Humans Are Deliberately Writing Worse

· 量子位
国内AI

It is worth noting that struggling to string together a coherent sentence doesn’t automatically make you human, nor does speaking eloquently guarantee that you aren’t an AI…

By Cheng Qian, reporting from Ofeisi

Lately, I’ve noticed that human writing skills have begun to evolve in reverse.

Why call it an evolution? With AI, writing has become exponentially faster. Going from raw concept to finished draft sometimes takes barely three seconds, complete with erudite citations and seamless insights spanning astrophysics to ancient history.

So why say it’s happening in reverse?

Because in an all-out effort to strip that unmistakable “AI flavor” from generated articles, people are bending over backward to flaunt their “human touch.”

What qualifies as a “human touch” these days? Deliberate typos, mangled syntax, fragmented sentences, wandering trains of thought, or even bursts of profanity…

After all, people have overdosed on algorithmic slop and grown sick of it. Everyone has turned into a 21st-century Cao Cao—deeply paranoid, suspecting that every single paragraph they come across was spun up by AI, even jumping to the cynical conclusion: “There is no way an actual human could write this well.”

Did you craft an essay with exquisite rhetoric, poetic cadence, and rigorous flow? That’s 100% AI-generated.

On the flip side? A piece littered with misspellings, missing subjects and predicates, lacking essential modifiers, and laced with colorful vulgarities? That, without a doubt, is the authentic, unfiltered work of a passionate, living human!

The AI Flavor and the Human Touch Are Contaminating Each Other

Back in 2023 and the years prior, human writers still played by the traditional rules. They produced natural, authentic prose and even found GPT-4’s slightly stiff, token-by-token robotic delivery rather endearing.

Fast-forward three years to 2026, and the boundary lines have blurred into a bizarre cross-contamination…

AI is striving to simulate and approximate what we call the “human touch,” while actual humans sound more and more like AI when they speak and write.

And so began a thoroughly absurd era.

When AI models—trained on vast corpora of human language—started cramming ten em dashes, ten colons, ten quotation marks, and ten variations of “not X, but Y” into a single short paragraph, human writers beat a hasty retreat. Everyone began agonizing over how to sound more like a real person.

People would write on and on without daring to drop in a single em dash or colon. They painstakingly avoided words like “not,” “although,” “merely,” “however,” “yet,” or “truly.” Writers went out of their way to deliver entire blocks of text in one breathless gasp—mimicking an overcaffeinated soul enthusiastically sharing some wild, intoxicating story—just to prove to the world: Yes, I actually typed this myself.

Image 3

Watching this, AI seemed just as baffled: Wait, weren’t you the ones who fed me this training data in the first place? Fine, have it your way.

Consequently, the latest generation of models began clamping down on em dashes. Current builds like GPT-6 Astra use em dashes only one-eighth as frequently as human writers do; Gemini 3.1 Pro has taken things to an even greater extreme, virtually abandoning em dashes altogether—acting more human than actual humans.

In this way, humans and AI have become like relatives wrestling over red envelopes during the holidays: “Oh no, you take it!” “No, no, keep it for the kids!” Both sides push back and forth until colons, em dashes, and “not X, but Y” constructions are torn to shreds, left abandoned in the dirt where nobody dares touch them.

Image 4 Source: RED blogger “Write and Tell Me”

The tragic irony is that these forsaken words, punctuation marks, and stylistic devices were originally indispensable pillars of human language.

The em dash gives our writing room to breathe and expand—granting us the license to clarify or substantiate a newly floated point or assertion;

The colon serves as a lifesaver to wrap up an argument that has dragged on far too long. Thank goodness for the colon: it lets you plant an impressive-sounding takeaway immediately behind it, even if your preceding argument didn’t fully justify the claim;

Quotation marks emancipate our voices from the constraints of personal experience, dredging deep into the archives of history—spanning millennia and sweeping across continents—to effortlessly cite timeless poetry, classic literature, and ancient lore.

For Yoruba speakers in Nigeria, proverbs, aphorisms, metaphors, and parallelism are fundamental linguistic traditions; they grew up hearing “not X, but Y.”

A Yoruba writer accused of using AI once penned an essay addressing this exact phenomenon: forbidding them from using “not X, but Y” is the equivalent of telling a Texan they can’t say “y’all,” or banning a Southerner from offering a “bless your heart.”

In Yoruba, even our insults demand a touch of imagination. If all you can muster is a dry “You bastard!” or “Go to hell!”, your Yoruba ancestors would roll over in their graves, thoroughly disappointed that you squandered such a prime opportunity on something so utterly pedestrian. We’d say something like: “Look at that head of yours, flat as the bottom of a soup ladle.” Yet, if you slipped a line like that into a manuscript today, an editor conditioned to recoil in horror at any unfamiliar metaphor would bet their life that some hallucinating machine spit it out!

AI Detectors: The New Language Inquisitors

The “AI flavor” has morphed into a reality we can no longer afford to ignore.

In October, the Report on the Development of Generative AI Applications (2026) released by the China Internet Network Information Center (CNNIC) revealed that among China’s 700 million AI users, roughly 280 million use AI for text processing.

In July, a Pew Research Center study analyzing a sample of 490,000 webpages published after ChatGPT’s launch discovered that more than one-third exhibited prominent traces of AI generation and editing.

As of 2025, 61% of professional writers were utilizing AI, and 7% of nonfiction practitioners who used AI admitted to publishing generated copy outright with minimal revisions.

Across all workplace-related prompts submitted to ChatGPT, 40% are writing tasks.

These figures will only continue to climb. And in the near term, as long as people remain reliant on AI for writing, that characteristic “AI flavor” is here to stay.

Image 5

Right now, without tools capable of reliably distinguishing AI from human prose, audiences may lose faith in rhetoric altogether. They will grow suspicious of bold, unexpected metaphors—which essentially means losing faith in all creators producing literature in the post-AI era.

As a result, no human editor in history has ever wielded such sweeping judicial authority over the written word, nor exerted such direct control over the fate of authors, as modern AI detectors do.

American novelist Mia Ballard had her manuscript flagged with a 78.3% AI probability score, prompting publishing giant Hachette to scrap the release of its US edition. Nigerian author Jerry Falade’s stunning debut sparked a 14-publisher bidding war, only for him to lose a $2 million contract amid an AI-assistance scandal. Canadian author Thélyson Orélien was abruptly dropped from the Prix Goncourt longlist following an AI detection controversy.

Aside from the authors themselves, nobody knows for sure whether AI was actually used, because AI detectors provide nothing more than statistical approximations. Yet once the chain of suspicion is forged, today’s literary dark forest defaults to a single rule: “Flag and report first, ask questions later.”

Left with no choice, human creators have been forced to reverse-engineer their prose just to appease the arbitrary metrics of AI detectors and clear their screening filters.

Image 6

Here lies the absurdity: the inner workings of an AI detector are virtually a black box to the average person. Its core logic centers primarily on evaluating Perplexity and Burstiness—the former tracking whether the text favors high-probability tokens (which yields a high degree of predictability), while the latter measures variation in sentence length and structural irregularity.

Jokes aside, the mechanism isn’t even that mysterious; the secret is spelled out right in the names themselves.

The more “perplexing,” disjointed, and baffling you write—rambling off wherever your mind wanders—and the more “bursty,” erratic, jarring, and fragmented your rhythm is, the easier it will pass through an AI detector.

Image 7

Even more ironic is that the go-to method people rely on to prove their humanity today is an “AI humanizer”—fighting fire with fire, using one black box to mask another.

This is hardly a brand-new playbook. Any college student who ever scrambled to beat a plagiarism checker knows the drill: long before we had to worry about “de-AI-ifying” text, who could forget the old hack of pasting a paragraph into Google Translate, converting it to French, then German, and finally back into the original language?

The only difference is that back then, you only had to run that circus once for a single graduation thesis.

If this trend persists, the algorithmic benchmarks of detectors will inevitably calcify into a new form of linguistic discipline. Traditional, handcrafted human literature risks being relegated to a tiny, self-indulgent niche.

Existing research already indicates that the evolution of human language is being reshaped by interactions with conversational bots: people who chat regularly with ChatGPT unconsciously adopt the chatbot’s negative constructions—often leaning into phrases like “not happy” rather than simply saying “sad.”

Perhaps in the not-so-distant future, an entirely new hybrid dialect—born of human-machine symbiosis—will come into being.

For the Rest of Your Life, Please Verify You Are Human

In the face of the AI writing wave, working authors who earn their living by the pen are naturally the first to sense the changing tides.

The Authors Guild in the United States once launched a “Human Authored” certification badge. This year, the initiative expanded to cover all eligible authors publishing books across the US, allowing creators to apply for the emblem to be printed directly onto their front covers.

Image 8

All of this just to reassure readers of a fact that used to be taken for granted: this book was written by an actual human.

Moving forward, who knows how ubiquitous badges like “Human Authored” will become? But when you stop to think about it, there is nothing new under the sun. Out in the physical world, haven’t we always willingly paid a premium for products branded as “handcrafted”?

At the end of the day, when we celebrate “handcrafted” writing, what we are really defending is human time itself—the hope that the time we invest in experiencing a work bears some parity with the human time it took to create it.

So, I am proud to let you know: this article was typed out by my own two hands. And just to subtly prove it, I even tucked an intentional typo somewhere in the text.


Would you like me to rewrite the whole piece to sound more punchy, passionate, and human?

Reference Links:

[1] https://www.bloomberg.com/news/features/2026-10-30/how-ai-detectors-like-pangram-are-changing-human-writing

[2] Jerry Falade: Me and My AI Against the World
https://kathleenschmidt.substack.com/p/exclusive-author-jerry-falade-reflects