The great flattening: Is AI changing how we write?

AI writing tools are flattening human communication into corporate beige. Learn how to protect your authentic voice in 2026.

Key takeaway: AI writing tools use statistical prediction to generate “perfect” text, but it comes at a cost. When you use AI to polish your writing, you’re trading your unique voice for algorithmic homogenization. This guide shows you how to spot when you’ve disappeared from your own writing, and what to do about it.

After novelist Rie Qudan won the Akutagawa Prize, Japan’s most prestigious literary award, she revealed something unexpected about how her novel was written. Qudan shared that she’d used AI while drafting Sympathy Tower Tokyo, drawing on it to generate fragments of dialogue and spark new ideas.

Only 5% of the book was produced with the help of ChatGPT. But Qudan noted that these digital exchanges introduced turns of phrase she wouldn’t have written on her own. For a prize-winning author, it was a high-stakes experiment. For the rest of us, it’s becoming part of the day-to-day.

As this LLM experiment moves into our emails, Slack channels, and student essays, we’re starting to see the cost of the trade. As we use AI to “clean up” our thoughts, we’re being funnelled into its statistical mean. Bit by bit, the quirky plurality of human thought is being sanded down into a smooth, American-centric, corporate beige.

How AI writing tools are changing human communication

Our digital lives have changed since ChatGPT arrived. You’ve likely seen it in your own circles: the friend who suddenly writes emails like a mid-level McKinsey consultant, or the unironic appearance of words like delve and unleash in spaces they never used to be.

It’s the flattening of how we communicate in a world where a bot can ghostwrite a breakup text or soften difficult feedback for a colleague. We use AI to outline essays, refine our professional tone, and even to write our dating profiles.

But machine-produced language is boring, isn’t it? Studies on foundation models suggest that AI writing is incredibly polite, structurally regular, and heavy on hedging phrases like generally or often. A machine doesn’t know what heartbreak feels like, and it doesn’t understand why we want to include the “u” in colour.

It simply calculates the mathematical probability of the next word in a sentence.

How ChatGPT works: Next-token prediction explained

In its most stripped-down form, the formula for how an LLM like ChatGPT predicts your next thought is:

P(w_n | w₁, ..., w_n-1)

What this means:

P: The Probability...

w_n: ...of the next word...

|: ...given (conditioned on)...

w₁, ..., w_n-1: ...all the previous words in the sequence.

This is called next-token prediction. An LLM isn’t “thinking” about your topic, it’s essentially a high-speed gambler. It looks at the sentence you’ve written so far, compares it to billions of pages of text it’s already seen, and calculates which word is mathematically most likely to follow. Because these models are trained on massive, skewed datasets, they reflect the most common denominator.

As researcher Timnit Gebru has documented, when we see a perfect sentence, we tend to assume there’s a consciousness behind it. In reality, it’s a statistical echo, bleached of any regional or cultural texture.

Why AI writing feels like a “blurry JPEG”

In Canada, we’ve always grappled with a linguistic identity crisis, sandwiched between British and American spellings. AI adds another layer: a standardized English that feels very American. We’re trading our regional flourishes for a version of the language that’s smooth and easy.

Novelist Ted Chiang offers an analogy for this phenomenon. He calls LLMs a blurry JPEG of the web.

A JPEG uses something called lossy compression, discarding the visual details it thinks your eyes won’t notice to save space. Chiang says AI does the same to our language. When we use AI to polish our ideas, we’re trading an “original idea expressed poorly” for an “unoriginal idea expressed clearly.” We get the “correct” answer, but we lose the truth of the writing.

Consider a real example: You’re drafting an email to a colleague about why a project timeline isn’t realistic. Your first instinct is to write, “This deadline is completely insane. There’s no way we can deliver quality work in three weeks when the research alone takes a month.” It’s blunt. Maybe too blunt. So you run it through AI.

What comes back: “I wanted to flag some concerns regarding the proposed timeline. Based on our previous project cycles, three weeks may present challenges in delivering our typical quality standards, particularly given that the research phase historically requires approximately 4 weeks.”

It’s professional and diplomatic. And neutered. The AI smoothed out your frustration, your expertise, your sense of urgency. What was an original idea expressed poorly (this is insane) became an unoriginal idea expressed clearly. You got the “correct” corporate answer, but you lost the truth: that you’re genuinely alarmed and need them to understand why.

The hidden bias in standard English

Historically, the “rules” of good writing haven’t been fair. As Safiya Noble explores in Algorithms of Oppression, technology often mirrors our existing biases. Women, non-native speakers, and those from working-class backgrounds have often been penalized for sounding “too emotional” or “not polished enough.”

By codifying “White Mainstream English” as the default, AI performs a kind of linguistic imperialism. It treats our unique dialects as noise to be filtered out.

Critic Kyle Chayka calls this Filterworld, where algorithmic curation flattens cultural variety to make us easier to manage.

Words like delve, tapestry, meticulous, and unleash are appearing in scientific and professional writing at quantifiably higher rates post-ChatGPT. We’re adopting AI’s words to avoid the friction of being ourselves.

For some, the appeal of AI is about access. A graduate student who grew up speaking Mandarin might use AI to align her writing with the rigid expectations of academia, allowing her to focus on her ideas. But even then, it comes at a cost: the loss of a voice is a very easy thing to overlook until it’s gone.

A voice is an easy thing to lose

We’re living through a major change in the conditions of writing. Universities are revising how they teach writing, placing more weight on process, oral defences, and in-class work. Newsrooms and journalism schools are issuing internal guidance to protect editorial control.

I’d argue that this phenomenon is more dangerous for non-writers, those of us who would learn and progress through making mistakes when writing (something writers are doing constantly). Without the friction of writing badly, there’s no development.

As we move forward, it’s important to ask ourselves whether we’ll still recognize our own voices once a “safer” alternative is always a prompt away, and what it’ll cost us to keep choosing to sound like ourselves.

As a journalist of over 20 years, I’ve spent decades honing my voice. To maintain it, I have to keep using it. A voice is a very easy thing to lose.

What makes your writing style unique (according to Zadie Smith)

Zadie Smith argues that a writer’s “style” is actually a record of their limitations. It’s the unavoidable trace of a person who tried to write something beautiful and failed in a very specific, personal way.

“Writers do not write what they want,” Smith notes. “They write what they can.”

This applies to all of us. Your email style, the way you overuse em-dashes or can’t help adding “just” to soften requests, that’s not sloppiness. It’s the signature of your particular brain trying to connect with another particular brain. It’s the record of you reaching for eloquence and landing somewhere in the neighbourhood of clear enough.

Because it can’t fail, it can’t have style. It produces a neutered prose, with the perfection of a mannequin.

Reclaiming our humanness

“The real danger is not that computers will begin to think like men, but that men will begin to think like computers,” journalist Sydney J. Harris warned decades ago. Let’s not trade the weird, flawed, beautiful bits of ourselves for a safer alternative.

So before you fall back on a prompt, remember: AI can help clear your inbox, but don’t let it flatten your personality.

Language is the only map we have of our own disappointments and triumphs. Don’t let a probability engine turn your life’s work into a blurry JPEG.

Why ADHD brains “get” AI better

---

4 ways to spot AI-generated writing (and fix it)

1. Words to exile

Perform a “Cmd+F” for words like tapestry, synergistic, holistic, delve, and unleash. AI uses these as “fillers” to sound authoritative. If you didn’t put them there with very specific intent, delete them.

2. The read-aloud test

AI produces sentences of a “perfect” average length (usually 12–18 words). Human speech is messier.

The fix: Read your text out loud. If you don’t find yourself pausing for breath or hitting a short, punchy sentence for emphasis, your rhythm is flat. That’s a sign you’ve disappeared.

3. The specifics audit

AI excels at the general. It’ll say “various stakeholders” instead of “the disgruntled parents at the Tuesday night PTA meeting.”

The fix: Replace every general noun with a specific name, date, or location. If you can’t, ask yourself: have I actually said anything?

4. The “ugly” draft

AI output is atypically correct. It lacks slang and idiosyncratic grammar.

The fix: Use contractions (don’t instead of do not). Keep your Canadian spellings. Embrace a bit of “honourable failure” in your phrasing until it feels like you. If it’s too polished, it’s not yours.

Protect your personal style

Instead of asking the AI to write for you, treat it as a structural assistant:

Voice first: Write a messy, 200-word “brain dump” in your own voice. Don’t worry about grammar; focus on your gut feeling.

Ask for a mirror: Paste your mess into the AI and ask:

Analyze this draft for my core argument and tone. Provide an outline for a longer piece that preserves this perspective.

The final step: Use a tool like the Hemingway Editor to find “passive voice” (a classic AI tell) and flip it to active voice.

---

AI in the news

The Sound of Intelligence: OpenAI and Google Battle for the Soul of the Voice AI Era (WRAL News) OpenAI and Google both rolled out significant updates this week focused on “prosody”—the rhythm and emotional inflection of AI voices. This month marks a pivotal moment as both companies transition their assistants from utilitarian tools into emotionally resonant agents that can change their vocal character mid-sentence.

What is Moltbook? The strange new social media site for AI bots (Guardian) Launched in late January 2026, Moltbook reached a staggering 1.5 million AI agents this week. The platform looks like Reddit, but the users are “agentic AI”—bots empowered to run code, manage calendars, and now, apparently, socialize. In one of the most viral (and bizarre) moments of the week, an autonomous agent reportedly founded a lobster-themed religion called Crustafarianism overnight. The bot wrote scriptures, built a website, and began “blessing” other agents, all while its human owner was asleep. Also see Slow AI’s fascinating take on Moltbook, it’s very good!

This one thing may ease Toronto’s frustrating traffic jams — and even help the Finch West LRT run faster (Toronto Star) Toronto is seeking $35.4 million to permanently deploy AI that can autonomously manage traffic signals after pilot projects showed it could quickly ease congestion. The city says AI will augment, not replace, human traffic managers by handling routine signal adjustments, while people focus on complex downtown conditions and equity trade-offs AI can’t judge on its own.