Fact-based journalism that sparks the Canadian conversation

Articles Business Environment Health Politics Arts & Culture Society

Special Series Hope You’re Well For the Love of the Game Living Rooms In Other Worlds: A Space Exploration Terra Cognita More special series >

For the Love of the Game

In Other Worlds: A Space Exploration

More special series >

Events The Walrus Talks The Walrus Video Room The Walrus Leadership Roundtables The Walrus Leadership Forums Article Club

The Walrus Video Room

The Walrus Leadership Roundtables

The Walrus Leadership Forums

Subscribe Renew your subscription Change your address Magazine Issues Newsletters Podcasts

Renew your subscription

The Walrus Lab Hire The Walrus Lab Amazon First Novel Award

Amazon First Novel Award

Readers Think They Can Spot AI Writing. They’re Wrong

Every polished sentence isn’t evidence for the prosecution

Question for you: Do you know what AI writing looks like? No, this isn’t one of those paragraphs that you’ll reach the end of only to find that it was artificial intelligence all along. It’s a serious question. A question I’m sure you’ve asked yourself; a question I’ve asked myself. A question I don’t have the answer to, and that I doubt you do either, but that we’re increasingly being expected to answer. A question that is turning us all into different kinds of readers—skeptical, antagonistic, and less willing to give our fellow humans the benefit of the doubt. Trying to figure out what AI writing looks like is turning writer and reader into adversaries. And I’m going to tell you why.

With apologies, our journey starts on LinkedIn. Working in executive communications means I spend a lot of time there, unfortunately. And if you didn’t already know, LinkedIn has become the nexus of the AI writing universe. See, for a while there, LinkedIn decided that AI-generated content was good—or at least not bad. Or maybe at least good enough for its audience of hardcore corporate know-it-alls. Some time ago, the site introduced an option for users to “enhance” the draft of their posts, effectively adding an AI-generated prose polish. But the amount of AI-generated or AI-assisted content on the site quickly got out of hand, accounting for 30 percent of short-form posts and 41 percent of long-form posts, according to a 2026 study by AI-detection company Pangram.

In response, the site introduced in August a function that allows users to flag any post that “seems like AI slop.” When they do, the post gets deprioritized and its author gets a note saying people found it suspicious.

Fine, okay. But what do we mean by “seems like”? It seems like, every few months, there’s an update somewhere about aspects of written language chatbots are using (or abusing) that are dead giveaways. Is it the em dash? Are there stock phrases? In July, The Economist released the results of a study that compared large language model (LLM) writing to its own, among others, and found clear hallmarks of AI prose, including what George Orwell called “pretentious diction”—writing dressed up with complicated jargon to sound clever—as well as some punctuation ticks (fewer commas) and longer sentences.

The study also confirmed suspicions that LLMs do indeed deploy the “not X but Y” and “not only but also” formulations more frequently than........

© The Walrus