Write Drunk, Edit Sober: The New Rules of AI Prose

For Franz Kafka, putting words on a page meant “unending torments.” George Orwell compared it to a “long bout of some painful illness.” Hunter S. Thompson said it was the exact opposite of sex, noting it was “only good when it’s over.” Ernest Hemingway rarely produced thousands of words in a day, and he required much more alcohol to do it.

Today, writing is easy. Large language models (LLMs) churn out thousands of words a minute without breaking a sweat and never get a writer’s block. They draft a third of new websites and have pushed the number of new Amazon e-books to 300,000 a month. But for human writers, this endless flood of text is both a blessing and a curse.

If you claim to be an author but secretly use AI, you risk looking like a fraud. There are even claims that AI-generated prose won the Commonwealth Short Story prize, tricking judges with its “quiet authority.” (The Commonwealth Foundation swiftly denied this.)

Wordsmiths are spooked by this linguistic spectre. AI can mimic a Shakespearean sonnet or write a dull office manual. Spotting this ghost writer, however, is getting tricky. Software like Pangram boasts a 99.98% success rate in catching AI text, earning it a partnership with the blogging platform Substack. Yet, these detectors are black boxes that sometimes guess wrong, and they can’t explain how they reached their conclusions.

You might think you can spot AI just by looking for a few annoying habits. But relying on old tricks is no longer enough. Karolina Rudnicka, a linguist at the University of Gdansk, points out there’s no single style of AI writing, just as there’s no single style of human writing.

To find out what AIs really sound like today, The Economist ran a massive test. They asked 14 versions of ChatGPT, Claude, Gemini, and Grok released since 2024 to rewrite their articles. They checked 1.2 millions words across 55,940 sentences. To make sure they found AI quirks and not their own, they compared the results to human journalism from CNN, the New York Times, and the Washington Post, plus hit novels from 1950 to 2022.

The findings were surprising. AI updates fast, and its writing style changes with every patch. For example, people used to joke  that AIs loved em-dashes. Today, only Claude overuses them. ChatGPT now uses fewer dashers than a typical human. When asked about this, an older version of ChatGPT cheered, “Ha—great question!” Today, the updated model soberly replies that dashes are “best used sparingly.”

AIs have dropped old favourites like “delve” and “rich tapestry.” Now, they hide behind long, fancy words. They use bulky terms like “significant,” “increasingly,” and “consequences.” They prefer rare words like “interdependence” and scientific jargon like “parameter.” Gemini and Claude are especially guilty of this.

George Orwell hated this kind of writing. He called it “pretentious diction,” mocking writers who used Latin words instead of punchy Anglo-Saxon ones to sound clever. AIs do exactly this, stuffing their sentences with Latinate suffixes.

They also hate punctuation. Much like the novelist James Joyce, AIs write long sentences without many commas or semicolons. Their favourite word is “and.” When they want to sound lively, they lean hard on the “rule of three” to group ideas, or use predictable phrases like “not only but also.”

Why do they write this way? Tommie Juzek, a researcher at Florida State University, notes that AIs learn from human feedback. They repeat the tricks that impress people and drop the ones that don’t.

How do we fix the flood of slop?

If AI keeps improving, it might soon write as well as a professional. Until then, reading its wordy output drains your time and brain power. This puts a huge premium on editing skills. AI is making us all editors, and here’s how to manage your new robot staff.

First, give good instructions. The author George R.R. Martin said an editor’s job is to “clip the wings” of writers and “tell them in which direction to fly.” Martin called editors a writer’s natural energy. A sycophantic bot, however, will never push back. Tell the AI exactly what you want and make it rewrite the text until it gets it right.

Second, cut the fluff. AIs love to sound mart while saying nothing. If a bot writes about the “bi-directional flow” to “expediate grid enhancements,” swap those Latinate words for sharp Anglo-Saxon ones. Cut the tired clichés, like “Achilles’ heels.” And if an AI uses the rule of three to say, “The mystery has been solved—not partially, not ambiguously, but definitely,” just delete the extra words. End the sentence at “solved.”

Finally, know who will read the text. Bots don’t understand tone. If you’re using AI to write an email breaking bad news to your boss, tell the bot to drop the exclamation marks. A chatty AI will pin you to the will like a bore at a party.

An old writing rule says you should “write drunk, edit sober.” As humans and machines team up, it looks like AI will get to have all the fun, while we’re stuck doing the sober work of cleaning it up.