The Structure of This Sentence Is a Dead Giveaway That AI Wrote It

For as long as people have been using AI to churn out text, other people have been coming up with “tells” that something was written by AI. Sometimes it’s punctuation that comes under suspicion. (The em dash is generally considered the shadiest.) Other times it’s words that robot writers seem to love and overuse. But what if the biggest giveaway that a text was written by AI isn’t a word, phrase, or punctuation mark, but a particular sentence structure instead? Why is it so hard to make AI writing sound human? The idea that certain rhythms of sentences might be a sign of AI writing first came to my attention through my work as a professional word nerd. Recently, I a potential new client contacted me about helping to polish up some of their writing. As an editor, that’s not unusual. But like several recent inquiries, this assignment came with an AI-age twist. The client had conducted a good amount of research for a work project and then asked a popular LLM to synthesize the findings. Afterward, they checked it for factual errors and removed anything that seemed an obvious red flag for AI writing. But the text still just didn’t sound human. Could I fix it? I agreed that despite the client’s considerable efforts, something still sounded off about the text. I also concurred it wasn’t immediately easy to spot what it was. All the commonly cited tells of AI writing had been removed. There wasn’t an em dash or a delves in sight. Still, it felt like it came from a bot, not a human. The problem was clearly deeper than word choice. I faced this dilemma from the perspective of a communications pro. But there are plenty of others scratching their heads over the same issue. These are the entrepreneurs, marketers, and others who want to use AI to speed up their workflows but don’t want to annoy others with robotic off-note emails and reports. The group also includes writer Sam Kriss. AI tells are more than weird words and punctuation In a fascinating article in The New York TImes Magazine, Kriss delves into the stylistic tics that are certain, frequently infuriating, tells of AI writing. Unlike more quantitatively focused recent studies, he doesn’t focus on easy-to-measure features like the frequency of certain words or punctuation marks. Instead, he investigates the larger patterns in AI writing that contribute to its uncanny and often deeply annoying feel. AI, for instance, lacks any direct experience of the physical world. As a result, AI writing tends to be full of imprecise abstractions. There are a lot of mixed metaphors. Bots also overuse the rule of three. (Lists of descriptors or examples are generally more satisfying for the reader in groups of three.) Phrases that are common in one country or context are reproduced in others where they sound foreign. If you’re either a language lover despairing about the current flood of AI slop or a practically minded professional looking to use AI without irritating human readers, the article is definitely worth a read. But one of Kriss’s observations in particular set alarm bells ringing in my mind. “It’s not X. It’s Y” “I’m driven to the point of fury by any sentence following the pattern ‘It’s not X, it’s Y,’ even though this totally normal construction appears in such generally well-received bodies of literature as the Bible and Shakespeare,” he writes. Kriss goes on to cite instances of this “It’s not X, it’s Y” sentence construction in everything from politicians’ tweets to pizza ads. Appearances in great literature notwithstanding, the recent flood of examples has transformed this phrasing into a sure-fire way to know you’re reading something written by a machine. Hmmm, I thought, reopening my client’s document. Sure enough, when I reread my new client’s oddly mechanical writing, I saw that particular sentence construction in nearly every paragraph. One AI tell that’s easy to scrub Getting rid of all the giveaways that a particular text is written by AI is difficult. It might just take you longer to do a thorough scrub job than to just actually put in the intital effort to write the thing yourself. (Which is, as a side note, what I often tell clients looking for this sort of editorial work.) Plus, writing is good for your brain. In other instances of more mechanistic writing, keeping AI style might not matter. Who cares about the literary merits of the executive summary of a data analysis if the numbers and the takeaways are correct? If that’s the case, don’t sweat the odd, “It’s not X. It’s Y.” But if you’re producing ad copy, a presentation, or persuasive content and you want the reader to feel like a human actually wrote it, Kriss’s article is a helpful reminder. Sure, certain words or language ticks might be more common in AI writing. But the overall problem is usually deeper. If you really want to try to make AI language passably human, you need to worry not just about word choice and eliminating hallucinations. You need to look more deeply at the way the sentences are constructed. And you definitely want to avoid “It’s not x. It’s y.” As a bot might put it, this sentence structure isn’t just a cliché. It’s now a dead giveaway that AI wrote the text. EXPERT OPINION BY JESSICA STILLMAN @ENTRYLEVELREBEL