Opinions

From ‘Is It Cake?’ to ‘Is It AI?’: Why Writers Are Deliberately Sounding Less Perfect

BYRuman Baig

August 26, 2026

Remember a few years ago when the internet was obsessed with figuring out whether something was cake or an inanimate object? Everything from a tomato to a shoe could secretly be cake because of the sheer precision involved. Today, we’re playing a different guessing game: Is this written by a person or by AI? I catch myself asking that question more times than I’d like to admit in a day. Whether I’m reading an article, an Instagram caption, or a LinkedIn post, AI-generated language has become so ubiquitous that it has seeped into everything. But when did AI become the benchmark for good grammar, punctuation, and vocabulary— the very things we were taught made us better writers in the first place?

Early on, a mentor gave me advice that has stayed with me ever since: if a reader has to Google a word to understand your point, you’ve failed to make it. Another taught me how punctuation could shape rhythm and emotion, how an em dash or a comma could make a sentence feel as smooth as butter on toast. Those lessons made my writing more restrained, more intentional. They took years to develop. Ironically, they’re now some of the very things that make people suspect you’ve used AI. An editor recently asked me to avoid words like “quiet” and “enduring” because they “smelled of ChatGPT.”

Out of curiosity, I ran articles I’d written in 2017, 2018, and 2019 through an AI detector. Some came back as AI-generated, despite predating ChatGPT’s public release by years. So I tried something older still: the US Constitution. According to the detector, roughly 93 per cent of it was AI-generated — a finding that, it turns out, isn’t just me being unlucky with one tool. Ars Technica ran the same experiment in 2023 and got the same bizarre result, for the same underlying reason: detectors trained to flag “low-perplexity” text — writing that is unusually predictable, evenly paced, grammatically clean — will flag anything that fits that profile, whether a machine wrote it in 2024 or a room full of eighteenth-century lawyers wrote it in 1787.

This isn’t only writers second-guessing themselves in private. Over the past year, AI’s fingerprints have turned up in published work with embarrassing regularity. In May 2025, readers of the fantasy romance novel Darkhollow Academy: Year 2 spotted a stray editorial note in chapter three (left in, apparently by accident) instructing the model to rewrite a scene to mimic the style of a specific bestselling author. Screenshots spread across Reddit and Goodreads within days, and the author later admitted to using AI for editing she couldn’t otherwise afford. In November 2025, Pakistan’s leading English-language newspaper, Dawn, published a business story that ended not with a kicker but with a chatbot cheerfully offering to draft “an even snappier front-page style” version next time. And this July, freelance journalist Emma Grae flagged an AI-generated passage left in a Marie Claire Australia piece, which was deleted soon after — no correction, no acknowledgment, just gone.

None of these were writers deliberately trying to pass AI off as their own; they were people who forgot to delete a paragraph. But that’s precisely what makes them useful evidence. They tell us AI-flavored writing isn’t a paranoid hallucination some of us are having about our own prose. It’s sitting in published copy across genres and continents, waiting to be caught. I’ve caught myself doing the opposite: leaving in the odd typo or oddly placed comma, like I’m trying to fool a literary sniffer dog into believing there’s a human behind the words.

Writer Answesh Banerjee, who writes for Harper’s Bazaar and Vogue India, says the suspicion has changed how he works. “I’ve always had a flowery style — long sentences, em dashes. Over the last year, as people started identifying certain patterns as AI language, I’ve found myself cutting back on those instincts,” he says. He recently had a press release he wrote from scratch flagged as AI-generated by a client’s detector, and it was rewritten on that basis alone.

It would be easy to read this as an occupational hazard confined to a small, online-adjacent circle of culture writers. The data suggests otherwise. A study by Imperial College London, tracking two million job postings across 61 countries, found demand for freelance writing fell by roughly 30 per cent within eight months of ChatGPT’s launch — the steepest decline of any category studied. Upwork’s own platform data shows writing-project postings down 32 per cent year-on-year as of early 2026, the largest drop of any category on the site. This isn’t a handful of magazine writers feeling twitchy. It’s a labour-market shift, and the anxiety about sounding “too AI” is a downstream symptom of a much bigger one — sounding replaceable.

Journalist Arpita Hota, whose bylines include Vogue Arabia and Cosmopolitan Middle East, says AI has made her more protective of her own fingerprints, not less. “The things that make my writing mine are the weird observations, the cultural references, the moments only I would’ve noticed. If a piece reads like anyone could’ve written it, I’ve probably edited too much of myself out of it.” She’s also started cutting personal anecdotes that once anchored her stories, worried they’ll read as “too polished,” It’s a genuinely strange thing for a writer to have to worry about.

Even setting aside the Constitution stunt, the research on AI detectors is damning. A widely cited 2023 study that tested 14 different AI detectors, testing multiple leading detection tools, concluded they were “neither accurate nor reliable” and too easy to fool with light paraphrasing — the authors recommended against using them as evidence of misconduct at all. In other words, the machines writers are being edited to appease can’t reliably tell the difference either.

On Aug. 2 2026, the artificial intelligence company Anthropic started hiding an invisible watermark inside text written by its newer Claude models. It’s a response to new EU transparency rules, and it applies worldwide, not just in Europe. Here’s the basic idea: whenever the model has two equally good word choices, like “overcast” or “grey,” it now uses a secret key to help pick one. To a reader, the text looks completely normal. But anyone with the key can scan it afterwards and work out how likely it is that Claude wrote it. Older Claude models are being brought into this gradually, so the watermark isn’t yet universal across every version. The watermark also barely shows up if Claude has only lightly edited your own writing, since there aren’t many AI-picked words left to detect. And it fades out completely if the text is heavily rewritten or paraphrased afterwards.

The implications cut both ways. In theory, a functioning watermark is a more honest referee than any perplexity-based detector, it doesn’t care whether your sentences are long or your vocabulary is “quite” enduring, only whether Claude actually chose the words. If detection tools built on real watermarks replace today’s guesswork, writers may no longer need to flatten their own style to dodge a false accusation — the tool would finally be checking the right thing. Anthropic has also said a public detection API is coming, which OpenAI has not yet matched for text. But sceptics have already pointed out the gap, it only catches Claude, not the dozen other models writers might use, and the company itself concedes a full rewrite defeats it. It’s a genuine step toward accountability, not a solved problem.

A piece should sound like me. But does authenticity really mean I need to add a joke about my situationship in an article about Korean skincare, just to prove I wrote it? Susan Sontag once argued that interpretation is the revenge of the intellect upon art. In the age of AI, writers have started applying that same scrutiny to themselves — not “is this good?” but “does this sound machine-made?” A functioning watermark might, paradoxically, be the thing that lets us stop asking.

From ‘Is It Cake?’ to ‘Is It AI?’: Why Writers Are Deliberately Sounding Less Perfect

Suggestions

Articles

Topics

From ‘Is It Cake?’ to ‘Is It AI?’: Why Writers Are Deliberately Sounding Less Perfect

Related Stories

More From Opinions

Suggestions

Articles

Topics

Newsletter

The only newsletter you won't unsubscribe from

Follow Us