Here is a caption I could publish tomorrow:

"Golden hour in Oslo never disappoints. From vibrant cityscapes to serene fjords, this city offers something for everyone. Stay tuned for more captivating moments!"

And here is the one I would actually publish:

"Took the camera out after dinner because the light over the fjord turned that ridiculous orange again. Oslo does this maybe ten evenings a year. I dropped everything for it."

The first one is grammatically perfect and completely dead. You have read a thousand captions like it, and lately, you have probably suspected most of them were written by AI. You were probably right. But here is the uncomfortable part, and the reason I am writing this: I used to write like that long before ChatGPT existed.

The essays that exposed me

In 2018 I joined an IELTS workshop in the Philippines. My classmates were mostly nurses preparing to move to the UK, and I was the odd one out with an English degree, there because a diploma was not enough. The Philippines is an English speaking country, but English is still our second language, so we are asked to prove ourselves with a certificate anyway. I trained for months. I already knew my dashes (-) from my en dashes (–) from my em dashes (—). I had studied Halliday's frameworks at university, learned the inverted pyramid for news writing, tried literary writing, even wrote my own poems. I was, by every rubric, a trained writer.

Recently I dug up my old IELTS practice essays. Reading them now is genuinely uncomfortable, because they sound exactly like ChatGPT. The tidy thesis in the opening paragraph. The balanced transitions, firstly and furthermore and in conclusion, marching in formation. Every paragraph the same shape and length. Every claim delivered with the same polite, even confidence. High band score energy. No pulse whatsoever.

That is when something clicked. AI did not invent this voice. It learned it from us, from millions of exam essays, corporate blogs, and academic papers written to formula. Researchers at Stanford analyzed nearly a million scientific papers and found that words like realm, intricate, showcasing, and pivotal exploded in frequency after ChatGPT launched (Liang et al., 2024). In PubMed, the word delve appeared 349 times in 2020 and 2,847 times in 2023 (Stokel-Walker, 2025). The machine sounds like a well trained writer because well trained writers fed it. So when people ask how to stop sounding like AI, the honest answer is older than AI itself: stop writing to the template you were rewarded for in school.

The tells, according to research and my own ear

Working in digital marketing, I read and write captions for a living, and I have developed a sense for when something smells machine made. It turns out most of my instincts match what researchers have documented.

The word tells are the famous ones. Delve, realm, meticulous, pivotal, vibrant, "in today's fast paced world." One study found these words so contagious that they are now leaking into human speech. Researchers at the Max Planck Institute for Human Development analyzed over 360,000 YouTube videos and 771,000 podcast episodes and found people unconsciously adopting GPT vocabulary after 2022 (Brinkmann et al., 2025). We trained the machine on our words, and now it is training us back. Which means word lists alone will not save you, because the lists keep moving.

The rhythm tells matter more. AI text often has what researchers call low burstiness, meaning every sentence lands at roughly the same length with the same structure (Shammas, 2025). Human writing is messier. We ramble a little when an idea excites us. Then we stop short. My own observation, after months of working closely with these tools, is that the failure has two opposite flavors depending on the model and the prompt. Sometimes the output goes staccato, all short punchy lines stacked like bricks ("She opened the file. She read it twice. Something was off. She knew it."). Other times it turns weirdly poetic, reaching for word combinations that are grammatically fine but that no actual person would say out loud. Both come from the same place: the model is performing "good writing" instead of saying something.

Then there is the structure tell I see everywhere on LinkedIn: "It's not about X. It's about Y." Once you notice the negative parallel, you cannot unsee it.

And the biggest tell of all is absence. No city, no specific evening, no fjord turning ridiculous orange. Nothing only this writer could have written. Generic text is the natural output of a system that has no Tuesdays.

Why this matters if you do marketing

This would just be a writing nerd's hobby horse if readers did not care. They care. Around 62 percent of consumers say they are less likely to trust social media content they know was AI generated, and AI written emotional content can actively backfire on brand loyalty (SmythOS, 2026). Worse, once your audience starts suspecting AI, they stop absorbing your message and start hunting for evidence instead (PR Daily, 2026). The caption stops being a caption and becomes a forensic exhibit.

For a brand, that is a slow leak in the trust tank. For a freelancer or a job seeker, it is worse, because your words are the product.

The overcorrection is getting weird

There is now a counter trend, and you have probably seen it on LinkedIn. People writing everything in lowercase, even the word "i." Ellipses scattered everywhere... like this... for no reason. Deliberate typos left in as proof of life. The logic is simple: AI capitalizes correctly and spells perfectly, so imperfection must equal human.

(Fun fact, since we are here: English originally wrote the first person pronoun in lowercase. Back in Middle English, around Chaucer's era, the word shrank from "ich" to a single letter, and a lone little "i" kept getting visually lost on the handwritten page. Scribes started writing it tall so it would survive, and the capital stuck. So the capital I was never about ego. It was a legibility hack from the manuscript days.)

I understand the impulse, honestly. Nobody wants to be mistaken for a machine, and humanity is becoming almost allergic to its own polished writing. But typing in lowercase is not a voice. It is a costume, and a costume everyone is wearing stops working fast. The moment "imperfect" becomes the new template, AI can imitate that too, and then we are back where we started, just with worse spelling.

The real lesson in the curve, I think, is this: authenticity was never in the typography. It is in the specificity. A perfectly punctuated caption about the dinner you abandoned for the light over the fjord will always read more human than a lowercase "i" attached to a generic thought. You do not have to break your grammar to prove you exist. You just have to say something only you could say.

How I actually do it

I want to be clear about something: I am not anti AI. I use it daily. I built my own desktop assistant with it. I vibe coded the website you are reading this on. The problem is not using AI to write. The problem is publishing what AI writes without putting yourself back in.

My workflow is not complicated. I let AI draft when drafting is the boring part. Then I read the result out loud, and anywhere I stumble or hear a phrase I would never say, I rewrite in my own words. I hunt down the parallel structures and the marching transitions, survivors of my IELTS years, and break them. I add the one detail only I know, the dinner I abandoned, the friend who asked me to bring my camera, the specific ridiculous orange. And I prompt better in the first place, because vague prompts are how you get vibrant cityscapes that offer something for everyone.

Funny enough, becoming a more human writer in the AI age meant unlearning some of my formal training. The frameworks were never wrong. They taught me control. But control is only half of writing, and the other half is sounding like someone. Other writers might disagree with me on this, and that is fine. I would just rather read a slightly crooked sentence from a real person than a flawless paragraph from no one.

The machines learned to write from our most obedient writing. The least we can do is stop being obedient.