How to Tell If Something Was Written by AI (Spoiler: You Usually Can’t — and the Detectors Can’t Either)

That’s the stuff you can sometimes spot. And that’s the stuff that makes people hate the result.

This article is in two parts. One: what AI detectors and “tell” lists are actually good for. Two: a practical, human way to spot the kind of AI writing that wasn’t worth the writer’s time — and how to fix it so it doesn’t smell like a bot.

Why Ai Detectors Are A Bad Bet

Most AI detectors have false positives all the time. Run the US Constitution through a few of them; more than one will decide AI had a hand in it. Same with plenty of published novels and academic papers written years before GPT existed.

So: there is currently no AI detector you should trust enough to stake your reputation on an accusation. Use them as a loose signal at best. Don’t use them as proof.

What *does* help is knowing the writer’s usual style. If someone’s previous work reads one way and the new submission reads like a different person — smoother, more generic, more “blog post” — that’s a clue. It’s not proof of AI. It’s a reason to look closer.

How To Spot Lazy Ai Writing (The Human Method)

You’re not trying to “prove” AI. You’re trying to notice text that was clearly generated with low human involvement — because that’s what readers (and graders) react badly to.

Weird stuff

Look for out-of-place phrases that sound like instructions. Someone might have pasted a short prompt into the doc: “do this next,” “add more detail,” “continue.” That’s a dead giveaway.

Facts that don’t hold up

AI reads well but makes stuff up. If there’s a specific fact or detail, check it. Finding a mistake doesn’t prove AI — but it’s the kind of error that shows up a lot in unedited AI output.

ChatGPT’s default voice

The most obvious is “As an AI language model” — that’s the model literally saying it. But the default tone is long-winded and a bit like a lecturer. So you get a lot of:

• “It’s generally recommended”
• “It’s important to note”
• “Remember,”
• “In conclusion” / “Wrapping up” / “So there you have it.”

Sentence starters that show up in generic, un-styled AI essays:

Ah, … / Indeed, / Now, / See, / You see, / Moreover, / Furthermore, / However, / To sum it up, / When it comes to… / It might not be easy, / There you have it folks.

Fiction and flowery stuff

In stories, default AI leans into colorful metaphors — and then repeats the same kind (a dozen nautical references, or everything “like a” something). It tends toward overblown drama, instalove, and a certain juvenile sweep. It loves “in today’s digital era,” “a rich tapestry of,” “is a dance between,” and “plethora.” And it has this annoying positivity: “We’ll get through this… together!” or “Together we can face anything!” — or nonfiction sign-offs like “Stick around as we dive into more intricate facets. Happy writing, and remember, it’s all about the journey!” Nobody talks like that unless they’re selling something.

So: spotting this stuff points less to “AI” in some abstract sense and more to *very lazy use of AI*. With careful prompts and real editing, you often can’t tell. The goal isn’t to catch every AI user. It’s to catch the writing that nobody bothered to make human.

What About Turnitin And Detectors?

Students are often taught to write in that same safe, stilted, clean style — the kind that gets a B and sounds a lot like default AI. So “sounds like AI” and “sounds like average student writing” overlap a lot. The best, most distinctive college essays are an easy style for AI to copy too.

So as a measure of quality or “did they cheat?”, it’s going to be very hard to prove AI use unless someone leaves an obvious clue. You can compare to their previous work and see a big shift — but that’s not proof, and accusing someone is serious.

Ethically, the line is already blurry. Is it cheating to use Grammarly or ProWritingAid or Word’s AI to rephrase and polish? Different people will answer differently. The real question for teachers might be: do you care more about the final product or the process? If it’s process, you need different assignments — drafts, exercises, in-class work — not just a detector.

How To De-Ai Your Own Writing (If You Draft With Ai)

If you use AI as a draft tool and don’t want it to be obvious:

• Delete or heavily rewrite the first paragraph (and often the first 2–4 sentences). That’s where the model voice is strongest.
• Break the symmetry. Don’t do three neat points; do two and a weird aside. Vary paragraph length.
• Remove “It’s not X — it’s Y” entirely, or use it once at most.
• Swap a lot of smooth transitions for harder cuts. Cut “Additionally,” “Moreover,” “In conclusion.”
• Add one specific detail a model wouldn’t invent: a place, a sense, a petty opinion.
• Let one sentence be ugly: a fragment, parentheses, a bit of heat.
• Stop asking the reader what they think unless you actually want answers.
• Cut most of the hedges: “might,” “could,” “often,” “in many cases.”
• No “Great question!” or “You’re absolutely right!” openers. No “Happy to help!” or “What do you think?” closers.

That’s the fast checklist. There are longer lists (we keep one in-house for our humanizer tool) — but doing even half of this will take most of the default “AI smell” out.

Bottom Line

You can’t reliably prove something was written by AI. Detectors can’t either. What you *can* do is notice when writing sounds generic, over-structured, and barely touched by a human — and you can fix that kind of writing so it sounds like a person wrote it. That’s the game: not “detect AI with certainty,” but “make (or judge) writing that doesn’t feel like it was generated with low human involvement.”

If you want a one-page checklist you can paste into your editor or our humanizer tool, we’ve got that. It’s in the same place as our anti-slop rules: built from a ton of beta-reader flags and “this sounds like AI” complaints, so we could fix the same bullshit over and over in one place.