The short version: Most people think they can spot AI writing at a glance, and most people are wrong more often than they'd like. In an informal test I ran with roughly 300 people on LinkedIn last year, the average score was six out of ten, barely better than guessing. The tells are real but they're subtler than "too many em-dashes", and the uncomfortable truth is that well-edited AI writing now beats badly-edited human writing in blind tests almost every time.
Why I even started testing people on this
I read a lot of copy. Client drafts, guest post pitches, ghostwritten LinkedIn posts, the lot, probably 40 to 50 pieces a week between my own business and the consulting work I do. Somewhere around late 2024 I started noticing I couldn't call it as fast as I used to. I'd flag something as "obviously AI" and the writer would tell me they'd written every word themselves, just tired and rushing. And I'd flag something as "clearly human, look at the voice" and it turned out to be a first draft from ChatGPT with a paragraph moved around.
So I built a quick test. Five short passages, mixed sources, sent to my email list and posted on LinkedIn, no context given. People guessed human or AI for each one. I collected just over 300 responses. Average score: 6.1 out of 10, which is only marginally better than a coin flip across five questions (random guessing gives you 5 out of 10 on average). The people who scored highest weren't marketers, oddly. They were editors and English teachers, the ones whose actual job is reading closely for a living.
The test: can you tell which is which?
Here are five short passages. Read each one and decide: human-written, or AI-generated first draft. Answers and the reasoning are below, no peeking ahead.
- Passage 1: "In today's fast-paced digital landscape, businesses must adapt quickly to stay ahead of the competition. Leveraging the right tools and strategies can make all the difference in achieving sustainable growth."
- Passage 2: "I lost the Comet Print account because I sent the proof on a Friday and didn't check my inbox until Monday. Cost me eleven grand. I check email on weekends now, every hour, whether I like it or not."
- Passage 3: "Effective time management involves prioritizing tasks based on urgency and importance. By implementing strategies such as the Eisenhower Matrix, individuals can enhance productivity and reduce stress."
- Passage 4: "My nan used to say you can't polish a turd but you can roll it in glitter, and honestly that's most of what I do for a living now."
- Passage 5: "The quarterly figures show a 12% increase in customer retention, which we attribute to the new onboarding sequence launched in March. Two clients cancelled anyway, both citing price, so I'm not celebrating yet."
The answers
Passage 1 and Passage 3 are AI, both raw and unedited. Passage 2 and Passage 4 are human. Passage 5 is the tricky one: it's AI-drafted, then edited by a human who added the specific detail about the two clients cancelling. That last sentence is what makes it pass as human, and it's exactly the kind of edit that means most detection tools and most people get it wrong.
If you scored 4 or 5 out of 5, you're already reading more closely than most of the people I tested. If you scored 2 or 3, you're bang average, which is the whole point.
What gives it away (and what doesn't)
Forget the em-dash myth. Plenty of humans use them, including me, in this exact sentence. The real tells are structural, not cosmetic. Unedited AI writing tends to explain the thing rather than show it. "Effective time management involves prioritizing tasks" is a definition. "I lost the Comet Print account because I checked my inbox forty-eight hours too late" is a fact with a cost attached. Humans, when they're being honest, name the number, the client, the mistake. AI, left alone, stays at the level of the concept.
There's also a rhythm tell. AI-generated paragraphs tend to run at a similar length and land on a similar cadence, three clauses, a comma, a soft conclusion. Human writing, especially from someone who's tired or annoyed or excited, breaks rhythm constantly. Short sentence. Then a long one that keeps going because that's how the thought happened in their head. I wrote a full breakdown of the patterns that give AI writing away after reading client copy professionally for years, and the short version is: look for specificity and broken rhythm, not for buzzwords.
What doesn't work anymore: word lists. There was a stretch in 2023 and early 2024 where "delve" and "tapestry" and "in today's world" were reliable flags. They still show up in lazy output, but anyone using AI seriously now runs a second pass asking it to cut those words, and it complies instantly. If you're still scanning for "delve" as your main strategy, you're testing 2023's AI, not the one people are using now.
The bit most people writing about this won't say out loud
Here's the uncomfortable part. Detection tools like Originality.ai, GPTZero and Copyleaks get flagged for false positives constantly, and it's not evenly distributed. Studies on this have repeatedly found that non-native English writers get flagged as "AI-written" at far higher rates than native speakers, because their sentence patterns happen to be simpler and more evenly paced, which is also what unedited AI output looks like. A student writing carefully in their second language and a lazy AI draft can score almost identically on these tools. That should worry anyone using detection software to make hiring or grading decisions, and mostly it doesn't get mentioned because it complicates the tidy "just run it through a checker" advice.
The other uncomfortable part: in blind tests, heavily edited AI writing beats badly edited human writing most of the time. I've sat in on hiring reviews for freelance writers where the "obviously human" sample was AI output that had been rewritten by someone who knew what they were doing, and the "obviously AI" sample was a rushed human draft from someone talented but having a bad week. Origin isn't the thing that matters. Quality of the final edit is. If you're building your entire content vetting process around "can I spot the AI," you're solving the wrong problem. The right question is "is this specific, is it accurate, does it sound like one person with one point of view," regardless of what wrote the first draft.
How to get better at this, step by step
If you want your score to go from average to reliably good, here's what worked for me and for the editors who scored highest in my test:
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
- Step 1: Read the first line alone. Cover the rest. Does it make a specific claim or does it set a scene generically? "Quarterly figures show a 12% increase" is specific. "In today's competitive market" is not.
- Step 2: Count the concrete nouns. Names, numbers, dates, places. Human writing, when the writer lived the thing, tends to have one every two or three sentences. AI drafts without editing often go a full paragraph with none.
- Step 3: Look for the mistake. Real experience includes friction: something that went wrong, a number that disappointed, a plan that didn't work. AI defaults to smooth outcomes unless specifically prompted otherwise.
- Step 4: Check the rhythm out loud. Read it aloud for fifteen seconds. If every sentence is roughly the same length and shape, be suspicious. Human speech patterns are lumpy.
- Step 5: Ask what only this person could know. Not what's true in general, but what's true specifically to them: a client name, a dollar figure, a date, a regret. If you can't find one sentence like that in 300 words, treat it as unedited AI regardless of who claims to have written it.
- Step 6: Practice on your own drafts. Run something you wrote yourself six months ago and something AI generated on the same topic side by side. Most people are surprised how similar their tired first drafts look to AI output. That's not an insult, it's information.
None of this requires technical training. If you're building this skill because you want to write or edit for a living, you don't need a degree first, you can pick most of it up the same way I did, by reading a lot and comparing closely. There's a good starting point on where to learn copywriting for free as a beginner if you want structure around it rather than just trial and error.
Where this matters
This isn't a party trick. If you hire writers or editors for your business, this skill decides whether you're paying for the right thing. Clients often ask me whether they should pay editing rates or writing rates for someone cleaning up AI drafts, and the honest answer is it depends on how much rewriting is happening, which is a whole different job description. I covered exactly how those two roles differ in how freelance writing and editing jobs differ from each other, and the distinction gets murkier every year as more "writing" work is heavy editing of AI drafts.
If you're the one hiring, ask for a sample written live, in a call, on a topic you choose. Fifteen minutes, no AI tools open. It tells you more about someone's actual skill than any detector score ever will. And if you enjoy this kind of close pattern spotting generally, it's the same muscle used in logical reasoning and decision-making games, which is worth building deliberately if you run a business and need to spot problems before they're obvious; there's a good case for that in how logical puzzle games sharpen decision making for entrepreneurs.
My own rule now, after years of getting this wrong more often than I'd like to admit: I stopped asking "did AI write this" and started asking "would I bet money there's a real person and a real fact behind this sentence." That question is harder to game, and it's the one that protects the quality of what you're publishing.
Frequently asked questions
Can AI detection tools like Originality.ai or GPTZero be trusted?
They're useful as one signal, not a verdict. They get regularly caught flagging non-native English speakers and heavily edited human writing as AI, so treat a flagged result as a reason to look closer, not proof of anything on its own.
What's the single biggest tell that writing is unedited AI?
Lack of specific, personal detail. Real experience comes with names, numbers, dates and small failures attached. Unedited AI stays at the level of general concept because it has no lived experience to draw a specific detail from.
Does it matter if content was originally drafted by AI?
Less than most people assume. What matters to readers and to search engines alike is whether the final piece is accurate, specific and useful. A well-edited AI draft with real facts added beats a rushed, generic human draft almost every time.
How can I test my own skill at spotting AI writing?
Take five short pieces of writing you don't know the source of, guess human or AI for each, then check. Do this weekly with real samples from your inbox or from published articles, and track your score over a month. Most people improve noticeably after doing this ten or twelve times.
