The short version: You detect AI written content in your marketing by reading for missing specifics, checking rhythm and sentence-length variation, running it through two different detection tools rather than one, and asking whether a real person with real experience could have written that exact sentence. Detection tools help but none of them are reliable enough to use alone, and the bigger issue most businesses miss is that flat, generic AI content damages trust even when nobody formally “catches” it.
Why this is suddenly everyone’s problem
Two years ago I’d get the odd freelance draft that felt a bit stiff. Now I’d say close to a third of the content crossing my desk, from client blog drafts to LinkedIn posts written by agencies charging good money, has the same fingerprints. Not because writers got lazier. Because ChatGPT got good enough that it’s the fastest way to hit a word count, and word count is still what a lot of contracts pay for.
The problem isn’t that AI wrote a first draft. I use it myself for outlines and for pulling ChatGPT prompts for business content when I’m stuck. The problem is content going out under a brand name that reads like nobody who works there ever touched it. Readers can feel that even when they can’t name what’s wrong. That’s the real risk to detect, not the technical one.
The uncomfortable bit nobody wants to say out loud
Here’s what I’ll say that a lot of “how to spot AI content” posts won’t: AI detection tools are not reliable enough to make firing, publishing, or legal decisions on their own, and if you’re treating a 94% detection score as fact, you’re one false positive away from accusing a real human writer of cheating.
Turnitin’s own research, published when universities started panicking about ChatGPT essays, put their false positive rate at around 1% per sentence, which sounds small until you realise that compounds across a full document and disproportionately flags non-native English writers, because their sentence patterns are naturally more uniform. Originality.ai markets itself at 99%+ accuracy on GPT-generated text, but that figure drops noticeably on text that’s been lightly edited by a human afterward, which is exactly what most competent freelancers now do. So the tools are a signal, not a verdict. I use them as one input among several, never the only one.
The freelancer story that changed how I check things
Last year I was reviewing blog drafts for a client in the professional services space, a firm that prided itself on 30 years of hands-on case history. I’d hired a freelance writer through a referral, decent portfolio, reasonable rate. The first draft came back on time, clean grammar, hit the brief, 1,400 words.
Something felt off but I couldn’t name it for about ten minutes. Then I spotted it: every single case study reference in the piece was generic. “A client in the manufacturing sector faced significant challenges.” No name, no number, no year, no specific outcome, even though the brief I’d sent included three real case files with actual figures. The writer had skimmed my brief, fed the topic to ChatGPT, and never opened the case files at all. Every sentence was grammatically perfect and factually empty.
I ran it through Originality.ai (91% AI probability) and GPTZero (flagged as “likely AI”) but honestly, by that point the tools were confirming what a careful read had already told me. I sent it back with the three case files attached and a note: “I need the £340,000 saved on the Bristol project in paragraph four, not ‘significant savings.'” The rewrite that came back was still AI-assisted, I’d bet money on it, but now it had the specifics that made it usable. That’s the actual fix most of the time. Not rejecting AI content outright. Demanding the specifics back into it.
The nine tells I read for before I open a detection tool
- No proper nouns. Real experience has names, places, dates, exact figures. AI drafts default to “a company,” “many businesses,” “one client.”
- Rule of three everywhere. AI loves listing things in threes: “faster, smarter, and more efficient.” Real writers vary this without noticing.
- Perfectly even paragraph lengths. Humans write short punchy bits then a long one then a short one. AI drafts often sit at a suspiciously consistent four to six sentences per paragraph.
- Hedging without opinion. “It’s worth considering that this approach may offer certain advantages” instead of “this works” or “this doesn’t.”
- Summary sentences that restate the previous point. “In essence, this shows the importance of X” tacked onto the end of a paragraph that already made that point.
- Overuse of “boasts,” “landscape,” “ever-evolving,” “tapestry,” “testament to.” These show up in AI output at a rate no natural writer matches.
- Zero contradictions or caveats. Real experts say “this usually works but here’s when it fails.” AI drafts rarely volunteer the exception.
- Transitions that feel bolted on. “Moreover,” “furthermore,” and “additionally” stacked one after another, sentence after sentence.
- Advice with no cost, timeframe, or number attached. “Invest in quality content” instead of “budget £150 to £400 per 1,000-word article from a specialist writer.”
The step-by-step check I run
This is what I do with every piece of content that comes into my business from a freelancer, an agency, or a team member, before it goes anywhere near publish:
- Read it once for voice, not accuracy. Does it sound like a person who has done this work, or like a summary of the topic written by someone who read three articles about it?
- Check for at least three specific, checkable facts. A name, a number, a date, a real result. If there are none in a 1,000-word piece, that’s your first flag, no tool needed.
- Run it through two detectors, not one. I use Originality.ai and GPTZero together. If both flag it high, I trust that more than either alone. If they disagree wildly, I trust neither and go back to my own read.
- Paste a paragraph into a plain text editor and read it aloud. AI sentences often scan fine silently but sound flat or oddly formal spoken aloud. This one catches more than people expect.
- Ask the writer one specific question about the source of a claim. “Where did the 340,000 figure in paragraph four come from?” A person who did the research answers in seconds. A person who didn’t will stall or paraphrase the brief back at you.
- Check the structure against your own brand voice guide. If you don’t have one, this is the moment to write one, because “check it sounds like us” only works if “us” is defined somewhere.
What the detection tools measure
Worth understanding this before you trust a percentage: most detectors, including Originality.ai, Copyleaks, and GPTZero, work by measuring “perplexity” and “burstiness,” essentially how predictable each word choice is given the words before it, and how much sentence length and structure vary across a document. AI text tends to be low-perplexity (predictable word choices) and low-burstiness (consistent sentence rhythm). Human writing is messier on both counts. I’ve written more on how AI writing detection tools work if you want the full mechanics, because understanding the method changes how much weight you give the score.
That mechanic is also exactly why lightly edited AI content slips past detectors. Someone runs the draft through ChatGPT, then manually swaps ten words, shortens two sentences, and adds one personal aside. Perplexity and burstiness both shift enough to drop a 95% AI score down to 40% or lower, without the content gaining any substance. I’ve seen this happen in real time with a client’s marketing assistant who was proud of having “beaten” a detector, when what she’d really done was launder the same empty paragraph. The tool moved. The quality didn’t.
Where this shows up in different marketing content
Blog posts are the obvious one but this problem is everywhere now:
- LinkedIn posts with that unmistakable “I used to think X. Then Y happened. Now I believe Z.” structure, repeated across a dozen supposedly different founders’ pages in the same week.
- Email newsletters that open with “I hope this finds you well” energy even when the brand voice everywhere else is direct and a bit cheeky.
- Facebook group posts that read like ad copy dropped into a community space, which gets flagged and ignored by group members within seconds. If you’re posting into groups, it’s worth reading how to post content to a Facebook group without wasting the post, because AI-flat copy in that context gets you removed by admins, not just ignored by readers.
- Product descriptions that all “boast” the same three benefits in the same order across an entire catalogue, because someone fed a spreadsheet of product names into one prompt.
- Christmas and seasonal campaigns where AI-written copy and AI-generated visuals get bolted together without anyone checking they match the brand’s actual tone. If you’re using tools for seasonal creative, my piece on building festive photo prompts with ChatGPT covers where AI helps versus where it flattens everything into the same generic sleigh-bells stock photo look.
What to do when you find it
Don’t panic and don’t nuke it. Here’s the order I work through:
- Add the specifics back in yourself if it’s fast. A ten-minute fix beats a full rewrite most of the time. Swap “many businesses” for the actual client name and number.
- Send it back with exact instructions, not vague feedback. “This needs more detail” gets you another generic draft. “Tell me the exact outcome of the Bristol project in paragraph four” gets you a usable one.
- Have a direct conversation about process, not accusation. I ask writers straight out: “Did you use AI for this, and if so, what did you add on top of it?” Most people tell the truth if you ask like that instead of ambushing them with a detection score.
- Set the expectation before the next brief, not after. I now say explicitly in every freelance contract that AI-assisted drafts are fine but must include verified specifics from the source material I provide. That one sentence has fixed 90% of the issue for me going forward.
If you’re running content through a small team and this keeps happening despite clear briefs, it’s often less a writer problem and more a process gap, and it’s the kind of thing worth getting outside eyes on. I’ve walked several clients through fixing this exact workflow issue as part of working with an AI implementation coach, because the fix usually isn’t “ban AI,” it’s building a checklist like the one above into your actual editorial process so it happens every time, not just when something feels off.
The line I hold
I don’t care if a first draft came from ChatGPT. I care if the finished piece could only have been written by someone who did the work, ran the numbers, or lived the experience. That’s a much higher and much more honest bar than “did a detector flag this,” and it’s the one your readers are quietly applying whether they know it or not. If you want your own drafts to clear that bar, the same checklist above works in reverse. Before you check other people’s writing, run it against yours: could someone who’s never touched your business tell your paragraph apart from one written by a competitor using the same prompt? If not, that’s your first fix, and it’s worth reading through what makes AI generated writing sound human before you publish another piece under your own name.
One more small practical thing: whatever editor you draft in matters here too. A cluttered free tool with no readability scoring makes flat AI paragraphs harder to spot because everything looks the same on the page. I’ve written up which free content editor I use and which ones I’d skip, and the right one makes sentence-length variation, one of the biggest AI tells, visible at a glance.
Frequently asked questions
Can Google penalise my site for AI written content in marketing?
Google has stated publicly that it ranks content on quality and helpfulness, not on how it was produced, so AI content itself isn’t penalised. What gets penalised is thin, unhelpful, or duplicate content, which AI produces a lot of when it’s used without real facts added. Focus on specificity and usefulness, not on whether a bot could technically have written a sentence.
What’s the most reliable free tool to detect AI written content?
GPTZero offers a free tier that’s reasonably solid for a quick check, and pairing it with a manual read for specific facts and names gets you further than any single tool alone. No free or paid tool is reliable enough to use as your only check, so treat the score as a prompt to look closer, not a final answer.
Should I tell clients or readers when content was AI-assisted?
There’s no legal requirement in the UK or US to disclose AI assistance in standard marketing copy, unlike in some regulated areas like advertising claims or journalism. That said, I disclose it when a client asks directly, because trust matters more than a technicality, and most people care far less about the tool used than about whether the content helps them.
How do I stop freelancers from submitting AI content without checking?
Put it in the brief upfront: AI-assisted drafts are fine, but every claim, statistic, and case reference must come from source material you provide and must be checkable. Ask one specific factual question about their draft before approving payment. Writers who did the real work answer in seconds; writers who didn’t will stall.
