Making AI-assisted content read as human: what detectors flag
Short answer Detectors flag text that is too predictable: the most likely next word, sentence after sentence. Swapping words with a paraphraser doesn't fix that, and Turnitin says it catches common ones. What works is real editing: add facts only you know, check every claim, cut generic openings and closings, and write the way you explain jobs to customers.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callWhy do AI detectors flag text?
Because AI writing is predictable. A chatbot picks the most likely next word, again and again. People don't. Turnitin, a detector many schools run, explains its own model this way:
"Human writing, on the other hand, tends to be inconsistent and idiosyncratic, resulting in a low probability of picking the next word the human will use in the sequence." (Turnitin Guides: AI writing detection FAQs)
So a detector doesn't know who typed a page. It measures how expected each word is and guesses. That means some human writing gets flagged too. Turnitin lists what tends to trip it, even on human text:
| What gets flagged | What it looks like on a business page |
|---|---|
| Lists without much variation | A services paragraph where every sentence starts "We provide..." |
| Text that repeats itself | The same promise in the intro, middle and end |
| Paraphrase with no new idea | A rewritten competitor page |
| First and last few sentences | "In today's fast-paced world..." openers, "In conclusion..." closers |
Turnitin also says it only scores prose of at least 300 words, ignores bullet points, and on short pieces the call is "mostly 'all or nothing'" (Turnitin Guides). It shows no score between 1% and 19%, to avoid false positives, and it says its score "should not be used as the sole basis for action" (Turnitin Guides).
Plain, careful writing can look "too predictable" as well. A Stanford study found popular detectors "consistently misclassify non-native English writing samples as AI-generated" (Liang et al., Patterns, 2023). If English is your second language, a flag says little about who wrote your page.
Do paraphrasers and humanizer tools get past detectors?
Not reliably, and they make the page worse. Turnitin says it tested common word spinners and can still tell:
"We also tested our model on some of the more easily available and popular text spinners such as Quillbot, Grammarly (free paraphraser), and Scribbr and our model is able to identify AI-generated text as likely paraphrased when run through these tools." (Turnitin Guides)
A spinner changes words, not substance. The page still says what every other page says, now in odder words. Grammar fixes are a different story: Turnitin says spelling and grammar edits from Grammarly were in most cases not flagged, but its writing and paraphrasing features likely will be (Turnitin Guides).
How do I edit an AI draft so it sounds like me?
Give it what no model has: your work. In order:
- Add three facts only you know. Your price range, how long a typical job takes, the mistake you fix most often. A roofer: "Most hail claims we see in May are on 15-year-old shingles."
- Check every number, rule and name. Google's guidance says it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing" (Google Search Central). Models predict words; they don't look things up.
- Delete the first and last paragraph. AI openers and summaries are the most generic part, and Turnitin says those sentences drew more false positives. Start with the answer.
- Say it the way you say it on the phone. Read each paragraph aloud. If you'd never say a sentence to a customer, rewrite it in your words.
- Cut repeats. One promise, once, with proof (a photo, a review, a license number).
- Turn generic lists into specifics. "Drain cleaning" becomes "Drain cleaning, including root-cut sewer lines in older homes."
- Put a real name on it. Google's checklist asks: "Do pages carry a byline, where one might be expected?" (Google Search Central).
Google's own self-check is a good final test:
"Does your content clearly demonstrate first-hand expertise and a depth of knowledge (for example, expertise that comes from having actually used a product or service, or visiting a place)?" (Google Search Central)
If the draft could run on any competitor's site unchanged, it isn't done. Prompting the model with your voice up front saves editing time; see ChatGPT prompts that keep your brand voice.
Does a detector score matter for my website?
Not for Google. Google judges whether the page is helpful and reliable; it says nothing about third-party detector scores; see AI detectors and SEO and Does Google penalize AI content?. Where a score can matter is a person: a client, editor or publisher who runs your text through a checker. The edits above answer them too, because they make the page yours.
Should I say I used AI?
Where readers would wonder. Google says disclosures "are useful for content where someone might think 'How was this created?'" (Google Search Central). A service page usually doesn't need a label. A guest article, a client report or anything with a contract that mentions AI use is different: follow what the client or publisher asked for.
Example
A plumber's service page
The AI draft opens with "Plumbing issues can be stressful" and lists eight services. The owner deletes the opener, starts with "Same-day repairs in most of the county," adds his usual price range for a water heater swap and two photos. The page now says things no other plumber's page says.
A bookkeeper writing for a client newsletter
The client's editor runs submissions through a checker. Her AI draft scores high. Instead of a humanizer, she adds a story from tax season, fixes one outdated filing date the model got wrong, and cuts the summary. It reads like her because now it is.
A contractor whose own writing gets flagged
He wrote his "About" page himself in short, plain sentences. A client's free checker says "AI." Short, simple text is exactly where detectors are weakest. His proof is the content: job photos, permit numbers, his name.
FAQ
Can a detector prove I used AI?
No. Turnitin itself says its score shouldn't be the only basis for action, and studies show human text gets flagged. It's a probability, not proof.
Does Grammarly make my writing look like AI?
Spelling and grammar fixes mostly don't, per Turnitin's tests. Its generate and paraphrase features likely do (Turnitin Guides).
Will asking ChatGPT to "write like a human" fix it?
It may change the score, but it adds no facts. The Stanford team found simple prompts could get text past detectors, which is one more reason the scores mean little (Liang et al., 2023).
How long should I spend editing an AI draft?
As long as it takes to add your facts and check every claim. For a service page, that's usually the owner's knowledge, not hours of rewording.
Different for you?
- Worried Google will penalize AI pages: Does Google penalize AI content?
- Want to know what detector scores mean for SEO: AI detectors and SEO
- Want drafts that start closer to your voice: ChatGPT prompts that keep your brand voice
- Want to know which human parts Google rewards: AI content plus human input
- Want a team to write and keep your pages accurate: SEO agent