ChatGPT undetectable content: why detector scores mislead
Short answer Not reliably, and a website doesn't need it. Detectors guess from how predictable the words are, from patterns a model was trained on, or from a hidden watermark. Their scores are estimates, not proof: rewording moves them, and plain human writing gets flagged. Google judges whether a page helps the reader, not who typed it.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callCan you make ChatGPT content undetectable?
Sometimes, for one tool, for a while. Research shows rewording AI text can push detection rates down a lot: one research team found its "recursive paraphrasing method can significantly reduce detection rates" (Sadasivan et al., 2023). But detectors update, each one scores differently, and nothing guarantees tomorrow's result.
For a business website, that chase misses the point. Google doesn't rank pages by who wrote them:
"Our focus on the quality of content, rather than how content is produced, is a useful guide that has helped us deliver reliable, high quality results to users for years." (Google Search Central Blog, February 8, 2023)
What Google does punish is mass-produced filler. Its spam policies describe "large amounts of unoriginal content that provides little to no value to users, no matter how it's created" (Google Search Central: Spam policies). A "humanized" page can still be exactly that. For what to do instead, see AI detectors and SEO. This page explains the scores themselves.
How do AI detectors score text?
There are four main methods. Most tools combine them and don't say how.
| Method | What it checks | Weak spot |
|---|---|---|
| Perplexity | How likely an AI model was to pick these exact words | Clear, simple writing looks "predictable" too |
| Burstiness | How much that predictability changes across the text | Short or uniform pages (a price list, an FAQ) look flat |
| Trained classifier | Patterns a model learned from labeled human and AI samples | Only as good as its training samples; a black box to you |
| Watermark | A hidden signal the AI company adds while writing | Only finds text from that company's model, and heavy rewrites weaken it |
Perplexity and burstiness. GPTZero, one of the first detectors, defines perplexity as "a measure of how likely an AI model would have chosen the exact same set of words as found in the document," and burstiness as how much "writing patterns and text perplexities vary over the entire document" (GPTZero). Low surprise, evenly spread, reads as AI. GPTZero itself says it stopped using them in autumn 2023, when it moved to a "deep-learning based architecture" (GPTZero).
Trained classifiers. Turnitin, the tool many schools use, runs a model that flags sentences it judges "likely generated from a Large Language Model" and also text changed by "an AI paraphraser tool or an AI bypasser tool" (Turnitin Guides). In other words, it is trained to catch humanizers too.
Watermarks. Google's SynthID changes the word-choice odds slightly while Gemini writes, leaving a pattern a matching detector can read (Google DeepMind: SynthID). Google lists the limits:
"Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language." (Introducing SynthID Text, October 23, 2024)
OpenAI developed a text watermark but, as of its August 2024 update, was still only considering it. It said rewording with another model or translation made it "trivial to circumvention by bad actors" (OpenAI). So no public tool can confirm a ChatGPT watermark today.
What does a score like "80% AI" mean?
It depends on the tool, and that is the first way scores mislead. Two tools can show the same number and mean different things:
- Share of the text. Turnitin's percentage is the amount of "qualifying text" (sentences in paragraphs of long-form writing) its model flagged (Turnitin Guides). 80% means most sentences were flagged, not that it is 80% sure.
- Confidence. Many other tools show how sure they are that the whole text is AI. 80% then means "probably," with a 1-in-5 chance it's wrong even by the tool's own measure.
Turnitin also hides low results. Its testing found "a higher incidence of false positives when the percentage is between 0 and 19," so since July 8, 2024 it shows an asterisk instead of a number below 20% (Turnitin Guides). It needs at least 300 words of prose to score at all, and it says its results "should not be used as the sole basis for adverse actions against a student" (Turnitin Guides). Most website pages, such as a service page or contact page, are shorter than that or aren't long prose.
Why do detector scores mislead?
Five reasons, each from the sources above:
- A score is a guess, not a record. Except for a watermark, no detector knows how the text was made. It compares wording to patterns.
- Plain writing gets flagged. Researchers testing several widely used detectors found they "consistently misclassify non-native English writing samples as AI-generated" and may "penalize writers with constrained linguistic expressions" (Liang et al., 2023). Short, clear sentences, the kind that read well on a phone, are what the method distrusts.
- Rewording moves the score both ways. The same study found "simple prompting strategies" could "effectively bypass GPT detectors" (Liang et al., 2023). If a prompt can flip the result, the result says little about the text.
- Small error rates add up. Even OpenAI noted that its watermark's low false-positive rate would still produce "a large number of total false positives" across lots of text (OpenAI). Say a tool wrongly flags 1 human page in 10: check 30 pages you wrote yourself and expect about 3 "AI" results.
- Tools disagree. Each uses its own method, training samples and cut-off. A page can pass one and fail another the same day.
Do AI humanizer tools work?
They can lower a score. That is all they are built to do. They swap words and reshuffle sentences; they don't add a fact, a price, a photo or a step from your own jobs, which is what Google's helpful-content guidance asks for (see AI content that Google likes). Turnitin says it now looks for bypasser-tool output specifically (Turnitin Guides), so a pass today may fail after the next update.
If a page needs a humanizer to feel human, the fix is to make it more yours, not to disguise it.
What should I do with AI-drafted pages instead?
| Instead of | Do this |
|---|---|
| Rewriting until a detector says "human" | Add one thing only your business knows: a real price range, a local rule, a common mistake you see |
| Paying for a humanizer | Have a person check every number, name and rule |
| Publishing 50 pages fast | Publish the 10 services customers actually ask about |
| Hiding that you used AI | Put a real person's name and role on the page; see E-E-A-T for a small business |
Example
A roofer's service pages
The owner drafts six service pages with ChatGPT, then adds his own photos, his warranty terms and which shingles he installs. A free detector says "72% AI." Nothing to fix: every page is specific and accurate, and the score has no place in Google's ranking rules.
A bookkeeper flagged for her own writing
A prospect pastes her 200-word "About" page into a detector and gets "AI generated." She wrote it herself in short, plain sentences. That is the false-positive pattern the Liang study describes, and at 200 words it is below Turnitin's 300-word minimum for a score at all.
A clinic that bought a humanizer
The office manager pays $20 a month to run blog posts through a humanizer. Scores drop; the posts get wordier and two treatment names come out wrong. The $20 and the hours would do more as one page answering the question the front desk hears every day.
FAQ
Can Google tell if my content was written by ChatGPT?
Google hasn't said it tries to, and its stated rule judges quality "rather than how content is produced" (Google Search Central Blog). For text, Google's SynthID watermark covers Gemini output, not ChatGPT.
Is GPTZero or Turnitin more accurate?
Their numbers aren't comparable: they measure different things. Turnitin reports a share of flagged sentences; many others report confidence. Neither tells you whether a page helps a customer.
Will paraphrasing ChatGPT text in Quillbot make it undetectable?
It may lower some scores, as research shows for paraphrasing in general (Sadasivan et al., 2023). Turnitin says it also detects text changed by AI paraphrasers (Turnitin Guides). Either way, it doesn't make the page more useful.
Should I run every page through an AI detector before publishing?
No. A score can't tell you if a page is accurate or helps a customer. Check the facts, add what only you know, and watch the page's clicks and calls.
Different for you?
- You want to know if detector scores touch rankings: AI detectors and SEO
- You worry Google penalizes AI pages: Does Google penalize AI content?
- You want a checklist for AI-drafted pages: AI content that Google likes
- You'd rather a team write and maintain your pages: SEO agent