Bypass AI detection? Why humanizing beats prompt tricks
Short answer Sometimes, against some detectors, for a while. Prompts like "vary your sentence length" target perplexity and burstiness, two signals GPTZero, which popularized them, stopped relying on in 2023, and Turnitin now flags humanizer output. Google doesn't rank by detector scores anyway. The edits that last add what only you know: facts, prices, examples, your voice.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callCan a prompt bypass AI detection?
Sometimes, briefly. A 2023 study found "simple prompting strategies" could "effectively bypass GPT detectors" (Liang et al., Patterns, 2023). But detectors get retrained, each one scores differently, and you rarely know which one a client or editor runs. A trick that passes today is a gamble tomorrow.
| Prompt trick | What it changes | Why it doesn't hold |
|---|---|---|
| "Vary sentence length" | Rhythm of the text | Targets burstiness, which GPTZero no longer relies on |
| "Use rare words" | Word predictability | Makes the text harder to read; the facts stay generic |
| "Add fragments and typos" | Surface style | Looks careless to a customer; the next detector update may catch it |
| Run it through a "humanizer" | Rewords every sentence | Turnitin now flags likely humanizer output |
None of these add anything a reader can use. That is the real problem with text that "sounds like AI".
What are perplexity and burstiness?
They are the two measures early detectors were built on. GPTZero, which popularized them, defines them this way:
"a measure of how likely an AI model would have chosen the exact same set of words as found in the document" (GPTZero, on perplexity)
"a measure of how much writing patterns and text perplexities vary over the entire document" (GPTZero, on burstiness)
In plain words: perplexity asks "how predictable is each next word?" Burstiness asks "does that predictability jump around, the way it does when a person writes?" A line like "We offer high-quality service at competitive prices" is very predictable. "We fixed 40 furnaces in January, mostly cracked heat exchangers" is not. Low perplexity and steady rhythm used to read as "AI".
Do detectors still look at perplexity and burstiness?
Not as the main test. GPTZero added this note to its own explainer:
"As of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture." (GPTZero)
The same page says the statistical signals remain "one of the seven 'indicators'" in its model. So prompting for "high burstiness" tunes one small input of a learned model. Models like this pick up patterns no one lists, which is why you can't write a prompt that reliably beats them.
Why do humanizers and paraphrasing tricks stop working?
Because it's an arms race, and the detector side keeps moving.
- Paraphrasing did work, on older detectors. One research team built a rewriting model that "drops detection accuracy of DetectGPT from 70.3% to 4.6%" (Krishna et al., 2023). Another found repeated rewording "can significantly reduce detection rates" across many detectors (Sadasivan et al., 2023).
- Detectors answered. Turnitin defines AI bypassers as "AI tools that rewrite AI-generated text and revise it with the intention of evading detection." In August 2025 it added bypasser detection (in its Originality add-on) and said it "will flag if a student has likely used an AI bypasser to 'humanize' their AI-generated content" (Turnitin, August 28, 2025).
- Detectors also misfire on people. The Sadasivan paper shows human text can be made to look AI-written, and the Liang study found detectors biased against non-native English writers. A "pass" or "fail" says little either way. See how accurate AI detectors are.
So a humanizer buys a score from one tool on one day. It doesn't buy a better page.
Does bypassing AI detection matter for Google?
No. Google judges the page, not the tool that typed it. Its spam policy targets content that "provides little to no value to users, no matter how it's created" (Google Search Central: Spam policies). Passing a detector doesn't protect a thin page, and failing one doesn't hurt a useful page. The full policy is in Does Google penalize AI content?
What Google does ask of AI drafts:
"It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." (Google Search Central: Guidance on using generative AI content)
How do I make AI writing sound human?
Edit for the reader, not the detector. The prompt that helps most is context: paste your own notes, prices and customer questions before asking for a draft. Then edit. Each of these also happens to make text less predictable, because generic text is what AI defaults to.
| Edit | Example for a roofer |
|---|---|
| Add a fact only you know | "Most hail claims we see in May are on 15-year-old shingles" |
| Replace vague claims with numbers | "Quotes in 48 hours" instead of "fast turnaround" |
| Cut the stock opening and closing | Delete "In today's fast-paced world" and "In conclusion" |
| Check every fact, price and rule | Fix the permit fee the draft guessed |
| Read it aloud in your own voice | Swap words you'd never say to a customer |
| Add your name and a real photo | A byline and a job-site picture, not stock |
For a repeatable way to keep your voice, see AI content creation that keeps your voice. For the trust signals Google names, see E-E-A-T for a small business.
Example
A cleaning company's blog
The owner prompts "write with high burstiness" and gets choppy posts that still say nothing specific. She switches to adding one customer question and one price range per post. The posts get shorter and more useful; no detector involved.
A consultant whose client runs a checker
A client's detector flags a report. Running it through a humanizer would only change words. He adds the client's own numbers and his recommendations, then tells the client he drafts with AI and checks every figure. The client cares about the second part.
A contractor's service pages
The AI draft reads smoothly but lists permit rules from another state. Fixing them, adding job photos and a local price range does more for customers and for Google than any detector score.
FAQ
Is it illegal to bypass AI detection?
Generally not. What usually applies is a school's, client's or publisher's own rules, so read your contract or their policy.
Do AI humanizer tools work?
Against some detectors, some of the time. Turnitin now flags likely humanizer use, and the tools change wording, not facts or experience.
What does "low perplexity" mean in my content?
Your words are easy to predict. That's usually a sign of generic phrasing, which readers skim past, whoever wrote it.
Should I tell clients I use AI?
If they asked, or your contract covers it, yes. Being upfront beats hoping a detector says "human".
Different for you?
- You want to know if detector scores affect rankings: AI detectors and SEO
- You worry Google will penalize AI pages: Does Google penalize AI content?
- Your AI drafts don't sound like you: AI content creation that keeps your voice
- You want a team to write and check your pages: SEO agent