Do AI Detectors Work on Paraphrased Text?

Here’s a thing people get wrong about AI detectors: they assume a detector reads your text the way a teacher does, hunting for clumsy ideas or robotic phrasing. It doesn’t. It measures a statistical fingerprint. And that single fact explains why paraphrasing an AI draft sometimes sails right past a detector and sometimes gets caught dead in the open.

So, do AI detectors detect paraphrased text? It depends entirely on what you mean by “paraphrased.” Swapping a few synonyms is a different act than actually rewriting, and detectors treat them very differently. There’s also a wrinkle most people miss: the big tools now run a dedicated check for AI-then-paraphrased text and report it as its own score, separate from plain AI-generated. So a spinner can light up a flag the original draft never would have. Let’s get specific about why.

Key takeaways

• Detectors don’t grade meaning. They measure statistical signals like perplexity and burstiness, so changing words while keeping the original structure usually leaves the fingerprint intact.
• Light paraphrase often still trips detectors. Synonym swaps barely move the math, and tools like Turnitin and GPTZero now flag AI-paraphrased text on purpose, in a separate score bucket.
• Heavy, genuine rewriting changes the statistics because you change the *structure*, not just the vocabulary.
• Scores are probabilities, not proof. Detectors throw false positives, and a Stanford study found they’re biased against non-native English writers.
• The honest goal isn’t permanent “evasion.” It’s writing that reads like a person because a person actually shaped it.

What a detector is actually looking at

Most AI detectors lean on two ideas. We dig into the full mechanics with real examples in how AI detection actually works; here’s the short version.

Perplexity measures how predictable your word choices are. Language models pick statistically likely next words. So when GPTZero explains its own approach, it uses the example of “Hi there, I am an AI ___” — a model almost always fills that blank with “assistant” (low perplexity, very predictable), while a human might write “detector” or something odder (higher perplexity). Predictable text reads as machine-like to the math.

Burstiness measures how much that predictability *varies* across the document. People write unevenly. A short sentence. Then a long, winding one that doubles back on itself before it lands. Models tend to hum along at a steady, even level of predictability, and low variation looks machine-made.

Notice what’s missing from both: meaning. A detector doesn’t know whether your argument is good. It’s looking at the *shape* of the prose — word probabilities and how they fluctuate. Hold onto that, because it’s the key to the whole paraphrasing question.

Why light paraphrase usually doesn’t help

Imagine you ask a model to write an intro and it hands you this:

> “Artificial intelligence has fundamentally transformed the way modern businesses operate, enabling unprecedented levels of efficiency and innovation across virtually every industry.”

Now you “paraphrase” it the lazy way — open a thesaurus in your head and swap words:

> “Artificial intelligence has radically changed how contemporary companies function, allowing remarkable levels of productivity and creativity across nearly every sector.”

Look at what changed. The vocabulary. Look at what didn’t. The sentence is still one long, balanced, evenly-paced unit. Same clause structure. Same rhythm. Same smooth, low-variation flow. “Transformed” became “changed,” “efficiency” became “productivity,” but the statistical fingerprint — predictable word patterns, steady burstiness — is basically untouched. To a detector, you handed it the same skeleton wearing a different shirt.

That’s why synonym-swapping is the weakest possible move. As GPTZero puts it, the tool doesn’t just check vocabulary; it analyzes sentence structure, rhythm, and statistical patterns. You changed the layer the detector cares about least.

The detectors started watching for this on purpose

This is where it gets pointed. Vendors noticed people were running AI text through spinners, so they built detection for exactly that — and they surface it as its own category.

In July 2024, Turnitin rolled out an AI paraphrasing detection feature. Its own description is blunt: it identifies text that was “likely AI-generated and then revised using an AI-paraphrasing tool or AI word spinner, such as Quillbot.” The report splits the score into two buckets — AI-generated only, and AI-generated-then-paraphrased — and highlights *where* a spinner was probably applied. That second bucket is the part to notice. A spinner doesn’t just fail to hide your text; it can manufacture a brand-new flag.

GPTZero’s Advanced Scan does something similar, aimed at catching text that was originally AI-generated and then rewritten, with sentence-level highlights that point at the suspect passages.

One caveat before you over-read those features: a vendor saying “we detect paraphrasing” is not the same as a guarantee. These remain probabilistic models that output a likelihood, and they misfire in both directions. Treat the new paraphrase score as one more signal, not a verdict.

Still, if your plan is “generate with a chatbot, run it through a spinner, submit,” understand that this is precisely the workflow the big detectors built features to catch. You’re not finding a loophole. You’re stepping on a tripwire they installed on purpose.

So what actually moves the needle? Real structural change

The reason *genuine* rewriting changes detector output is the same reason light paraphrase doesn’t: the statistics follow the structure.

When you actually rewrite something in your own voice, you don’t keep the sentence as a skeleton. You blow it up. You merge two sentences into one and split another into three. You drop a concrete example the model never had. You cut the throat-clearing. You start a sentence with “But.” You let your natural unevenness back in — that’s the burstiness returning, not because you’re gaming a metric, but because that’s what real human drafting looks like.

Take the bloated AI intro from earlier. Genuine rewrite:

> “Every company says AI changed everything. Mostly it changed two things: how fast they answer emails and how confidently they exaggerate in pitch decks. Here’s what actually shifted on the ground.”

That’s not a paraphrase of the original. The ideas got reorganized, a specific (slightly contrarian) observation got added, the rhythm went jagged, and a chunk of vague filler died. The perplexity climbed because the word choices are less predictable. The burstiness climbed because the sentences vary wildly in length and shape. The fingerprint genuinely changed — because the writing genuinely changed.

That’s the whole hinge of the question. Paraphrasing edits the clothes. Rewriting throws out the skeleton.

The honest catch: a lower score isn’t proof of anything

I’d be selling you something if I stopped there. Two cautions, both important.

First, detectors are probabilistic. They output a likelihood, not a verdict. Turnitin says so itself — its FAQ warns the model “may not always be accurate and may misidentify human-written, AI-generated, and AI-paraphrased text,” and shouldn’t be the sole basis for action against a student. These tools throw false positives at genuinely human writing too, especially clean, formal, or non-native English prose. A Stanford study (Liang et al., 2023) found detectors flagged more than half of human-written TOEFL essays as AI while rarely misfiring on US eighth-grade essays. If you want the full picture, we cover it in why detectors flag human writing as AI. The short version: a score is a signal, not a sentence.

Second, the whole category is shakier than the marketing suggests. OpenAI built its own AI-text classifier, then quietly pulled it in July 2023, citing a low rate of accuracy. By its own numbers it caught only about a quarter of AI text and falsely flagged about 9% of human text. When the company that makes the AI gives up on detecting it, that should reset your expectations about how authoritative any single score really is.

So don’t chase a number for its own sake. Chasing the number is what sends people spinning text in circles. The healthier question is whether the writing is actually yours.

A quick worked scenario

Say you drafted a discussion-board post with a chatbot and a detector flags it at 90% AI. You’ve got three roads:

1. Synonym swap. Run it through a spinner. Maybe the score dips to 70%, maybe it doesn’t budge, and a paraphrase-aware detector might now add a *second* flag for “AI-paraphrased.” You spent ten minutes and the post still doesn’t sound like you. Worst trade in the deck.

1. Surface human polish. Fix a few words by hand, add a contraction, call it done. Slightly better, still mostly the AI’s structure. The score wobbles unpredictably.

1. Rebuild it. Read the AI draft once for the ideas, then close it and write the post from scratch in your voice — your example from class, your actual opinion, your messy sentence lengths. Now the text is structurally yours. The score usually drops a lot, but more to the point, it *reads* like you, which is the part a human grader and a real reader will notice.

Only one of those three is real work. It’s also the only one that holds up when the detectors get smarter next quarter.

Where a humanizer fits (and where it doesn’t)

A good humanizer accelerates road #3, not road #1. It can break up monotonous rhythm, vary sentence length, and strip the obvious AI tells so you’re starting from something that already reads more naturally — then you bring your own facts, voice, and judgment on top. What it can’t do is turn skipping the thinking into a free lunch. We’re pretty direct about this in why you shouldn’t treat an AI humanizer like magic. A tool that changes structure beats a tool that changes synonyms. Neither replaces the part where the ideas are actually yours.

Frequently asked questions

Do AI detectors detect paraphrased text? Often, yes. If the paraphrase is light — mostly synonym swaps and small reorderings — the underlying statistical pattern stays close to the original, so the detector still flags it. Some tools now run a dedicated AI-paraphrasing check. Heavy rewriting that changes structure and rhythm is harder to score confidently, but no detector is reliable enough to treat its output as proof either way.

Does running AI text through Quillbot or a spinner beat detectors? Not dependably. Turnitin publicly added AI-paraphrasing detection in 2024 specifically to catch text that was AI-generated and then run through a spinner, and it reports that category separately. Word spinners change vocabulary while leaving sentence architecture intact, which is exactly what the statistics measure. You may move a score, but you’re playing whack-a-mole, not improving the writing.

Why does my own rewriting sometimes still get flagged as AI? Two reasons. First, if you only swapped words and kept the AI’s sentence shapes, the fingerprint barely changed. Second, detectors produce false positives on genuinely human writing too, especially clean, formal, or non-native English prose — the Stanford study found detectors misclassified over half of human-written TOEFL essays as AI. A flag is a signal, not a verdict.

What’s the difference between paraphrasing and genuinely rewriting? Paraphrasing keeps the original sentence as a skeleton and changes the clothes. Genuine rewriting throws out the skeleton: you merge or split sentences, reorder ideas, add a specific example, cut filler, and let your own rhythm take over. The first leaves the statistical pattern mostly intact. The second changes what the detector is actually measuring.

Are AI detector scores reliable enough to act on? No, not on their own. Major detectors output a probability, not a definitive judgment, and Turnitin itself warns its model can misidentify human, AI, and AI-paraphrased text and shouldn’t be the sole basis for action. OpenAI even shut down its own classifier in 2023 over low accuracy. Treat any score as one data point alongside human judgment.

The bottom line

Paraphrasing and rewriting feel like the same thing until you understand what a detector measures. Change the words and keep the bones, and the fingerprint follows you — and these days a paraphrase-aware tool might even tag the spin as its own offense. Change the bones, the structure, the rhythm, the actual thinking, and you’ve written something new. That’s the only version that holds up over time and reads like a person to the people grading it.

Want to see where a draft stands before you decide what to fix? Check a draft in the free tool and look at which sentences light up — then rewrite those for real.

Try it yourself

You don’t have to take a detector’s word for it. Paste your draft into PaperBleach’s free AI checker and humanizer to see how it scores — and make it read more like you actually wrote it.