Why does paraphrasing make your Turnitin AI score go up?

Paraphrasing tools are built to cut your similarity score, and they do that by swapping your words for other words that mean the same thing. The replacement words tend to be high-frequency expressions, the kind of language that AI generators also default to. Turnitin's detector reads word-probability patterns, and when your rewritten text lands closer to that distribution, the AI percentage goes up. This is not a bug. It is the mechanism working as designed on a different target than the one you were aiming at.

The similarity score and the AI percentage are, in Turnitin's own words, "completely independent and do not influence each other." They measure different things, they are computed by different processes, and nothing you do to one is guaranteed to help the other. When they pull in opposite directions, the paraphrasing tool you used to fix the first one can walk you into the second.

Turnitin states this plainly in its detection FAQs: "The Similarity score and the AI writing detection percentage are completely independent and do not influence each other." The similarity score tells you how much of your text matches other text in the database. The AI percentage tells you how much of your text looks like it was written by a machine. These are not two views of the same property.

A paraphrasing tool is designed for the first target. It takes a sentence that matches something in the database and rewrites it so it no longer matches. That lowers your similarity score. But the tool achieves this by replacing your original word choices with alternatives, and those alternatives are drawn from the most common, most likely expressions for each meaning. Your original phrasing might have been idiosyncratic. The replacement is, by design, the opposite of idiosyncratic.

The AI percentage, meanwhile, is computed from something else entirely. We traced how that calculation works in what AI detectors measure, and the short version is that the detector slices your text into overlapping segments, gives each one a probability score, and pools those into a document-level number. The words the paraphrasing tool picked are now the words being scored. If those words happen to cluster in the high-probability zone, the number goes up.

People often say Turnitin's detector measures "perplexity" and "burstiness." That is not quite right. Turnitin's FAQ says: "Our model is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions." The next sentence in the same paragraph reads: "Instead, it learns statistical patterns from our training data."

But the same FAQ page, in a separate section, says this: "Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers." So the model is not computing a perplexity score and calling it that. It is, in its own words, reading word probability. Perplexity is a measure of word probability. The distinction matters because the first quote, quoted alone, sounds like the detector ignores word probability altogether. It does not. Why detectors disagree covers how different tools land on different numbers for the same text, and word-probability modeling is central to that.

What this means for paraphrasing is straightforward. Your original sentence had a word-probability profile shaped by your own vocabulary, your own habits, the specific things you happened to read before you wrote it. The paraphrased version has a profile shaped by whatever the tool selected, and the tool selects from the center of the distribution, not the edges. The detector, which is trained on word probability, sees the rewritten sentence as closer to its model of machine output. Not because you used AI, but because the statistical footprint of a paraphrased sentence looks more like one.

This is where it gets uncomfortable. Turnitin's FAQ lists the kinds of text that tend to produce false positives, and the description reads like a spec sheet for what a paraphrasing tool produces:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."

The sentence directly after that is the one that matters most: "If our indicator shows a higher amount of AI writing in such text, we advise you to take that into consideration when looking at the percentage indicated." Turnitin is telling instructors to discount high AI scores on text that matches this description.

A paraphrasing tool, by design, does not develop new ideas. It takes your existing ideas and rewords them. It also tends to reduce structural variation, because it smooths style across the document rather than introducing new sentence shapes. The tool is not trying to make your text more varied. It is trying to make it less matching. Those two goals do not align, and the second one pushes the text toward exactly the profile Turnitin says is prone to false positives.

We went deeper into what makes polished writing trigger detectors, and the mechanism there is the same one at work here. Smoothing is a statistical signal.

Turnitin states that it would rather miss AI text than falsely flag human writing. The FAQ says: "In order to maintain this low rate of 1% for false positives, there is a chance that we might miss some AI written text in a document. We're comfortable with that since we do not want to incorrectly highlight human-written text as AI-written."

That tilt is real, and it is the reason a fully human-written document has a low probability of being flagged. But it does not protect paraphrased text in the way people assume. The 1% false positive rate applies to text that does not match the false-positive profile. Text that does match the profile, the kind Turnitin itself describes in the quote above, is not the text the 1% figure was measured against. The detector is tilted toward caution in general, but the features that tilt triggers on are the same features a paraphrasing tool installs.

If your text was paraphrased by an AI-driven tool rather than a simple synonym swapper, there is a separate pathway. Turnitin says it "can also identify instances where AI-generated text may have been modified by AI paraphraser or bypasser (also called humanizers) tools to evade detection." That is not a false positive. That is a true positive on a different signal, and the detector has a named capability for it. Paraphrased with AI and still flagged walks through what happens when the tool you used was itself an AI model doing the rewriting.

There is a meaningful difference between a grammar checker that fixes your spelling and a generative tool that rewrites your sentences. Turnitin draws this line explicitly. About Grammarly, the FAQ says: "Based on tests we conducted on human-written documents with no AI-generated content in them, in most cases, changes made by Grammarly (free & premium) and/or other grammar-checking tools were not flagged as AI-written by our detector." The key qualifier is "in most cases." The exemption is not absolute.

And the FAQ continues: "Please note that this excludes content generated by Grammarly's generative AI-powered features, including draft generation, paraphrasing, summarizing, and other features. Content produced using these features will likely be flagged as AI-generated by our detector." Spelling and punctuation fixes are one category. Paraphrasing and rewriting are another. The second category is where the AI score climbs.

Turnitin does not publish a list of which paraphraser and bypasser tools it has been trained and tested on. The FAQ says: "However, to safeguard the integrity of our solution and its effectiveness in maintaining academic honesty, we're unable to disclose the names of these tools." So nobody outside Turnitin can tell you which specific tools are detected and which are not, and any claim that a named tool "beats" the detector is a claim nobody can verify against Turnitin's own testing. We do not make that claim, and we do not recommend trusting anyone who does.

The AI percentage goes down when the word-probability profile of your text moves away from the distribution the detector associates with machine output. That happens when you introduce vocabulary, sentence structures, and ideas that a generator would not have reached for on its own. Paraphrasing without adding new ideas does the opposite. It moves the profile toward the center.

If you have already run your text through a paraphrasing tool and the AI score went up, the path forward is not another paraphrasing pass. Another pass replaces one set of high-frequency words with another set of high-frequency words. The profile does not move. You need to rewrite the flagged passages with your own word choices, your own sentence shapes, and ideally your own additional analysis or examples. That is what shifts the probability distribution the detector reads.

The practical version of this, with the report in front of you, is in how to read a Turnitin AI writing report. The report shows you which segments are flagged, and those are the segments where your word choices need to stop being the average and start being yours.

HumanPen is a document rewriter, and the relevant design choice here is about what gets rewritten and what does not. You can upload the document together with its AI writing report, and only the passages the report flagged are rewritten. The rest of the document stays as you wrote it.

The reason that matters for this article is that a full-document paraphrasing pass is the worst-case scenario for the mechanism described above. It rewrites everything, including passages that were not flagged, and it does so by moving all of them toward the high-frequency center. A scoped rewrite touches only the segments the detector already flagged, and it does so with the goal of shifting their word-probability profile rather than smoothing them further.

Billing counts only the words actually rewritten. If a later report still flags passages, those can be re-run at no charge. We do not tell you what the next report will say, because nobody can. But the mechanism is the mechanism, and a scoped rewrite works with it rather than against it.

Does paraphrasing always increase the AI score? Not always, but often enough that the pattern has a name. Turnitin lists "text that has been paraphrased without developing new ideas" among the profiles prone to false positives. What that profile turns on is whether new analysis was added, not which editing operations produced the sentences. Text that stays inside the same ideas keeps a word-probability profile the detector has been trained to associate with AI output. Whether your specific score goes up depends on what your original text looked like and what the tool replaced it with.

I used Grammarly to fix spelling. Will that raise my AI score? Turnitin says that in most cases, changes made by Grammarly and other grammar-checking tools "were not flagged as AI-written by our detector." The qualifier "in most cases" is doing work in that sentence. If you used Grammarly's generative features like paraphrasing, summarizing, or draft generation, the FAQ says that content "will likely be flagged as AI-generated."

Why does my similarity score go down but my AI score goes up? Because they measure different things. Turnitin says the two scores are "completely independent and do not influence each other." Paraphrasing lowers similarity by removing matching text. The replacement words can raise the AI score because they tend to be high-frequency expressions, and the detector reads word-probability patterns that overlap with those of AI-generated text.

Can I just run the paraphrasing tool again to fix the AI score? Running it again replaces one set of common words with another set of common words. The word-probability profile does not move in a useful direction. What shifts the profile is rewriting the flagged passages with your own vocabulary, your own sentence structures, and your own additional ideas.

Does Turnitin detect which paraphrasing tool I used? Turnitin says it has been "trained and tested to detect leading paraphraser and bypasser tools" but does not disclose which ones. The FAQ says sharing a list "would make it easier for students to evade our system." Nobody outside Turnitin can verify whether a specific tool is on that list.