Paraphrasing Lowered Similarity but AI Score Went Up: Why It Happens

The Similarity score and the AI writing detection score are separate measurements that do not influence each other. Turnitin states this directly: "The Similarity score and the AI writing detection percentage are completely independent and do not influence each other."

When you paraphrase to bring the similarity percentage down, there is no mechanism that also pulls the AI score down. The two numbers can move in opposite directions, and often do. Paraphrasing changes the words in your document, reducing string matches in the similarity database. Those same changes can alter word probability patterns in ways that make the AI detector more likely to flag the text. The result is the scenario many students encounter: similarity goes down, AI score goes up.

To understand why this happens, we need to look at what each system measures. The similarity report compares your submitted text against a database of existing content. As Turnitin explains, "The Similarity score indicates the percentage of matching-text found in the submitted document when compared to Turnitin's comprehensive collection of content for similarity checking." Paraphrasing reduces these matches because you are changing the strings the system searches for.

The AI writing detector does something fundamentally different. It does not search for matching text. Instead, it analyzes the statistical patterns of your word choices. Here is how Turnitin describes the process: "When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated. Each qualifying sentence within these segments inherits the segment's score. Since segments overlap, some sentences may have multiple scores, which are then pooled into a single score. These sentence scores are further aggregated and used to compute the overall document AI writing score."

The similarity system looks for strings that already exist elsewhere. The AI system looks for probability patterns in how words are chosen. Paraphrasing changes the strings (helping similarity) but also changes the probability patterns (which may hurt the AI score). These are two different targets, and hitting one does not mean you hit the other.

Here is where the situation gets particularly frustrating. Turnitin's own documentation explicitly lists paraphrasing as a characteristic of text that can produce false positives. The full statement reads: "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."

Turnitin follows this with the advice: "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."

This is the core of the problem. When you paraphrase to lower similarity, the most common approach is to swap words for synonyms and rearrange sentence structure. This is exactly the kind of "paraphrased without developing new ideas" text that Turnitin flags as prone to false positives. You are not adding new analysis, new arguments, or new evidence. You are replacing words with other words that mean the same thing. The ideas stay the same, the structure often stays similar, and the text now fits a profile the AI detector associates with machine-generated content. The changes you made to avoid similarity matches may have produced text that looks statistically more like AI output to the detection model, not less.

A common question is whether the AI detector measures things like "perplexity" or "burstiness," terms that circulate in discussions about AI detection. Turnitin addresses this directly: "Our model is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions." In the next sentence, they add: "Instead, it learns statistical patterns from our training data."

This does not mean word probability is irrelevant. The same documentation continues with a critical clarification: "Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."

Together, these two statements tell us something important. The model does not compute perplexity as a named metric, but it is trained on word probability patterns. When you paraphrase by choosing the most obvious synonym for each word, you may be selecting the most statistically predictable word in that position. If the original word was an unusual choice and the replacement is the most common choice, the word probability pattern becomes more predictable, not less. To a classifier trained on word probability, more predictable sequences can look more like AI-generated text. This is the mechanism behind the paradox: paraphrasing makes text less similar to existing sources, but the specific word choices involved can make the text more statistically predictable, which is the opposite of what lowers an AI score.

The practical takeaway is that lowering similarity and lowering an AI score are two different goals. They require different changes to text, and the changes that help one can hurt the other. Paraphrasing alone, especially synonym swapping without new ideas, is the kind of approach that lowers similarity while potentially raising the AI score.

The similarity report is about whether your strings match existing content. The AI report is about whether your word probability patterns resemble machine-generated text. Trying to solve both with the same editing pass, particularly a pass that stays inside the ideas already on the page, is where the conflict arises. If your AI score went up after paraphrasing for similarity, the cause is likely that the paraphrasing changed your text in ways the AI detector reads as more predictable, while also fitting the false-positive profile of "paraphrased without developing new ideas."

Here is what to take away from this:

1. The similarity score and the AI score are independent measurements. Lowering one does not lower the other.
2. The similarity system searches for matching strings. The AI system analyzes word probability patterns. These are different signals.
3. "Paraphrased without developing new ideas" is explicitly listed by Turnitin as a false-positive trigger. Synonym-only paraphrasing fits this profile.
4. The AI model is not programmed for burstiness or perplexity as named metrics, but it is trained on word probability. Picking predictable synonyms can make text look more machine-generated.
5. Lowering similarity and lowering AI score are separate goals. Paraphrasing alone can help one and hurt the other.

When eligible passages can be re-run at no charge, you have room to iterate. The key is to treat the two scores as separate problems that need separate approaches.