Why Prompt Instructions Cannot Reliably Lower Your AI Score

A prompt is a text instruction processed by an AI model. When you ask an AI to "rewrite this to sound more human" or "add burstiness and perplexity," the model generates new text based on that instruction. The new text still has word probability patterns, and those patterns are what Turnitin's classifier reads. The detector does not see your prompt. It only sees the output. Whether the output scores lower depends on whether its word probability profile differs enough from what the model learned to associate with AI-generated text, and you cannot control that by writing a better instruction. The model decides the word choices, and the model's word choices are what the detector is trained to recognize.

Turnitin's FAQ 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 detector is reading statistical features of word sequences in your text. It does not see the prompt you used. It does not know whether the text was generated, rewritten, or edited by a human. It sees the final word sequences and classifies them.

A common type of instruction tells the AI to "increase burstiness" or "vary perplexity." 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: "Instead, it learns statistical patterns from our training data."

And from the same page: "Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."

The model does not compute a metric called burstiness and check it against a threshold. So instructing an AI to "increase burstiness" is asking it to adjust something the detector may not be measuring in the way you think. What the detector does measure is word probability patterns learned from training data. Whether an AI generating text with "more burstiness" actually shifts those patterns in a favorable direction is not something you can verify by reading the output. Does Turnitin use perplexity and burstiness is the longer version of that question.

There is a more fundamental problem. If you use an AI to rewrite text that was flagged as AI-generated, you are producing text that was modified by an AI tool. Turnitin's FAQ says:

"Furthermore, 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."

The detector is not only looking for raw AI-generated text. It is also looking for text that was run through a paraphraser or bypasser tool. When you write a prompt instructing one AI to rewrite another AI's output, the result is functionally a paraphraser output. The detector is trained to identify that kind of text. That outcome has its own write-up: I paraphrased AI text with another AI and Turnitin still flagged it.

Turnitin's FAQ also lists text that is prone to false positives:

"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 next sentence: "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."

The third item is the one to notice. "Paraphrased without developing new ideas" describes what happens when an AI rewrites text by shuffling words without adding substance. The detector sees uniform structure and word patterns that resemble machine-generated paraphrase. So an AI rewrite that only changes surface wording while preserving the same structure can move the text closer to the false-positive profile, not further from it. The edits that actually change the statistics is what the alternative looks like by hand.

To summarize what we have covered:

• A prompt is processed by an AI model, which produces text with its own word probability patterns. The detector sees the output, not the instruction.
• The model does not compute burstiness or perplexity as named metrics, so instructing an AI to "adjust" them may not affect what the detector reads.
• Turnitin actively detects text modified by paraphraser and bypasser tools, which is what AI-rewriting-AI produces.
• Rewriting that only shuffles words without adding substance can move text closer to the false-positive profile.
• What changes the score is changing the word probability profile of the flagged passages, not writing a better prompt.

If you have a Turnitin report showing which passages were flagged, import it and work specifically on those passages. Eligible passages can be re-run at no charge.