Does Turnitin Detect Mistral AI Text? What the Documentation Says

Turnitin maintains a published list of models whose output its AI writing detection model can identify. That list includes GPT, Gemini, Claude, LLaMA, Mistral, Deepseek, Nova, Grok, o1-mini, and tools based on these LLMs as well. LLaMA sits next to Mistral for the same reason both are open-weight, and does Turnitin detect LLaMA runs the same reading. The wording is straightforward: Turnitin's AI writing detection model for English submissions can detect content from these models.

The documentation also notes: "We will continue to expand our detection capabilities to other models in the future." So the list is not frozen. It grows as new models are released and the detection system is updated to recognize their output.

For Mistral specifically, the answer to "does Turnitin detect it?" is yes, it is on the list. But what "detect" means in practice involves a more nuanced process.

Understanding what happens when you submit a document helps explain why the results look the way they do. Here is the mechanism Turnitin describes:

"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 system does not scan for a single telltale feature. It segments your text, scores each segment for AI-likelihood, pools overlapping sentence scores, and aggregates everything into one document-level percentage.

The documentation is also explicit about what the model does not do. It does not rely on named metrics like burstiness or perplexity, a denial we unpack in does Turnitin use perplexity and burstiness to detect AI: "Our model is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions." Instead, the model learns statistical patterns from training data. The classifiers are trained to detect differences in word probability and are adept at the particular word probability sequences of human writers.

This means the detection is probabilistic, not deterministic. The score reflects the model's confidence that text resembles AI-generated patterns, not a hard binary judgment.

Being on the detectable list does not mean every AI-generated sentence gets caught. Turnitin prioritizes keeping false positives low, and that tradeoff has a cost on the other side.

"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. For example, if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing."

So a document submitted with Mistral-generated text might receive a score lower than the actual AI content percentage. The detector might identify 50% as AI-written when the true amount is closer to 65%. This is by design: the system leans toward not flagging human text, which means some AI text will pass through unflagged.

This gap matters if you are trying to understand why a document with known AI content received a lower-than-expected score. The arithmetic behind that 1% target is worked through in what a 1% false positive rate means.

The model that detects Mistral text today is not necessarily the same model that will be running next month. Turnitin's documentation acknowledges this:

"As we iterate and develop our model further to better detect newer LLMs, it is likely that our detection capabilities will also change, affecting the AI percentage.. However, for a submitted document, the AI percentage will change only if it's re-submitted again to be processed."

Two practical takeaways from this. First, the same document could receive different scores if resubmitted after a model update, which is why comparing two reports beats comparing two scores. Second, once a document has been processed, its reported score stays fixed unless it goes through the system again. If you are comparing scores across time, differences could reflect model updates rather than changes to the document itself.

Mistral is a family of models, not a single fixed product. Tools built on top of Mistral models, or fine-tuned versions of Mistral, are also covered by Turnitin's "tools based on these LLMs" language. The detection model does not identify Mistral by name in the report. It produces a single document-level AI writing score based on statistical patterns, with no breakdown by specific model.

This means a report will not say "30% Mistral, 10% GPT." It will say something like "40% AI-generated" as an aggregate score. The model list tells you which families of models the detector is trained to recognize, not which specific model produced which paragraph. What the report does put on the page is covered in how to read a Turnitin AI writing report.

If you are working with Mistral-generated text and need to understand how Turnitin will treat it, here is the summary:

• Yes, Mistral is detectable. It appears on the published model list, and tools based on Mistral are covered as well.
• Detection is probabilistic. The system segments text, scores segments for AI-likelihood, and aggregates scores into one percentage. It does not use burstiness or perplexity.
• Some AI text will be missed. To keep false positives under 1% for documents with over 20% AI writing, the detector may undercount the actual AI content.
• Scores can change. If the model is updated and the document is resubmitted, the score may shift. A score from a previous submission stays fixed until reprocessing.
• No model-level breakdown. The report shows an aggregate AI percentage, not per-model attribution.

Eligible passages can be re-run at no charge.