Turnitin says it can detect some DeepSeek output, but only what its own pages state counts. Its English list, updated 2026-11-03, names DeepSeek V3.2. DeepSeek's current V4.1 Flash and V4 Pro are not on that list, and Turnitin does not process Chinese.
What Turnitin states about DeepSeek
- The English FAQ lists DeepSeek V3.2 (released 2025-12) among the models it "can detect content from". No other DeepSeek version appears in that list.
- Its Arabic launch blog, from 2026-08, lists DeepSeek R1, V3.1 and V3.2, and also names Qwen models.
- Turnitin says it is continuously updating its model to detect additional LLMs.
- The report does not name the model that wrote a passage. It gives an overall percentage and highlights sentences.
Turnitin's stated limits
- Languages. Long-form English, Spanish, Japanese, Modern Arabic, Dutch, German and Turkish. Paraphrase and humanizer detection is, per Turnitin, available only for English submissions.
- Length. At least 300 words of prose in a long-form writing format, up to 30,000 words. Below 300 words, Turnitin says the score may be less accurate.
- Content types. It does not reliably detect AI text in non-prose forms or code, and skips lists, headers, references, tables and formulas.
- Low scores. Scores of 1% to 19% are not shown as a number, because Turnitin says results there are less reliable and false positives are more likely.
- False positives. Turnitin states it aims to keep false positives under 1% for documents with over 20% AI writing, that it can miss some AI text, and that a document scored 50% could contain as much as 65% AI writing. These are Turnitin's own claims.
- Access. Inside Turnitin products such as Feedback Studio and LMS integrations, with no public paste box and no individual free tier.
Sources: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs and https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model.
What an independent second check adds
- Newer models. A model list lags releases. Veread, Fileverity's text model, reads writing style across the DeepSeek, Qwen and GLM families, and 94.9% of AI texts from models never used in its development were caught. See the fact sheet.
- Shorter texts. It works from 100 words.
- More languages. It covers English, French, German, Spanish, Chinese, Vietnamese and Arabic.
- A different view. It marks which passages read as AI-written and shows style similarity to DeepSeek and other families, a signal and not proof of authorship.
- Access without an institution. Anyone can run one free check on the DeepSeek AI detector.
Two tools that disagree are a reason to look closer, not to pick the louder one.
What to do with a flagged text
Treat any score as one piece of evidence next to drafts, sources and a conversation with the author. Style tells are listed in how to tell if DeepSeek wrote it, and both humans and AI use them.