Does Turnitin Detect ChatGPT, Claude, Gemini and DeepSeek? (2026)

Yes. Turnitin can detect ChatGPT, Claude, Gemini and DeepSeek output, and unedited text from any of them is usually what its AI writing indicator is built to catch. Turnitin's English AI writing FAQ now names specific versions from OpenAI, Google, Anthropic, Meta, Mistral, DeepSeek, Amazon and xAI, "and tools based on these LLMs as well."

The honest caveats matter just as much. The score is a prediction, not proof. It needs at least 300 words of prose, it hides results under 20%, it is weaker on short and heavily edited text, and it sometimes flags human writing. Below is what Turnitin says about each model family at the time of writing, why one detector can catch many models, and where it misses.

Does Turnitin detect AI? What the 2026 model list says

Turnitin used to describe its detector in terms of GPT-3 and GPT-3.5, the models behind the first ChatGPT. Its FAQ says that was the starting point and that it has since expanded to "a large number of other LLMs." In July 2026 it also consolidated a multi-model ensemble into a single model, which it says keeps the document-level false positive rate under 1%. Here is the English list, grouped by vendor, as we read it at the time of writing.

Two things follow from the list. First, a model's absence does not mean it slips through. Turnitin trains on patterns, and text from an unlisted model can still look statistically similar to text from a listed one. Second, a model's presence does not mean every sentence it writes gets caught. Turnitin says it accepts missing some AI text to keep false positives low: its own example is that a document scored at 50% "could contain as much as 65% AI writing."

Does Turnitin detect ChatGPT?

ChatGPT is the case Turnitin has worked on longest, and it is the best covered in the list, from GPT-4o through GPT-5.4. So yes, Turnitin detects ChatGPT in the sense that raw ChatGPT prose submitted as an essay or paper is the core thing the model is trained to score. That includes the "ChatGPT 5" generation people search for: GPT-5 and its mini, nano and point releases are named explicitly.

What raw ChatGPT academic prose looks like to a detector is easy to show with a real run. We put an AI-drafted public health abstract through our own detector checks. All seven sentences were flagged, and all five detector models in the check voted AI.

Urban heat islands pose a significant and growing threat to public health, particularly among vulnerable populations such as the elderly and individuals with chronic illnesses. This study investigates the relationship between neighborhood-level green space coverage and heat-related emergency department visits in three mid-sized cities between 2015 and 2022. Utilizing a combination of satellite-derived land surface temperature data, municipal tree canopy inventories, and hospital administrative records, we constructed a comprehensive dataset encompassing 214 census tracts. These results underscore the critical importance of integrating green infrastructure into urban planning and climate adaptation strategies.

Real scan from our case run. Full case: the abstract case study linked below.

None of those sentences is wrong. They are just predictable in the way chatbot prose is predictable: "pose a significant and growing threat", "utilizing a combination of", "underscore the critical importance". The full before and after is in our abstract case study.

Does Turnitin detect Claude?

Yes, by Turnitin's own account. Seven Claude versions are on the list, from Claude 3 Haiku to Sonnet 4.6 and Opus 4.5. Claude has a reputation on Reddit for sounding "more human" than ChatGPT, and its default style is often less list-heavy. But a detector does not judge charm. It scores how predictable each stretch of text is relative to what it learned from human and machine writing, and polished academic prose from any strong model tends to land in the same predictable zone. If you used a Claude version released after the list was last updated, assume the same general behavior rather than a free pass.

Does Turnitin detect Gemini?

Yes. Gemini has seven entries, including the Gemini 3 previews and Gemini 3.1 Pro preview. That matters for students who use Gemini inside Google Docs or Workspace, because "tools based on these LLMs" are covered too. The same goes for Gemini-written text pasted into a document: where it was generated does not change how it reads.

Does Turnitin detect DeepSeek, Copilot or Grok?

DeepSeek V3.2 is named, and so are Grok 4.1, Mistral Large 3, two LLaMA versions and Amazon's Nova 2 Lite. Microsoft Copilot is not listed by name, but Copilot runs on OpenAI models, and Turnitin says it covers tools built on the models it lists. Reasoning models such as DeepSeek R1 are not named either. Treat that as "not confirmed by Turnitin", not as "undetectable".

Why one detector catches many models

Turnitin's FAQ gives the basic idea: large language models generate text by picking the next highly probable word, so their output is consistent and probable, while human writing is "inconsistent and idiosyncratic." Turnitin adds that its current transformer-based model does not compute named metrics like perplexity or burstiness directly. It learns many statistical patterns at once from its training data, which mixes AI text with authentic academic writing.

Different chatbots are trained on overlapping data and tuned toward the same goal: fluent, safe, well-organized answers. That pushes them toward similar word choices and similar rhythm. It is why a detector trained on several models generalizes to others, and why you will see the same stock phrases flagged whether they came from ChatGPT, Claude or Gemini. Our explainer on how AI detectors work covers the mechanics.

Does Turnitin detect ChatGPT if you paraphrase it?

Often, yes. Turnitin added detection of AI text run through word spinners in December 2023, and since July 2024 the report has a separate category for AI-generated text that was "AI-paraphrased", with QuillBot given as the example. In August 2025 Turnitin added detection of "AI bypasser tools", which it describes as tools that try to make AI text look human. Both features apply to English submissions only, according to its model release notes. Turnitin does not say which tools it detects.

Grammarly is a separate case. Turnitin says spelling, grammar and punctuation fixes from Grammarly or other grammar checkers were, in most cases, not flagged in its tests on human writing. Text produced by Grammarly's generative features, such as drafting or paraphrasing, "will likely be flagged." We cover humanizers specifically in can Turnitin detect humanized text.

Where Turnitin's AI detection falls short

• Short text. Below 300 words there is no score. In documents of a few hundred words, Turnitin says the prediction is mostly "all or nothing", so a short piece that mixes your writing with AI help can be flagged as entirely AI.
• Edited text. Independent tests have found that manual edits and machine paraphrasing lower detection accuracy across tools. A 2023 study of 14 tools by Weber-Wulff et al. is the most cited example.
• Human text that looks formulaic. Turnitin lists text with little structural variation, repetitive text and paraphrase that adds no new ideas as more likely to be misread. Formal academic prose and some non-native English writing fit that description. Turnitin says it included second-language writers in training to reduce bias; independent research on detector bias against non-native writers explains why the concern exists.
• Lists and non-prose. Bullet points, code and bibliographies are not scored as qualifying text.

Turnitin's position on all of this is consistent: the percentage "should not be used as the sole basis for action." If you are on the receiving end of a flag you did not earn, read what to do if you are falsely flagged as AI.

FAQ

Can Turnitin detect ChatGPT 5?

Turnitin's English FAQ names GPT-5, GPT-5-mini, GPT-5-nano and the GPT-5.1 to GPT-5.4 releases among the models it can detect.

Can students see their Turnitin AI score?

Not by default. Turnitin says the AI writing indicator is visible only to instructors and administrators, though an instructor can download and share the report.

Does Turnitin detect AI in other languages?

It scores long-form English, Spanish, Japanese and Modern Arabic, each with its own model. Paraphraser and bypasser detection is English only.

Can Turnitin tell which AI model wrote the text?

No. The report shows how much qualifying text was likely AI-generated or AI-modified, not which chatbot produced it.

Checking your own draft first

If you drafted with ChatGPT, Claude or Gemini in a way your course or journal allows, the practical question is which of your sentences still read as machine text. Academic Humanizer's AI Detection Report answers that sentence by sentence, checked against Turnitin, GPTZero, ZeroGPT, Copyleaks, Sapling and Originality.ai, and the Humanizer rewrites only the flagged sentences while leaving the rest exactly as you wrote them. No tool can promise a Turnitin result, so read the output, keep your own argument and sources, and follow the disclosure rules that apply to you. Our guide on how to avoid AI detection in academic writing covers the full workflow.