Can Turnitin Detect Claude? (2026 Guide)

Quick answer: yes, according to Turnitin. Its AI writing FAQ lists Claude models among those its English detector "can detect," from Claude 3 Haiku through Claude Sonnet 4.6 and Claude Opus 4.5, "and tools based on these LLMs as well" (Turnitin FAQ, checked September 2026). What Turnitin doesn't publish is how often it catches Claude text, and no independent study has tested its current model on current Claude output, so any precise "Claude detection rate" you see online is a guess.

That matters if you use Claude because you heard it's "harder to detect" than ChatGPT or DeepSeek. The evidence for that idea is thin and dated: one small 2023 study, run on Claude 2, before Turnitin rebuilt its model. Below is what Turnitin actually documents, what the research shows, and what that means for you.

Why Claude Is Different From Other AI Models

Anthropic trains Claude partly with a method it calls Constitutional AI, in which "the only human oversight is provided through a list of rules or principles," and a model critiques and revises its own responses (Bai et al., 2022). People often describe the result as more hedged, more varied or less cliché-prone than other chatbots.

Whether that style changes Turnitin's score is a different question, and there's no published measurement that answers it. Turnitin describes its classifiers as "trained to detect these differences in word probability" between AI text, which tends to pick "the next highly probable words," and human writing, which "tends to be inconsistent and idiosyncratic" (Turnitin FAQ). A different surface style doesn't necessarily mean different word probabilities.

The one independent data point is old. In a study run in September and October 2023, seven detectors including Turnitin, taken together, correctly identified 17.7% of Claude 2 samples, compared with 23.9% for GPT-4 and 76.9% for Bard (Perkins et al.). That was five samples per chatbot, pooled across all seven detectors, and the authors warn their results "may not necessarily represent the current ability of any of the listed tools." It's a hint, not a Turnitin detection rate.

One real difference in 2026: a watermark

Claude now has something no detector score can replicate. Anthropic announced on 14 August 2026 that "future Claude models will generate text that contains a watermark," starting with models launched after 2 August 2026, with older models to follow (Anthropic). Checking it requires Anthropic's detection API, which is in private preview for eligible organizations "such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups." Anthropic says the watermark "doesn't identify anything to do with individual users."

It isn't part of Turnitin's AI score, and Anthropic contrasts it with detection services that look for "tells" in AI phrasing. What it means for students is covered in does Claude watermark text?

Detection Rates by Claude Model

Turnitin doesn't publish detection rates for Claude, or for any other model. What it publishes is a list of models its English detector "can detect," with release dates, and a promise to "continue to expand our detection capabilities to other models in the future" (Turnitin FAQ). Here's what that list says for each Claude tier.

Claude Opus

Turnitin's list names Claude Opus 4.5. It doesn't name other Opus versions individually. The list ends with "tools based on these LLMs," but Turnitin doesn't say how it treats a Claude release that isn't named, in either direction.

Claude Sonnet

Sonnet has the most entries: Claude Sonnet 3.5, Sonnet 3.7, Sonnet 4.5 and Sonnet 4.6 are all listed. As with Opus, versions that aren't named get no statement either way.

Claude Haiku

Turnitin lists Claude 3 Haiku and Claude Haiku 4.5. There's no evidence that Haiku's shorter, simpler output is treated differently. Turnitin's scoring rules apply to the document, not the model: it needs at least 300 words of prose, and it shows scores from 1% to 19% only as an asterisk.

Detection Rates Summary

Source: Turnitin FAQ, September 2026.

Two conditions apply to every row. The list is for English submissions; Turnitin also scores long-form Spanish, Japanese and Modern Arabic, but its detection of paraphrased or "bypassed" text is English only. And whatever the model, Turnitin says the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors."

Claude vs ChatGPT vs Gemini vs DeepSeek: Full Comparison

Here's how Claude sits alongside the other major families on Turnitin's list. The honest comparison is about coverage, because Turnitin publishes no detection rate for any of them.

Source: Turnitin FAQ, September 2026.

Turnitin's AI report doesn't say which model produced a passage either. It shows "an overall percentage of the document that AI writing tools or large language models may have generated or modified," with highlighted segments. So an instructor sees an AI percentage, not "Claude detected."

The only independent comparison across these families is the 2023 study above, where the seven detectors pooled caught Claude 2 text least often. It predates every model on Turnitin's current list and Turnitin's July 2026 move from "a multi-model ensemble into a single model." Rankings of which chatbot is "least detectable" that you'll find elsewhere rarely say what they tested, when, or how. For our take on that question, see which AI is hardest to detect.

Can You Prompt Claude to Avoid Detection?

It's a common question, so here's what's documented.

Turnitin says its detection includes "likely AI-generated content that may have been modified using a word spinner/AI paraphrasing or bypassing tool," and that it runs automatically on English submissions wherever the AI indicator is enabled (Turnitin FAQ). It announced this bypasser detection on 27 August 2025.

Older research did find that changing prompts moved detector scores. In 2023, Liang et al. found that asking ChatGPT to rewrite its own essays in more literary language sharply reduced detection by seven public detectors, which didn't include Turnitin. Perkins et al. found that some evasion prompts lowered detection across seven detectors, including Turnitin, but noted that texts made with one of the most effective techniques would "very likely receive poor marks in a real-world setting because of the high number of errors."

Those results are from 2023 models and 2023 detectors. None of them tells you what Turnitin's current model does with a Claude draft today.

Why Prompting Can't Fully Solve This

There are three problems with trying to prompt your way past a detector.

The detector reads the whole document. Turnitin splits a submission into overlapping segments and scores each one (Turnitin FAQ). Changing a few phrases doesn't change how probable the rest of the wording is. Our breakdown of what AI detectors look for explains the signals.

The target keeps moving. Turnitin says "as we iterate and develop our model further to better detect newer LLMs, it is likely that our detection capabilities will also change," and a document's percentage changes "only if it's re-submitted." A prompt that seemed to work on one version tells you nothing about the next.

It doesn't answer the real question. If your course doesn't allow AI-written work, a lower score doesn't make it allowed. If your course does allow some AI use, you don't need to hide it; you need to follow the policy and be able to show your process.

The SupWriter + Claude Workflow

If your course permits AI assistance, Claude can be genuinely useful for brainstorming, outlining, explaining a reading or giving feedback on your own draft. Here's a workflow that keeps you on the right side of the rules:

1. Read the policy first. What counts as acceptable AI use is set by your instructor and institution, not by Turnitin or any tool.
2. Use Claude for the parts you're allowed to. Keep the chat, so you can show what you asked and what you did with the answers.
3. Write in a document with version history. Google Docs keeps versions (Google help), and Word does for files stored in OneDrive or SharePoint (Microsoft support). Turnitin itself says its score "should not be used as the sole basis for action," and drafts are what settle disputes; see how to prove your writing is human.
4. If you rewrite AI-assisted text with a tool, disclose it where required. SupWriter's Claude humanizer rewrites drafts so they read more naturally, but results vary by detector, length and topic, and no tool can guarantee a Turnitin result. Keep your original draft.
5. Pre-check if you like, knowing its limits. Any AI detector you can run yourself uses a different model from Turnitin's, which students can't run, so its score is a hint, not a prediction (how we test).

If you're flagged anyway, remember that Turnitin's AI report is instructor-only but can be downloaded and shared, so you can ask to see it (can students see their Turnitin AI score?). It's also worth knowing what percentage of AI is acceptable on Turnitin, because scores under 20% are never shown as a number.

The Bottom Line

Turnitin says it detects Claude: its English model names Claude 3 Haiku, several Sonnet versions, Haiku 4.5 and Opus 4.5. It doesn't publish how often it succeeds, for Claude or anyone else, and the only independent comparison involving Claude dates from 2023. So nobody can honestly tell you Claude's "detection rate" on Turnitin today.

What is documented is enough to act on. Turnitin scores the share of your text it thinks is AI-written, only in English, Spanish, Japanese and Modern Arabic, only above 300 words, and warns instructors not to rely on the number alone. Claude's new watermark is a separate system that Turnitin doesn't use. If you use Claude, use it the way your course allows, keep your drafts, and let your process speak for you.