How Turnitin Works: Plagiarism Detection and AI Detection

TL;DR

Turnitin compares submitted text against a corpus of web content, published literature, and previously submitted student work. The result is a similarity score and a detailed match report. AI detection was added in 2023 but reliability is limited. Used at most major Anglophone universities; widely deployed globally.

How the core plagiarism detection works

The corpus

Turnitin maintains a large corpus of text used for matching:

• Internet content — web crawl covering public web pages
• Academic publications — partnerships with publishers (Elsevier, Springer, Wiley, etc.) provide access to published literature
• Student submissions — papers submitted by students at institutions using Turnitin are added to the corpus (with options to opt out per institution)
• Other licensed databases — newspapers, magazines, e-books

The corpus size is large — tens of billions of documents. Institutional client base contributes to corpus growth.

Matching process

When a student submits a paper, Turnitin:

1. Processes the document (extracts text, normalises formatting)
2. Compares against the corpus using phrase-level similarity matching
3. Identifies matched passages and their sources
4. Generates a similarity report

The similarity report

The report shows:

• Overall similarity percentage (e.g. "23% similarity")
• Individual matches highlighted in the text
• Source documents for each match
• Filter options to exclude quoted text, bibliographies, small matches

Instructor interpretation

The similarity percentage alone does not indicate plagiarism. Instructors interpret the report:

• 23% with most matches being properly quoted and cited: usually fine
• 23% with the same matches being uncredited copying: misconduct
• 5% with one large uncited passage: misconduct
• 50% but all properly attributed: fine

The interpretation step is critical. Turnitin generates evidence; humans determine whether the evidence indicates misconduct.

AI detection — added 2023

Following ChatGPT's late 2022 launch, Turnitin developed AI-content detection capability:

• Approach: statistical analysis of text features that distinguish AI-generated from human-written text. AI text has different word-distribution patterns, sentence-structure variation, and idiomatic characteristics.
• Output: a percentage estimate of AI-generated content in the submission
• Limitations: false positives (especially for non-native English speakers, who can produce text with patterns that flag as AI) and false negatives (lightly edited AI text and longer-form AI text often pass undetected)

Scarfe et al. (2024)

The University of Reading study (Scarfe, P., et al., 2024) submitted AI-generated work through normal coursework channels at Reading. 94% of submissions went undetected — meaning the combination of human review and automated detection caught only 6%. The result indicates that current AI detection technology is well below what would be needed for reliable misconduct prevention.

The detection-evasion dynamic

AI submissions resist detection through:

• Light editing: students editing AI text reduce its statistical signatures
• Iterative prompting: students using AI to generate text from extensive prompts produce more human-like output
• Paraphrasing: AI-generated content paraphrased by other AI or by the student passes more detection
• Hybrid drafts: combining AI and human writing produces text without clear statistical signatures

Detection vendors are iterating; the cat-and-mouse dynamic is structural.

Deployment globally

The AMI's R-Score Detection tools sub-component reflects deployment scope. The highest scoring countries:

• UK (R_det=90)
• Australia (R_det=85)
• US (R_det=80)
• Ireland (R_det=75)
• Canada (R_det=75)
• New Zealand (R_det=70)

These are Anglophone countries where Turnitin has near-universal university adoption. AI detection has been rolled out alongside the existing plagiarism detection capability.

Language coverage and alternatives

Turnitin's language coverage

• Strong: English (largest corpus)
• Good: Spanish, French, German, Italian, Portuguese, Polish
• Limited: many less-resourced languages

Language-specific alternatives

Some countries operate domestic detection systems:

• Antiplagiat (Russia) — Russian-language detection
• CopyKiller (South Korea) — Korean-language detection
• JSA (Poland) — Polish-language detection, mandatory for theses
• Compilatio (France) — French-language detection
• PlagScan — German-language detection, now part of Turnitin

These systems often complement rather than replace Turnitin, with institutions running both for different language documents.

Strengths and limitations

What Turnitin catches well

• Direct copying from publicly accessible web sources
• Direct copying from major published literature
• Cross-student copying within institutional and inter-institutional corpora
• Self-plagiarism (with appropriate corpus settings)

What Turnitin misses or struggles with

• Patchwriting and heavy paraphrasing
• Translation plagiarism (copying from foreign-language sources)
• Contract cheating (the original work is not in the corpus)
• AI-generated content (currently)
• Recently published content not yet indexed

Inherent limits

Turnitin can only match against its corpus. Work copied from sources Turnitin does not have access to (proprietary databases, recently-written content not yet indexed, private documents) cannot be matched. Contract cheating produces "original" text that Turnitin has never seen — making it Turnitin-invisible by design.

Sources

• Turnitin product documentation
• Scarfe, P., et al. (2024). University of Reading AI submission study
• AMI v1.5 methodology document
• Vendor and corpus partnership documentation

Full methodology | Download dataset

Related

Read the full methodology