Code Plagiarism Checker

About the Code Plagiarism Checker

Add 2 to 500 programs — one file, folder or ZIP file per student, in Python, JavaScript or TypeScript, Java, C or C++, C#, Go, PHP, Ruby or Kotlin — and see which of them share code. Each program is split into the tokens of its language, with comments and layout dropped and every variable, function, number and string name made the same, so a copy with new names, other comments, other indentation or its functions moved around still matches. A second, structural pass compares the programs’ statements in any order, which catches copies padded with extra lines.

The results are a ranked list of pairs with the share of each program found in the other, a similarity matrix, clusters of students who share code drawn as small graphs, and a side-by-side view that highlights every shared block in both programs. Starter code you handed out, import lines and code that many programs of a large class share are left out, so they never count. Everything runs in your browser: no program is uploaded. It compares only the programs you add, with each other — it does not search the internet or code sites and cannot tell whether code was written by AI.

Common uses

• Teachers and teaching assistants checking a class’s programming assignment for students who shared code.
• Checking resubmitted work against the earlier version, or a lab group’s programs against each other.
• Coding bootcamps and course staff screening exercise submissions before grading.
• Contest and hackathon organisers comparing entries with each other.

How it works

Each file is tokenised by its language’s rules: keywords, operators and the names of library functions such as print or System.out.println stay as they are; every name a student chose becomes the same placeholder (unless you switch that off), and every number, string and written-out list of data becomes another. Then every run of k tokens is hashed, and in each window of hashes the smallest is kept as a fingerprint. This is winnowing, from Schleimer, Wilkerson and Aiken’s paper “Winnowing: local algorithms for document fingerprinting” (ACM SIGMOD), the method behind Stanford’s MOSS service for programming classes; it guarantees that any shared run of at least t tokens is found. Each shared fingerprint is grown into the full shared run and the longest runs are taken first, so no token counts twice.

The share of a program is its tokens inside shared runs divided by all its compared tokens. The structural pass compares the programs’ statements as unordered collections (five tokens or longer, declarations left out) and pairs up functions whose statements are mostly the same, which shows functions moved around and copies with lines inserted between the real ones. A pair is ranked by the larger of its two shares, or by its share of statements in common when both programs have at least six statements and that is higher.

What is left out

• Comments, blank lines and layout, always, and Python docstrings.
• Boilerplate (on by default): import, include, using and package lines, and the line that opens the main program (public static void main(String[] args), int main(), if __name__ == "__main__":), which every program of a class has.
• Starter code: any run of tokens it shares with a program.
• Code many programs share (on by default, in a class of more than 10): runs found in more than 10 of the programs and in more than a tenth of them, such as a helper everyone copied from the lecture notes.

Left-out code is shown struck through in the side-by-side view, and a shared run never crosses it.

Reading the results

A high share is a reason to read the highlighted code, not proof that anyone copied. Short assignments have few good solutions, students who learn from the same examples write alike, and code from the course or a textbook is shared legitimately. Look at how long the shared runs are and whether they include the unusual parts — the odd variable layout, the same mistake, the same unneeded step. Use Strict sensitivity for long programs, where short shared runs happen by chance, and High for very short exercises.

Limitations

• It compares only the programs you add, with each other. It does not search the internet, code-sharing sites or a database, and it cannot tell whether code was written by AI.
• Programs are compared only with programs in the same language (JavaScript with TypeScript, C with C++). A program translated into another language is not found.
• Code that was restructured by hand — loops rewritten as recursion, logic moved into new functions — can share little at the token level. The structural pass helps with moved and padded code, not with rewritten algorithms.
• The lexers recognise each language’s comments, strings and operators but do not parse the program, so very unusual syntax (macros, PHP mixed deep into HTML) may be split less precisely.
• Up to 500 programs, 300 files and 1.5 million characters each. Files over 2 MB, binary files and folders such as node_modules, build or __pycache__ are left out. The side-by-side view draws the first 300,000 characters of each program; the reports cover all of it.
• The free preview compares the first 3 programs you add; a Premium pass (or unlocking this result) compares all of them.

Privacy

On this device; its files download once

Your programs are read and compared in your browser and never uploaded or stored. Only your option choices are remembered in this browser.