What MOSS misses

How MOSS works

MOSS reads each file as a stream of tokens, so names, comments and whitespace drop out. It hashes short runs of those tokens into fingerprints using winnowing, then reports the pairs of files that share the most fingerprints.

That is why renaming variables does nothing against MOSS. It is also why MOSS has these blind spots:

What the research says

None of this is our claim alone. MOSS's author says it plainly, and computer science researchers have measured each blind spot. In their own words:

1. “while Moss automatically detects program similarity, it has no way of knowing why codes are similar.”

   Alex Aiken, Stanford UniversityMOSS documentationRead it

   The same page adds that "it is a misuse of Moss to rely solely on the similarity scores."

2. “Mossad is effective at defeating four plagiarism detectors, including Moss and JPlag.”

   Breanna Devore-McDonald and Emery D. Berger, UMass AmherstOOPSLA 2020Read it

   Their tool rewrites a copied program automatically, in minutes, into variants that are "classified as no more suspicious than legitimate assignments."

3. “they are vulnerable to advanced obfuscation based on structural modification of program code that preserves the original program behavior.”

   Timur Sağlam and Larissa Schmid, Karlsruhe Institute of Technology2025Read it

   Written by the team behind JPlag, about plagiarism detectors in general, including AI-generated rewrites.

4. “a student using GPT-J ... can complete introductory level programming assignments without triggering suspicion from MOSS”

   Stella Biderman and Edward Raff2022Read it

   They also found the generated code "diverse in structure", so there is no single pattern for a similarity tool to look for.

5. “Existing detection methods, such as MOSS and JPlag (based on similarity metrics) and GPTzero (AI detection), have mixed success in identifying AI solutions.”

   Kevin Wang, Seth Akins, Abdallah Mohammed and Ramon LawrenceIEEE CSCI 2023Read it

   Measured across three courses: CS1, CS2 and databases.

6. “ChatGPT generates code non-deterministically, and thus, traditional similarity detectors might not suffice to detect AI-created code.”

   Muntasir Hoq, Thomas Price, Bita Akram and colleagues, NC State UniversitySIGCSE 2024Read it

   Models trained for the job did detect it: they report "accuracy above 90%" on CS1 submissions.

Rewritten copies

The example at the top uses four ordinary rewrites, each of which leaves the program's behavior unchanged:

• A counted for loop becomes an equivalent while loop.
• A compound assignment like total += n is expanded into total = total + n.
• A comparison is inverted and negated: n > m becomes not (value <= m).
• An indexed value is pulled into its own named variable.

None of these is sophisticated, and they no longer have to be done by hand. Devore-McDonald and Berger's Mossad does it automatically in minutes, and Sağlam and Schmid found detectors vulnerable to exactly this kind of behavior-preserving change, including rewrites done by AI. A person reading the two files side by side sees the same function immediately.

What Codequiry catches that MOSS cannot

A Codequiry check runs the same peer comparison MOSS does. Then it checks every file against GitHub, Stack Overflow, Gist, Software Heritage and the open web, and runs two AI detection models on it. These are the three cases a MOSS batch has nothing to say about.

Copied from Stack Overflow

A student finds an answer, renames everything, and submits it. Nobody else in the class used that answer, so MOSS has nothing to pair it with.

A Stack Overflow answer, renamed Same logic, same structure. Only the names changed.

Example

Student submissionbrackets.py

1def check_brackets(expr):2 match = {')': '(', ']': '[', '}': '{'}3 opened = []4 for c in expr:5 if c in '([{':6 opened.append(c)7 elif c in match:8 if not opened or opened.pop() != match[c]:9 return False10 return not opened

Stack Overflow answeranswer

1def is_balanced(s):2 pairs = {')': '(', ']': '[', '}': '{'}3 stack = []4 for ch in s:5 if ch in '([{':6 stack.append(ch)7 elif ch in pairs:8 if not stack or stack.pop() != pairs[ch]:9 return False10 return not stack

Classmates with this code: noneMOSS: no matchCodequiry: matched to Stack Overflow

Every name was changed (highlighted). Renaming does not fool a token comparison, but MOSS can only compare this file with the other files you sent, and nobody else in the class used this answer.

Copied from GitHub

Past students publish their homework, and the next class finds it. Unless you add every public repository to the batch yourself, MOSS never sees it.

Last year's homework, from a public GitHub repo Same messages, same odd log format, new names.

Example

Student submissionAccount.java

1public boolean withdraw(double amount) {2 if (amount <= 0) {3 System.out.println("Invalid amount entered.");4 return false;5 }6 if (amount > funds) {7 System.out.println("Insufficient funds. Balance: $" + funds);8 return false;9 }10 funds -= amount;11 log.add("W:" + amount);12 return true;13}

Public GitHub repositoryBankAccount.java

1public boolean withdraw(double amt) {2 if (amt <= 0) {3 System.out.println("Invalid amount entered.");4 return false;5 }6 if (amt > balance) {7 System.out.println("Insufficient funds. Balance: $" + balance);8 return false;9 }10 balance -= amt;11 history.add("W:" + amt);12 return true;13}

Classmates with this code: noneMOSS: no matchCodequiry: matched to GitHub

Only the names changed (highlighted). The exact error messages and the W: log entries are the kind of detail two people do not invent independently, and they point straight at the repository.

Two recent flags

Real flags from Codequiry. Student names, comments and strings are removed before anything is shown here.

What to do about it

You do not need to stop using MOSS. You need to stop treating its report as the whole answer, which is what the MOSS page itself says:

1. Keep the peer comparison, with the starter code passed as a base file.
2. Add a web check against GitHub, Stack Overflow and the web, and include earlier terms.
3. Add an AI check, and treat its result as a question for the student, not a verdict.
4. Read the top pairs yourself. Look for shared bugs, odd names, identical messages and dead code, which survive any rewrite.
5. Save the evidence before it expires.

Codequiry does steps 1 to 3 in one check and keeps every report. Comparing the free options instead? See MOSS vs Dolos vs JPlag, or read the full guide to detecting code plagiarism.

Frequently asked questions

Sources

1. Aiken, A. Moss: A System for Detecting Software Similarity. Stanford University. theory.stanford.edu/~aiken/moss
2. Schleimer, S., Wilkerson, D. S. and Aiken, A. Winnowing: Local Algorithms for Document Fingerprinting. SIGMOD 2003. PDF
3. Devore-McDonald, B. and Berger, E. D. Mossad: Defeating Software Plagiarism Detection. Proc. ACM Program. Lang. 4 (OOPSLA), 2020. doi:10.1145/3428206
4. Biderman, S. and Raff, E. Fooling MOSS Detection with Pretrained Language Models. 2022. arXiv:2201.07406
5. Wang, K., Akins, S., Mohammed, A. and Lawrence, R. Student Mastery or AI Deception? Analyzing ChatGPT's Assessment Proficiency and Evaluating Detection Strategies. IEEE CSCI 2023. arXiv:2311.16292
6. Hoq, M., Shi, Y., Leinonen, J., Babalola, D., Lynch, C., Price, T. and Akram, B. Detecting ChatGPT-Generated Code Submissions in a CS1 Course Using Machine Learning Models. SIGCSE 2024. doi:10.1145/3626252.3630826
7. Sağlam, T. and Schmid, L. Evaluating Software Plagiarism Detection in the Age of AI: Automated Obfuscation and Lessons for Academic Integrity. 2025. arXiv:2505.20158