New Breakthrough AI Detection. Our best yet for AI-written source code. Catches 90% of AI/GPT code with 1.3% false flags. 90% caught, 1.3% false flags. Read more Read more

AI assistants Copilot, Cursor and in-editor AI

Did Copilot or Cursor write this assignment?

AI coding assistants write code inside the editor, so there is no chat window to catch and nothing pasted from a browser. Codequiry reads the code itself: every file is scored for AI generation and the lines that drove the score are highlighted.

Check an assignment free Why editors are different

Works on any upload. No browser extension or student install needed.

src/LinkedQueue.java Example report

1public class LinkedQueue<T> {

2 private Node<T> head, tail;

3 private int size;

5 /** Adds an element to the back of the queue in O(1). */

6 public void enqueue(T value) {

7 Node<T> node = new Node<>(value);

8 if (tail == null) { head = tail = node; }

9 else { tail.next = node; tail = node; }

10 size++;

11 }

12 // Tab to accept: public T dequeue() { ...

Lines 5 to 11 read as AI-written 89%

Likely model family

GPT

64%

Claude

25%

Gemini

90%

of AI-written code caught in held-out testing (model v3e)

1.3%

of human-written code falsely flagged in the same test

model families the report can name, including GPT, Claude and Gemini

65+

programming languages scored

Why in-editor AI is harder

The code never leaves the editor, so look at the code.

Copilot and Cursor suggest whole functions as the student types, and a Tab key accepts them. Browser monitoring and chat logs see nothing. What remains is the code, and code written by the large language models behind these tools has patterns a trained detector can read.

Trained on the models behind them

The detector was trained on output from Claude Opus 5.5, GPT-5.6 Sol and GPT-5.5 Codex, model families that coding assistants run on.

Lines, not just a file score

The windows of code that read as generated are highlighted, so you can see which function the assistant probably wrote.

A model family estimate

The report estimates which family produced the code. It names the model family, not the product a student used.

What you get on every file

See where the assistant took over.

Students mix their own code with accepted suggestions. The report shows the parts that read as generated, next to the parts that do not.

Highlighted AI regions

Each file is read in overlapping windows, and the ones that score high are marked on the code itself.

Function-level view
Score per window and per file
Evidence you can show

Model family

A tentative estimate across families such as GPT, Claude, Gemini, DeepSeek, Llama, Mistral and Qwen.

Shown with its confidence
Family, not product
Useful context, never a verdict

Window writing record

Have students work in the Window editor to see how much of each file was typed by hand.

Web and peer checks

The same check compares every file with classmates, GitHub and the web.

Policy-friendly

If your course allows assistants for some tasks, the report shows where they were used rather than a pass or fail.

How to check an assignment

Four steps, no student install.

Collect the code

Download submissions from your LMS or have students work in Window.

Upload once

Drop the folder in. AI, web and peer checks run together.

Open the flagged files

Highlighted regions show which parts read as generated.

Decide with context

Compare against your course policy on AI assistants and ask the student.

Honest limits

Long suggestions are readable. Tiny ones are not.

A one-line autocompletion looks like code anyone would write, and no detector should claim otherwise. The signal is strongest where assistants do the most work: whole functions, classes and algorithms accepted in one go.

Whole function accepted at onceStrong signal

Class or module generatedStrong signal

Single-line completionWeak signal

Student edits the suggestion heavilyMixed

The report scores windows of code, so a file that is mostly the student's own work with one generated function shows exactly that.

Questions

Copilot and Cursor detection, answered.

Can you detect code written by GitHub Copilot?

Yes, as AI-generated code. Copilot runs on large language models, and the detector is trained on output from the model families these assistants use. Each file gets a probability and the regions that read as generated are highlighted.

Can you tell whether it was Copilot, Cursor or ChatGPT?

No tool can reliably tell the product apart, because they run on the same underlying models. The report estimates the model family, such as GPT or Claude, and shows its confidence.

What about short autocompletions?

A single accepted line looks like ordinary code and is not a meaningful signal. Detection is strongest on longer stretches such as whole functions or classes accepted in one go.

How accurate is it?

In held-out testing, model v3e caught 90% of AI-written code and falsely flagged 1.3% of human-written code. Use the result as evidence alongside your course policy and a conversation with the student.

Do students need to install anything?

No. The check runs on the code they submit. If you want a record of how the code was written, students can use the free Window editor.

See where the assistant wrote the code.

Upload your next assignment and get highlighted AI regions, a model family estimate and web and peer checks in one report.

Check an assignment free See pricing

Keep reading

More on AI-written code

Product names such as ChatGPT, Claude, GitHub Copilot, Cursor and Gemini are trademarks of their respective owners. Codequiry is an independent service and is not affiliated with, endorsed by, or sponsored by any of them.