AI Detection Breakthrough
.png){kind=logo}
Tap the image to see it full size.
Our best detection yet for AI-written source code is now live for all users. Model v3e is trained on Claude Opus 5.5, GPT-5.6 Sol and GPT-5.5 Codex submissions. It achieves 90% accuracy with a 1.3% false flag rate on human-written code, making our AI Code Detector the most accurate in the world.
Already a customer? Open your dashboard
- AI/GPT code caught
- 90%
- False flags, human code
- 1.3%
- Trained on
- Claude Opus 5.5GPT-5.6 SolGPT-5.5 Codex
How we built it
I'm Josiah, the founder of Codequiry. I studied at UC Berkeley, and I've spent the years since building the tools teachers use to tell whether the code in front of them is really their student's.
Here is how we built this one. We started with human code we could trust: student submissions written before ChatGPT existed, when nobody had an AI to lean on. Real assignments, with the odd variable names, the leftover print statements and the half-finished comments that people actually write.
Then we did the opposite. We took the newest models, Claude Opus 5.5, GPT-5.6 Sol and GPT-5.5 Codex among them, had them write code the way students ask them to, and fed our model how they write. Our main model learned to tell the two apart.
The part that surprised us is how well it generalizes. It doesn't memorize one model's habits, so it catches code pasted straight out of ChatGPT, and it keeps catching it as new versions come out. When it flags something, it shows you the exact block, like the class in the screenshot above, not just a number for the whole file.
I'll be honest about the rest of the market. Most AI detectors I've tested hand you 99% or 1% and very little in between, and a few small edits get code past them. That isn't detection. Ours is the best we've built, and in our testing it's the best out there.
It isn't finished. Claude and ChatGPT are where most copied code and AI contract cheating come from now, so that's where we keep training, to make sure it catches their code accurately every time a new model ships.
If you teach programming, try it on your next assignment.
Questions
- How accurate is Codequiry's AI code detection?
- Model v3e catches 90% of AI/GPT-written code and falsely flags 1.3% of human-written code.
- Which AI models is it trained on?
- Claude Opus 5.5, GPT-5.6 Sol and GPT-5.5 Codex submissions. It scores any code a student hands in, but detection can be weaker against models it was not trained on.
- Is an AI score proof that a student cheated?
- No. A score is a signal, not a verdict. Open the submission, read the highlighted lines and talk to the student before acting.
- How do I try it?
- AI detection runs on a Codequiry plan. Sign up, start your plan and upload an assignment. The AI Detection tab appears next to the plagiarism results.
How detection works, file by file: the AI Detector
Try it on your next assignment
AI detection runs on a Codequiry plan. Sign up, start your plan, upload a class's code and see the flagged blocks next to the plagiarism results.
Start free trial