Expert insights on AI code detection and academic integrity
As AI coding assistants become ubiquitous, learn how institutions are adapting to detect AI-generated code and maintain educational standards.
Codequiry Editorial Team
·
Jan 5, 2026
Stay ahead with expert analysis and practical guides
General
12 min
Dr. Sarah Chen
·
3 hours ago
A CS2 course with 412 students and eleven TAs ran peer similarity, web-source matching, and AI-generation scoring in one pass per assignment instead of three separate workflows. The interesting result wasn't the flag count. It was that the review meetings stopped arguing about whether a submission looked suspicious and started arguing about what the department's own AI policy actually permitted.
General
12 min
Rachel Foster
·
1 day ago
A department ran a web-source scan on 412 intro Python submissions and got 61 hits, 23 of which were noise. This is the eight-step workflow they built afterward, from normalizing LMS exports and budgeting GitHub's code search rate limit to ranking matches by token count instead of percentage and assembling an evidence packet that survives an appeal.
General
13 min
Rachel Foster
·
2 days ago
License detection spent twenty years matching strings against license texts and missing the code underneath them. The algorithms that fixed it came out of plagiarism research, not legal tech. A look at what changed after 2003, and what the AI boom reopened.
General
15 min
Dr. Sarah Chen
·
2 days ago
Running one scan and hoping it covers everything is how courses end up with both missed cases and false accusations. This walkthrough separates peer similarity, web provenance, and machine generation into three pipelines, then shows how to merge them into a ranked queue a TA can work through in an afternoon.
General
12 min
Marcus Rodriguez
·
3 days ago
I audited 1,847 programming submissions across three institutions over two terms. 3.3% were translations from another language, and automated similarity scoring caught exactly zero of them. Here is the method, the numbers, the tool comparison, and what actually surfaced the cases.
General
10 min
Priya Sharma
·
5 days ago
A lead TA at a mid-sized CS department gets 412 similarity pairs back from a single algorithms assignment and has six days to decide which ones matter. This is the triage method she uses: score stacking, cohort-normalized thresholds, and a written rubric that has survived two grade appeals.
General
13 min
Alex Petrov
·
5 days ago
We ran 412 bootcamp take-home submissions through four independent detection signals and followed 63 flagged candidates to a live code interview. This is what we found about which signals actually predict AI-generated code, which ones just predict "this person read a tutorial," and how to build an audit you can defend when a candidate pushes back.
General
10 min
Marcus Rodriguez
·
6 days ago
Meridian, a 40-person benefits SaaS, had 340,000 lines of contracted code and no way to answer question 34 on a diligence questionnaire. Here's the five-gate provenance audit that followed, what the tools actually found, and the clause in the MSA that would have made all of it cheap.
General
10 min
Alex Petrov
·
1 week ago
I ran 4,000 bootcamp submissions through web-source matching and 22% matched something public. Most of those matches were harmless. Here's what a real Stack Overflow diff looks like, why GitHub matches almost always come with a receipt, and how I stopped flagging half my class with false positives.
General
12 min
Alex Petrov
·
1 week ago
I seeded 48 known-copy pairs into 1,247 real bootcamp submissions and ran them through MOSS, JPlag, Dolos, and Codequiry. Token-based engines held up against renaming and reformatting, then collapsed once control flow changed. Here's the data, the methodology, and where every engine breaks.
General
17 min
Rachel Foster
·
1 week ago
Most CS departments teach students to cite what they copy but never to read the license attached to it. This reported guide walks through the licenses that actually show up in student repositories, the four failure modes that recur in grading, and a provenance rubric that separates academic honesty from legal obligation.
General
10 min
James Okafor
·
1 week ago
Take 30 hand-written Java solutions, push them through five levels of mechanical refactoring, and you get a controlled ladder for asking a simple question: how much does each similarity engine actually see? The answers differ by nearly a factor of four, and the differences map cleanly onto how we should be designing assignments.