Was this code written by AI?
Paste code, upload a file or a zip, or scan a GitHub repo or pull request. Kerata shows how likely it is that ChatGPT, Claude or another AI model wrote it, highlights the lines that look AI-written, and tells you how far to trust the answer.
/tree/branch/folder to scan part of a repo. Pull requests and commits: only the lines they add are scored.Accuracy, measured honestly
The cutoff for “likely AI” is set to protect people from false accusations. These numbers come from held-out data the model never saw, read live from the model's evaluation files.
Everything you need to make a fair call
Built for the moment someone asks “did you write this?”, when a bare percentage isn't good enough.
A calibrated score
The percentage is a calibrated probability, with a confidence level and a plain-language verdict. Short files get “too short to judge” instead of a guess.
The lines, not just a number
A dedicated line model highlights which lines read as machine-written and calls out AI-like sections pasted into human code.
Repos and pull requests
Scan a public repo, a folder of it, or just the lines a PR or commit adds. Every file gets its own score and highlights.
Honest about its limits
Every result explains what the score means and what it can't tell you. “Most similar to” vendor hints are labelled as weak leads.
Plagiarism and class checks
For instructors: rename-resistant peer similarity, past-term matches, edit history and AI signals in one ranked review list.
Private by default
Anonymous checks are never stored. Signed-in history is yours alone, protected by row-level security, and deletable anytime.
Who it's for
Instructors & TAs
Triage a stack of submissions: see which ones deserve a closer read, with the evidence to discuss with the student.
Hiring teams
Check take-home assignments and code samples before the interview, and ask better follow-up questions.
Maintainers & reviewers
Score what a pull request adds and flag sections worth a careful review.
Anyone curious
Got code from a contractor, a classmate or the internet? Paste it and see.
How it works
Give it code
Paste a file, drop a zip, or paste a GitHub link. Pull requests and commits are scored on the lines they add.
Two models read it
A fine-tuned code encoder blended with a stylometric model scores each file; a token-level tagger marks the lines that look machine-written.
You get evidence
A calibrated score, a confidence level, highlighted lines and sections, a per-file breakdown and a report you can print or share.
Questions
Is a high score proof that someone used AI?
No. Kerata measures how much code resembles code written by AI models, compared with human code. People follow tutorials, use autocomplete, write in a textbook style or use AI tools they were allowed to use. Treat a score as a reason to look closer and to talk to the author, never as a verdict on its own.
How accurate is it?
The numbers on this page are read straight from the evaluation files of the model this server runs. The cutoff for “likely AI” is deliberately strict to protect people from false accusations, so a fair share of AI-written files land in “mixed” instead. Short files get “too short to judge” rather than a guess. The methodology page has every number, per test set, and the known weak spots.
Which languages and AI models does it cover?
It is trained and evaluated on Python, JavaScript, TypeScript, Java, C, C++, Go, Rust, Kotlin, Swift, Ruby, PHP and Haskell. Other languages (C#, Dart) are accepted, but they haven't been evaluated, so treat their scores with extra caution. The training data covers 20+ OpenAI models, Claude, Gemini and open-weights models, including AI edits of human code and code written by AI coding agents in real repositories.
What does “most similar to” mean?
Only for code whose verdict is “likely AI”, a separate model estimates which vendor's style it resembles most (OpenAI, Anthropic or open-weights models). It is a lead with a strength label, not an identification.
Do you store or train on my code?
Code you check without an account is analyzed in memory and not stored. Signed-in scans are saved to your private history (you can rename, export or delete them anytime). We never train on your code.
What does it cost?
Kerata is free during the public beta. Without an account: 50 checks a day per network (pasted code, single files, public repos and pull requests; up to 20 of them repos or pull requests). Nothing is stored. Free account: 2,000 checks a month, saved history, zip uploads, class checks (20 a month, up to 400 students each) and an API key (2,000 calls a month).
Does it catch plagiarism too?
Yes. Signed-in instructors can run class checks: MOSS-style peer similarity that survives renaming, matches against past terms, edit-history signals and AI detection, merged into one ranked review list.
Is there an API?
Yes. Sign in, open the account menu → API keys, create a key and call the REST API from scripts, CI or your LMS. A GitHub Action scores the lines each pull request adds.
Free during the public beta
Free account: 2,000 checks a month, saved history, zip uploads, class checks (20 a month, up to 400 students each) and an API key (2,000 calls a month).
Without an account: 50 checks a day per network (pasted code, single files, public repos and pull requests; up to 20 of them repos or pull requests). Nothing is stored.