Free · Private · Zero-shot
The highest-scoring zero-shot AI detector on RAID, the largest public benchmark for AI-text detection. It runs on your own computer: no account, no word limit, and your text never leaves your device.
Accuracy
RAID tests detectors on 672,000 texts from 11 AI models, 8 kinds of writing and 11 attack types. Detect Law was scored by RAID's official evaluation.
Official results from RAID's evaluation service, 8 October 2026 (submission); listed on the RAID leaderboard once merged. Trained commercial detectors score higher on RAID. Detect Law is the top detector that needs no training on AI text.
Use cases
Screen essays and assignments on your own laptop. Student work never leaves your computer.
Check submissions before you publish. Drop in a .docx or .pdf and get an answer in seconds.
Audit content at any volume for free. One GPU checks about 7 thousand-word documents per second.
An HTTP API and an MCP server come built in. Add AI detection to your app or your AI assistant.
How to use it
One command on a Mac, Windows or Linux computer. No account, no credit card.
Paste any text, or drop a .txt, .md, .docx or .pdf file into the app.
A plain answer in seconds, not a made-up percentage.
The text shows statistical evidence of AI writing. Send it to a person for review.
The text was checked and an AI-sized effect was ruled out. Not proof of human authorship.
No significant AI evidence, but not enough to rule it out. Most human writing ends here.
Features
Runs entirely on your computer. Nothing is uploaded, ever.
No account, no word limit, no daily cap, no trial that runs out.
About 2 seconds for a 1,000-word document on a laptop. It stops as soon as it's sure.
Wrongly flagged 0.22% of human texts in our tests, about 1 in 450.
Hidden characters, look-alike letters and spacing tricks are cleaned before checking.
Web app, command line, HTTP API, and an MCP server for Claude and other assistants.
Plain text, Markdown, Word and PDF, with reference lists skipped automatically.
After a one-time download, no internet connection is needed.
Pricing
Detect Law runs on your own hardware, so there is nothing to meter and nothing to pay.
| Detector | per 1,000 words |
|---|---|
| Pangram | $0.50 |
| GPTZero | $0.09–0.15 |
| Originality.ai | $0.09–0.12 |
| Winston AI | $0.05–0.10 |
| Sapling | $0.015–0.03 |
| Detect Law | $0 |
The film
Get started
You get the app, a command line tool, an HTTP API and an MCP server. The first run downloads about 5 GB once.
pip install "git+https://github.com/devkancheti4-design/detect-law"
detect-law-server
Then open 127.0.0.1:8765 in your browser. Paste text or drop a file.
detect-law essay.docx report.pdf
curl -s 127.0.0.1:8765/v1/detect -H 'content-type: application/json' -d '{"text": "Paste the essay here"}'claude mcp add detect-law -- uvx --from git+https://github.com/devkancheti4-design/detect-law detect-law-mcp
check_textcheck_fileabout_detect_lawThe check still runs on your computer.
FAQ
Yes. It runs on your own computer, so there is no server cost to pass on: no account, no credit card, no word limit and no trial period.
Yes. Every check runs locally, and nothing you check is uploaded or stored anywhere else. After the one-time download it works offline.
On RAID's official test of 672,000 texts it caught 92.33% of AI-written text while flagging 5% of human text, 88.99% including adversarial attacks, and 86.54% at a strict 1% false-positive rate. That is the highest score of any zero-shot detector on RAID. Trained commercial detectors score higher on RAID; Detect Law needs no training on AI text and costs nothing.
On RAID it was tested on text from ChatGPT, GPT-4, GPT-3, GPT-2, Llama chat, Mistral, MPT and Cohere models, catching 82% to 100% depending on the model. Text from chat models (Llama 2 chat, ChatGPT, Mistral chat) was caught 99.6–100% of the time. Newer frontier models are harder to detect in short texts; longer documents give much stronger evidence.
AI evidence: the text shows statistical evidence of AI writing. No evidence: an AI-sized effect was ruled out. Undecided: no significant evidence either way, which is where most human writing ends up.
No. A verdict is statistical evidence, not proof, and a small share of human texts are flagged (0.22% in our tests, more for poetry). Use it to decide what a person should read closely, never to penalise anyone automatically. Human writing heavily polished by an AI tool can also be flagged.
Partly. Paraphrasing is the hardest attack: Detect Law caught 76.6% of paraphrased AI text on RAID, and paraphrased text is often marked Undecided.
English.
200 words or more works best. Very short texts are often Undecided, and longer documents give stronger evidence.
An Apple silicon Mac, or a Windows or Linux computer with Python 3.10+ and a C compiler. An NVIDIA GPU makes it faster but is optional. The first run downloads about 5 GB.