jevgram

jevgram is a command-line tool that checks if a document is AI-generated. Similar to pangram but a lot cheaper (you'll need API key from https://console.typesafe.ai/login - free, super easy, couple of clicks)

Supported inputs

• Plain text and source-like text files, such as .txt, .md, .tex, and similar files
• PDF files
• DOCX files

Minimal example:

Go to https://console.typesafe.ai/login and create an API key. Copy the key
$ export TYPESAFE_API_KEY=apikey_...
$ pipx install jevgram
$ jevgram README.md # try jevgram on this document
File: README.md
Type: text
Characters extracted: 3044
Characters sent to JEV: 3044
Model: jev-1.13.0
Prediction: AI
Probability AI: 79.00%
Probability human: 21.00%
Confidence: 58.00%

PDF extraction uses pypdf. DOCX extraction is handled directly from the Word XML contained in the .docx file.

Detailed instructions

Install (MacOS, Linux; Windows coming up)

pipx install jevgram

Get a TypeSafe API key

jevgram needs a TypeSafe API key to call JEV.

1. Login to TypeSafe: https://console.typesafe.ai/login

   (if you don't have an account, create one, for example just clicking on "Login with Google")

2. Click API Keys on the left, then Create API key on the right

3. Copy the key and save it somewhere private.

4. Set the key in your terminal:

export TYPESAFE_API_KEY="paste-your-key-here"

To avoid setting the key every time, add the export TYPESAFE_API_KEY=... line to your shell profile, such as ~/.bashrc or ~/.zshrc.

You can also pass a key directly with --api-key, but environment variables are safer/easier.

Usage

jevgram document.txt
jevgram paper.pdf
jevgram manuscript.docx --json

By default, long files are sampled down to 32,000 characters using the beginning, middle, and end of the extracted text. To send the whole extracted document, use:

jevgram document.txt --max-chars 0

To inspect extraction without calling JEV:

jevgram document.pdf --extract-only

To inspect the request body without calling JEV:

jevgram document.txt --dump-request

Output

Human-readable output includes the predicted class, AI probability, human probability, confidence when provided by JEV, and whether the input was truncated. --json produces machine-readable output for scripts.

Cost

This app is open-source/free on my end, but it uses JEV API, which is not free. Works like this: once you create a TypeSafe account and API key, you may get a credit for $5 free usage. jevgram is intended for personal use, and you can check a lot of documents for $5 (I tried dozens so far, I have yet to spend 1 cent). Once you use up the free credit, jevgram will fail and politely tell you to add credit card number to TypeSafe (to be fair, this is untested because I have not reached that point yet)

Notes

Authorship detection is probabilistic, so if you run jevgram multiple times on the same document, you'll get different (though pretty similar) probablities