AI Text Humanizer

A local-only tool that rewrites AI-generated text to sound naturally human-written. It uses a rule-based engine to strip telltale AI phrases and overused words, then optionally passes the result through a local LLM (via LM Studio) for a deeper rewrite. A built-in stylometric scorer estimates how "AI-like" text is before and after — no paid APIs, no cloud calls, everything runs on your machine.

Requirements

• Python 3.11+ (tested on 3.14)
• Windows 11 (commands below are PowerShell; the Python code itself is cross-platform)
• LM Studio (optional) — for the enhanced LLM rewrite pass

Setup

1. Create and activate a virtual environment

python -m venv venv
venv\Scripts\activate

2. Install dependencies

pip install -r backend\requirements.txt

3. Download NLTK data

python -m nltk.downloader punkt_tab averaged_perceptron_tagger_eng wordnet omw-1.4

(The app also auto-downloads missing NLTK data on startup, so this step is optional if you prefer to let it handle itself.)

4. Create your .env file

copy .env.example .env

5. Start the server

uvicorn backend.main:app --reload --port 8000

6. Open in your browser

Go to http://localhost:8000 — you should see the app with a status badge in the header showing whether LM Studio is connected.

Setting Up LM Studio (optional)

LM Studio adds an LLM-powered rewrite pass that produces more natural output than rules alone. The app works fully without it — you'll just see "rules-only mode" in the status badge.

1. Install LM Studio from lmstudio.ai
2. Open the Developer tab (the <> icon in the left sidebar)
3. Search for and download a Qwen3 8B instruct model — look for a GGUF file with Q4_K_M quantization (good quality, fits in 8 GB VRAM)
4. Load the model in the Developer tab
5. Click Start Server (default port: 1234)
6. The app's status badge should turn green and show the model name

The app auto-detects LM Studio at http://localhost:1234/v1. If you change the port, update LM_STUDIO_BASE_URL in your .env file.

How It Works

Rule-based engine (always runs):

• Strips ~70 known AI-telltale phrases ("it is important to note that", "in today's fast-paced world", "let's delve into", etc.)
• Swaps ~60 overused AI words for plainer alternatives (crucial → key, leverage → use, tapestry → mix, etc.)
• Varies sentence structure when it detects uniform rhythm
• Injects contractions at heavy intensity (do not → don't, it is → it's)

Local LLM pass (optional, requires LM Studio):

• Feeds the rule-based output through a local LLM with intensity-aware instructions
• Light: minimal edits. Medium: moderate rewording. Heavy: free rewrite.

AI-likelihood scorer (runs before and after):

• Computes type-token ratio, sentence burstiness, AI phrase density, punctuation entropy, and readability indices
• Optionally uses real per-token perplexity from LM Studio when available
• Produces a 0–100 score: <30 = "likely human", 30–59 = "mixed", 60+ = "likely AI"

Troubleshooting

LM Studio badge shows "not detected" Make sure you've loaded a model and clicked Start Server in LM Studio's Developer tab. Check that the port in LM Studio matches LM_STUDIO_BASE_URL in your .env (default: 1234).

First request is slow Normal — the model is loading into VRAM. Subsequent requests are fast.

Upload rejected with "unsupported file type" Only .txt, .docx, and .pdf files are accepted. Check the file extension.

Upload rejected with "file too large" Default limit is 10 MB. Change MAX_UPLOAD_SIZE_MB in .env to increase it.

NLTK errors on startup Run python -m nltk.downloader punkt_tab averaged_perceptron_tagger_eng wordnet omw-1.4 manually, or delete the nltk_data folder and let the app re-download on next start.

Project Structure

ai-text-humanizer/
├── backend/
│   ├── main.py              # FastAPI app, startup logic, route registration
│   ├── config.py            # pydantic-settings config from .env
│   ├── requirements.txt     # Python dependencies
│   ├── engine/
│   │   ├── rules.py         # Rule-based humanization pipeline
│   │   ├── llm_client.py    # LM Studio client (graceful fallback)
│   │   ├── detector.py      # AI-likelihood scorer
│   │   ├── file_parser.py   # .txt/.docx/.pdf extraction + export
│   │   └── data/
│   │       ├── ai_phrases.json   # Blocklist of AI-telltale phrases
│   │       └── synonyms.json     # Overused-word → natural-alternative map
│   ├── routes/
│   │   ├── humanize.py      # POST /api/humanize, GET /api/health-llm
│   │   ├── detect.py        # POST /api/detect
│   │   └── files.py         # POST /api/upload, /api/batch-humanize, /api/export
│   ├── models/
│   │   └── schemas.py       # Pydantic request/response models
│   └── tests/
│       ├── test_rules.py    # Rule engine tests
│       └── test_detector.py # Scorer tests
├── frontend/
│   ├── index.html           # Single-page UI
│   ├── style.css            # Dark neumorphism stylesheet
│   └── app.js               # Frontend logic (plain fetch, no framework)
├── .env.example             # Environment variable template
└── README.md

API Endpoints

Method	Path	Description
GET	/api/health	App health check
GET	/api/health-llm	LM Studio status + model name
POST	/api/humanize	Humanize pasted text (rules + optional LLM)
POST	/api/detect	Score text for AI-likelihood
POST	/api/upload	Upload and extract text from a file
POST	/api/batch-humanize	Upload multiple files, get a zip of humanized results
POST	/api/export	Export text as .txt or .docx download

License

MIT