DocForensics

Document Authenticity Detector

Repository

DocForensics is a browser-only web app that analyzes PDF, DOCX, and ODT documents to detect signs of AI-generated content or copy-paste. Files never leave the device.

Features

• Multilingual UI: English, French, German, Spanish, Galician, Catalan, and Basque
• Three analysis layers:
  • Metadata
  • Linguistic signals
  • Software fingerprints
• Segment-level style change detection for pasted or mixed-authorship blocks
• PDF provenance and structure checks, including C2PA manifest detection, embedded files, and suspicious tiny-text layers
• Deeper DOCX OOXML checks, including RSID coverage, RSID diversity, and paragraph style distribution
• Risk score from 0 to 100 with four levels: Low, Moderate, High, Critical
• Rich PDF export of the full session report with color-coded summaries, score breakdowns, and per-document detail
• Persistent local history of analyses restored after a page reload
• Multi-file upload, including ZIP archives expanded locally in the browser
• No backend

Supported formats

• PDF
• DOCX
• ODT
• ZIP archives containing PDF, DOCX, or ODT files

Legacy binary .doc files are not supported.

Analysis methodology

Metadata signals

Signal	Suspicion
Editing time under 1 minute for more than 100 words	Critical
Editing time under 3 minutes for more than 200 words	High
0-1 revisions	High
Creation and modification date almost identical	High
Empty or generic author field	Medium

Linguistic signals

Signal	Basis
Low burstiness	AI often generates more uniform sentence lengths
Low type-token ratio	AI tends to reuse vocabulary
AI transition phrases	LLMs overuse stock transitions
Slop patterns	Repetitive AI-heavy expressions
Low Shannon entropy	AI text is often more predictable
Low first-person pronoun ratio	AI tends to avoid personal language
No interrogative sentences	AI rarely uses rhetorical questions
Zero-width characters	Possible watermarking or manipulation
Local style shifts between segments	Possible pasted blocks or mixed authorship

Software fingerprint signals

Signal	Suspicion
AI tool in creator/producer/generator metadata	Critical
Programming library signature	High
Missing Word revision IDs in DOCX	Medium
Low RSID coverage or diversity in DOCX	Medium
Embedded files or suspicious tiny-text layers in PDF	Medium
Online converter signature	Medium
Many font families in DOCX	Medium

Tech stack

• React 19
• Vite 8
• TypeScript 5.9
• Tailwind CSS 4
• pdfjs-dist
• jszip
• fast-xml-parser
• mammoth

Setup

git clone https://github.com/gafapa/autenticador.git
cd autenticador
npm install
npm run dev

Open http://localhost:5173.

Build

npm run build

The production bundle is generated in dist/.

Disclaimer

This tool provides indicative analysis only and does not constitute definitive proof of AI authorship or plagiarism. No single signal is conclusive; multiple convergent indicators are required.

License

MIT