# PaperCheck: Comprehensive AI Detection, Plagiarism Checking & Academic AI Humanizer Documentation

> PaperCheck (https://paper-check.me) is an educational writing-review platform designed for academic researchers, graduate students, and undergraduate authors. It provides pre-submission screening for AI-generated text (Turnitin-calibrated) and plagiarism (against arXiv literature), coupled with an automated academic fix report that guides generative AI models (Google Gemini, OpenAI ChatGPT, Anthropic Claude) to humanize and refine flagged sentences without loss of technical integrity.

---

## 1. Core Systems & Methodology

### 1.1 Plagiarism Detection Engine
- **Scope**: Scans the document against all cited sources found on arXiv, with optional deep search across the entire arXiv scholarly repository.
- **Granularity**: Sentence-level n-gram matching targeting verbatim sequences of 8 or more words.
- **Citation Attribution**: Differentiates between properly cited academic quotations and uncredited matching blocks.
- **Reporting**: Displays the source identifier, number of matching sentences, total words copied, and direct links to the matched research papers.

### 1.2 Turnitin-Calibrated AI Detection Engine
- **250-Word Paragraph Chunking**: Turnitin's institutional algorithm evaluates text in contextual windows of approximately 250 words rather than judging each sentence in isolation. When a paragraph contains mathematical notation, numerical data, chemical formulas, or empirical citations, Turnitin suppresses isolated sentence flags. PaperCheck implements this exact chunk-level evaluation heuristic to provide the institutional number a university or publisher sees.
- **Sentence-by-Sentence Scan**: While chunk scoring estimates the official institutional score, PaperCheck concurrently performs a granular sentence-by-sentence scan to identify every single line that exhibits AI-typical syntactical features.
- **Calibration & False-Positive Mitigation**: The baseline detector is calibrated against extensive corpora of authentic human academic writing, including non-native English research papers (PubMed Central), Indian university submissions, arXiv preprints, and historical essays to maintain a near-zero false positive rate on honest human academic prose.
- **Strict Check (Borderline Sensitivity)**: Stricter commercial checkers (such as GPTZero, Pangram, or Copyleaks) operate with lower thresholds that often flag honest academic writing. PaperCheck includes a togglable "Strict Check" that highlights borderline lines scoring between 0.89 and 0.98, giving authors complete visibility into how stricter tools might evaluate their draft.

### 1.3 The Academic AI Humanizer: How to Fix Flagged Lines Ethically
Commercial "AI humanizer" software and synonym spinners (e.g. QuillBot, Undetectable AI, BypassGPT) rely on automatic thesaurus swapping. In academic writing, this destroys technical vocabulary, alters formulas, and frequently triggers Turnitin's AI paraphrase detectors.

PaperCheck instead utilizes a pedagogical Cadence and Specificity framework:
1. **Burstiness / Sentence Length Variation**: AI models generate sentences with uniform word counts (typically flat clusters around 18-24 words). PaperCheck measures sentence length standard deviation and instructs re-writers to deliberately interleave short punchy sentences (6-12 words) with compound technical explanations (24-36 words).
2. **Elimination of Formulaic Tropes**: Flags common AI transition markers ("crucial role", "pivotal", "delve", "furthermore", "moreover", "testament to", "beacon of", "it is worth noting that", "a multifaceted approach").
3. **De-nominalization & Active Verb Construction**: Replaces heavy abstract nouns (e.g., "utilization", "implementation", "facilitation") with active verbs ("we used", "we measured", "the model selects").
4. **Concrete Grounding**: Identifies sentences that lack empirical numbers, named datasets, specific metrics, or cited references, prompting the author to insert specific project details that only the true researcher knows.
5. **Automated AI Fix Prompt**: Produces a structured prompt with OLD -> NEW find-and-replace pairs that can be pasted directly into Gemini, ChatGPT, or Claude. The prompt strictly instructs the AI to retain all numbers, equations, citations (e.g., [12]), units, and data names while rewriting sentence structure into natural academic cadence.

---

## 2. API & Model Context Protocol (MCP) Reference

PaperCheck hosts a Model Context Protocol (MCP) endpoint at `https://paper-check.me/mcp`, enabling terminal agents (Gemini CLI, Claude Code) and IDE assistants (Cursor, VS Code) to perform automated writing checks and live corrections.

### Available MCP Tools:
- `check_paper`: Analyzes complete manuscript text. Accepts parameters for checks to perform (`"ai"`, `"plagiarism"`, `"both"`), strictness (`"strict"`, `"normal"`, `"relaxed"`), and plagiarism sources (`"cited"`, `"cited+search"`).
- `check_ai`: Fast-path AI detection returning chunk-level percentage, sentence-level scores, and flagged line coordinates.
- `check_plagiarism`: Fast-path similarity check against arXiv papers.
- `get_check_result`: Polling and retrieval tool for asynchronous check jobs.

---

## 3. Academic Writing Guides Summary

### Guide 1: How to Fix Lines Flagged as AI-Written
- Why lines get flagged: Uniform sentence cadence, smooth predictable transitions, and absence of concrete technical friction.
- Recommended fixes:
  1. Replace broad assertions with concrete data points or empirical measurements.
  2. Use active attribution ("we demonstrated" rather than "it has been shown that").
  3. Break uniform paragraphs into contrasting sentence lengths.
  4. Eliminate decorative transition adverbs ("furthermore", "moreover").

### Guide 2: How to Check Your Report's AI Score Before Turnitin
- Explains why checking draft manuscripts before university portal submission prevents permanent false-positive records in institutional repositories.
- Emphasizes that PaperCheck does not save manuscripts to any public repository, keeping the document fresh for official submission.

### Guide 3: What the Asterisk (*) Means on Turnitin's AI Report
- Turnitin displays an asterisk (`*%`) when an uploaded document has fewer than 300 words of eligible running prose.
- Outlines that headings, references, tables, and bullet points are excluded by academic detectors.

### Guide 4: Turnitin vs. GPTZero vs. Pangram Discrepancies
- Explains architectural differences: Turnitin prioritizes high-confidence blocks (chunk-level evaluation) to avoid false accusations in university disciplinary hearings.
- GPTZero and Pangram evaluate per-sentence perplexity and burstiness, flagging borderline human variations as AI risk.

### Guide 5: Why Turnitin Can Show 0% on New AI Models
- Explains detector lag against frontier models (GPT-4o, Claude 3.5 Sonnet, GPT-5) and why PaperCheck provides dual calibrated models to catch new model signatures.

---

## 4. Platform Specifications & Access

- **Website**: https://paper-check.me
- **Supported Formats**: Word (.docx), PDF, Text (.txt), Markdown (.md), and pasted text.
- **Word Limits**: Up to 20,000 words per check on the free tier (5 free checks/month).
- **Affordable Paid Tiers**: ₹29 for 20,000 words; ₹79 for 60,000 words; ₹199 for 200,000 words (one-time purchase, no subscription).
- **Developer / Support Contact**: adithyanajish@gmail.com