AI Content Detector

Part of the worlds-biggest-software-project initiative.

An open-source, AI-native detector for AI-generated text — built for academic integrity, editorial review, and compliance workflows where false positives carry real consequences.

The AI Content Detector identifies text generated by major LLMs (GPT-4/5, Claude, Gemini, Llama, Mistral) and reports calibrated probabilities at the sentence and document level. It is designed for educators, editors, and compliance teams who need defensible, audit-ready detection results rather than opaque marketing-grade scores.

Why AI Content Detector?

Independent testing of 30+ commercial detectors found only five scoring above 70% accuracy, with many misclassifying formal or neutral human writing as AI-generated.
Leading commercial tools have well-documented false-positive bias against non-native English writers (Stanford HAI study cited 61%+ misclassification of ESL student work by Turnitin's indicator).
Most incumbents are closed-weights proprietary SaaS; only Pangram Labs has released open weights, and those are CC BY-NC-SA 4.0 (non-commercial only).
No mainstream tool currently captures writing-process provenance, supports structured AI-use disclosure, or reads LLM provider watermarks (SynthID, OpenAI watermarking) as additional signals.
Pricing in the category is opaque and credit-metered; institutional buyers face procurement friction and per-scan costs that scale poorly across large submission volumes.

Key Features

Detection Core

AI-generated text detection across GPT-4/5, Claude, Gemini, Llama, Mistral, and other major LLMs
Sentence- and paragraph-level probability highlighting with colour-coded overlay
Calibrated confidence reporting using probability ranges rather than binary verdicts
Tiered classification labels (e.g. fully human / lightly assisted / moderately assisted / fully AI) for easier interpretation
Paraphrase and obfuscation detection to identify lightly reworded AI content

Integration & API

REST API with OpenAPI 3.x specification and API key authentication
Batch submission for high-volume programmatic processing
LTI 1.3 integration for Canvas, Moodle, Brightspace, and Google Classroom
Webhook support for automated submission and grading workflows

Institutional & Audit Tooling

Audit-ready PDF report export with methodology notes and uncertainty disclosure
Institutional admin dashboard with per-student and per-assignment summaries
Combined plagiarism + AI detection in a single submission workflow
Domain hints (academic, journalistic, technical) to mitigate false positives on formal prose

Fairness & Provenance

Non-native English bias detection and flagging to prevent unfair accusations
Multilingual detection targeting 10+ languages at launch
Writing-process provenance tracking via edit history and keystroke-dynamics integration (backlog)
Structured AI-use disclosure templates so authors can document legitimate AI assistance
Watermark-aware detection integrating with C2PA Content Credentials and SynthID where available (backlog)

AI-Native Advantage

The detector applies learned contextual models on top of classical perplexity and burstiness signals, with adversarial training against paraphrasing and "humanising" attacks that defeat current commercial tools. Domain-adaptive calibration provides separate thresholds for scientific writing, legal documents, journalism, and student essays to reduce false positives where they matter most. Bayesian or ensemble methods produce calibrated probability intervals rather than point estimates, and LLM-powered rationale generation explains exactly why a passage was flagged — increasing educator trust and transparency.

Tech Stack & Deployment

The project targets a REST API (OpenAPI 3.x, API key auth) with JSON responses, batch endpoints, and webhook callbacks, plus LTI 1.3 connectors for the major LMS platforms. Lightweight model variants are intended to run locally for privacy-preserving institutional deployments, alongside a hosted SaaS option. Open standards in scope include LTI 1.3, OpenAPI 3.x, and C2PA Content Credentials for watermark interoperability.

Market Context

The AI content detection space matured rapidly through 2024–2025 as ChatGPT, Claude, and Gemini reached mainstream adoption, producing a crowded field of commercial tools (Originality.ai, GPTZero, Winston AI, Turnitin, Pangram Labs, Copyleaks, Grammarly, Sapling, Scribbr, ZeroGPT). Pricing is dominated by credit-metered SaaS (Originality.ai at 1 credit per 100 words; Winston AI at 500,000 credits for $35; GPTZero API from $45/month for 300,000 words) and opaque institutional licensing (Turnitin). Primary buyers are universities and academic integrity offices, publishers and journal editors, content marketing teams, and enterprise compliance functions.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.