A local-first AI text detector built in Swift and MLX.
This is an experimental machine-learning project for identifying AI-like patterns in written text. The goal is to build and train our own detector rather than rely on third-party detection APIs.
The project focuses on:
- Low false-positive rates
- Calibrated uncertainty
- Robust evaluation across different models and writing styles
- Local and private inference
- Reproducible training and evaluation
- Swift
- MLX Swift
- Swift Package Manager
- Apple silicon
Run make test from the project directory. This uses xcodebuild to compile
MLX's Metal shaders and run the Swift tests. The current MLX Swift dependency
does not support building those shaders with the swift test command.
AI-text detection is inherently uncertain. A detector score should be treated as a signal, not as proof of authorship.
When the evidence is weak, the system should prefer uncertainty over false confidence.
MIT