AI usage policy

Training

AI training is part of the agreement every contributor accepts before recording. Arabic and English production models on api.mathla.org are trained from separate pools — never mixed.

What AI does on the platform

Quality checks, practice feedback, and search assistance may use models. These features support learning — they do not replace human reviewers for dictionary approval.

No false claims

We do not claim perfect recognition or translation. “Talk to me” and related products disclose limitations in-product. Always verify critical information with qualified humans.

Research labs

Research GPU environments are not connected to live contributor data without an approved export list. Sandbox experiments never train production brains directly.

Derived data

Recordings may be converted to pose landmarks, motion features, and metadata for model training and testing. Derived data is Mathla property and is never published without approval.