SandDune uses AI for a few precise jobs: reading your handwriting, answering your questions with citations, and generating Timeline entries and their titles. Here's exactly what that means for your data — no hand-waving.
What gets processed, when
- Transcribing a page: the photo is sent to a third-party AI provider, which returns the text. That's the only time your images are processed by AI.
- Asking (and Timeline nodes): the transcribed text of the relevant pages — never the images — is sent to an AI provider to compose the answer.
- Semantic search inside Ask: finding pages "by meaning" uses a search index built from your transcriptions, isolated per-user.
The no-training rule
Your data is never used to train AI models. We don't train models ourselves, and we use our AI providers under agreements that prohibit them from training on your data. This is the line we will not move.
Processing records (worth knowing about)
We keep records of AI runs — the text inputs, retrieved note excerpts, and outputs of transcription and Ask/Timeline runs, tied to your account — so we can measure and improve quality: transcription accuracy against your corrections, whether answers were grounded in the right pages. Two boundaries: your note images are not duplicated into these records, and the records are private to your account, never shared, never used for training, and deleted when your account is deleted. (This is disclosed in Privacy Policy §4 — we'd rather over-explain it here than have you find it in fine print.)
You stay editor-in-chief
AI output can contain errors. Everything AI writes in SandDune is reviewable and editable by you — transcriptions line by line, Timeline entries wholesale — and your corrections are what the rest of the product uses.
Why we don't name the providers here
We work with a small set of established AI providers and may adjust which ones as quality and terms evolve; the commitments above hold regardless of which is in use. What each provider category sees is itemized in Privacy Policy §5.