How we use AI

Radabase uses AI in a few narrow places, always as a drafting assistant for a contributor — never as an author, a diagnostician or a reviewer. This page says exactly where, so nothing about it is a surprise.

Where AI is used

Importing a report. If you drop your own report into the submission form, a language model reads the de-identified text and proposes values for the form fields — signalment, findings, diagnosis, history, outcome, teaching points, quiz differentials and a suggested difficulty level. Fields you have already filled in are kept.
Dictation. If you speak a case, a speech-to-text model turns the audio into text, and a language model proposes the same form fields from it.
Quiz alternatives. When a contributor asks for them, a model suggests three plausible wrong answers for a case's multiple-choice question. The correct answer is always the contributor's own diagnosis.
Grading typed quiz answers. If you type your own diagnosis in Quiz Mode, a language model compares it with the case's reviewer-approved diagnosis and marks it correct, partly correct or incorrect, with a one-line reason. It sees only your typed answer and that diagnosis — no images, no account details. It never decides the diagnosis, and you can always override its grade. Your answer is not stored.
Search keyword suggestions. A reviewer can ask a language model to suggest other words people might search for a case — synonyms, everyday terms and abbreviations (for example "bladder stone" for cystolith). The model sees only the diagnosis, species, body regions, modalities and imaging findings — never patient, owner or account details. It does not diagnose. Reviewers edit or remove suggestions before saving them — on the review page or while uploading a case; reviewers may also fill keywords for existing cases in bulk and tidy them afterwards. Keywords are used for search and are listed at the bottom of each case page.
Image and loop screening. Uploaded images are scanned automatically for burned-in text such as names, dates and ID numbers. For a cine loop, a handful of frames are sampled and scanned the same way — the clip itself is not sent anywhere. These flags are advisory only.

Where AI is not used

No case is written, diagnosed, approved, rejected or published by AI. Every diagnosis comes from the contributor. Every published case is checked by a human reviewer — a board-certified radiologist, or at times another relevant specialist. AI never sees the review decision and never makes one.

A human always confirms

Anything a model suggests is shown to the contributor first, field by field, and nothing is saved until they press apply. Every field stays fully editable afterwards, and a reviewer reads it again before the case goes public.

What happens to what you send

Reports are read entirely in your browser. The file is never uploaded and its raw text is never stored by us.
A cine loop is re-recorded in your browser before it is uploaded. The original file — with its embedded camera and patient metadata — stays on your machine. Only the new recording, which carries none of that metadata, is sent.
Dictation audio is sent for transcription only. It is never stored, and the transcript is discarded once the fields have been extracted.
Text is scrubbed of identifiers before any model call, and anything that still looks identifying afterwards is dropped rather than saved.
Only the structured fields you review and approve are stored with the case. Nothing you send is used to train any model.

Limits and mistakes

AI can mis-hear a word, over-summarise, or state something that was never in your report. That is precisely why nothing it produces is saved automatically. Read every suggested field before you accept it. If you spot something wrong on a published case, report the case and we will act on it.

Which models

Language and transcription models are accessed through Lovable AI, which routes requests to OpenAI and Google models on our behalf as a processor for Radabase. We do not send them images, contributor emails or account details — only the de-identified case text or dictation audio described above.

Questions about any of this? See our Privacy Policy and Terms of Use. Radabase is an educational reference — nothing here, AI-assisted or not, is clinical advice for an individual animal.