Scriba uses large language models (LLMs) to power case summarization, doctrine explanations, and practice question generation. This page explains which models we use, where they run, and what to trust.
Models used
- Google Gemini — routed via the Lovable AI Gateway. Used for chat, brief drafting, and study aid generation.
Grounded vs. general-knowledge answers
When the assistant answers based on a document you highlighted or pasted, it is "grounded" — the answer traces back to source text you provided. When it answers a general doctrinal question with no source pasted, it is running on general knowledge and can be wrong or outdated. Scriba surfaces a persistent amber "verify" strip in the AI drawer to make this distinction explicit.
Citation warnings
AI models routinely hallucinate case names, reporters, and pinpoint citations. Scriba scans AI output for anything that looks like a case citation (X v. Y) and renders a red "Verify these citations" chip that links straight to Google Scholar so you can check before relying on the citation.
What we send to the model
- The text of your prompt.
- The passages you explicitly attach or paste.
- Minimal context about which course / node is active (so answers stay on-topic).
We do not send other users' data, your full note archive, or any billing information to the model.
What we do not do
- We do not train third-party AI models on Your Content.
- We do not draft essays, exam answers, or graded assignments (see our Honor Code).
- We do not present AI output as legal advice.
Export an audit trail
Every brief can be exported as a Word document with an "AI Disclosure Report" that lists which passages were AI-assisted and which were student-written. Use it if a professor asks about your process.