Legal AI

AI Packets for Public Defense

A short public note on attorney-supervised AI: what may be summarized, what must be reviewed, and what stays out.

2026

The core design rule

The system should behave like a review packet.

That means every output needs a source reference, an extracted fact, an attorney check, and a risk note. Fluency is not enough. In public defense settings, a beautiful paragraph with a missing page citation is worse than a rough table that lets an attorney verify every claim.

First use case

The right first use case is triage, chronology, records organization, omission detection, and mitigation support. Those tasks still need judgment, but they are easier to supervise because the system can be asked to show where each point came from.

The useful question is practical:

Can it make file review faster without hiding uncertainty?

What a public pilot can show

SourceDocument, page, span, date, and record type remain visible.

FactEach extracted point has a status: present, absent, unclear, or not applicable.

ReviewAttorney action is a first-class field, not an afterthought.

BoundaryPrivileged, sealed, client-identifying, and office-confidential material stays out.

That is enough to evaluate the review discipline while keeping sensitive material out of public view.