Judiciary — Teaching the Bench

The Project’s judiciary workstream — the PennLaw-Judiciary AI Testbed — provides federal judges and their chambers with structured AI access, training, and ongoing support, governed by a formal Memorandum of Understanding. The work spans use-case research, chambers-facing tools, and a continuous feedback loop with participating chambers.

Per the Project’s pillar architecture, this work sits in Teach — judicial AI training is teaching the bench, not partnership-mediated Building.

The Judicial AI Portal

The public face of the work is the Judicial AI Portal, a plain-language resource for federal judges and chambers staff. It explains how AI tools differ from the legal research systems judges already trust, sorts observed chambers use cases by risk (drawn from the Testbed, reported in aggregate with chambers de-identified), and sets out guardrails for verification and confidentiality. It was prepared for the AI plenary at the 2026 Judicial Conference of the Sixth Circuit and went public on August 28, 2026. Openly licensed (CC BY 4.0).

What it is

A research partnership across multiple federal courts and the Delaware Court of Chancery, organized around shared infrastructure, structured onboarding, and a regular cadence of feedback.

• Participating courts: E.D. Pa., D.N.J., the Third Circuit, and the Delaware Court of Chancery — 15+ chambers and Pro Se offices across the four jurisdictions.
• Infrastructure: Shared workspace per chambers; orientation materials maintained across model upgrades; onboarding scripts; a formal Memorandum of Understanding.
• Cadence: Monthly Zoom feedback meetings, plus one-on-one sessions with individual chambers as needs arise.
• Deliverables: Orientation guide, best-practices documents distilled from clerk interviews, activity reports.

What we’ve learned (so far)

The Testbed is producing a working empirical picture of where AI helps in chambers and where it creates risk.

Where chambers get genuine value:

• Section-by-section opinion drafting (full drafts hallucinate; section-by-section consistently works)
• Procedural history and factual background from uploaded pleadings
• Summarizing party arguments across multiple briefs
• Oral-argument question generation
• Plea colloquy and scheduling-order scripts (repetitive / template tasks)
• Timelines and charts assembled from case records
• Proofreading and citation formatting
• Digesting voluminous pro se pleadings
• Rewriting content for different audiences
• Custom GPTs for specific motion types or doctrinal tests

Where AI fails — and where the failure mode matters:

• Full opinion drafts (hallucination)
• Independent legal research and case law retrieval (high hallucination rates on case citations)
• Nuanced or cutting-edge legal analysis
• Writing-style mimicry
• Stream-of-consciousness organization
• Working with sealed or multimedia content

Emerging governance issues:

• AI-generated filings from litigants — both pro se and represented — with fabricated quotes and plausible-sounding but legally unsound arguments.
• Judicial-ethics questions reaching the Judicial Conference level.
• The transparency question — what should the public know about how chambers use AI?

Status

Active and expanding. Monthly meetings with 15+ chambers continue; the orientation materials have been updated through model generations; new chambers are being onboarded. The Judicial AI Portal launched August 28, 2026, and a training curriculum built from it is in development for fall 2026.

Sensitivity

The most external-facing of the Project’s workstreams. Chambers context is presumptively confidential. Public materials describe the program structure and aggregated findings — not the work of any individual chambers.