CASE STUDY
(Meta)
Oversight Board
Designing Responsible AI for a Global Governance Body
FASTER CASE TURNAROUND
~30% Faster
INCREASED CASE COMPLETION
~18% More cases
GLOBAL MEDIA ATTENTION
High-visibility
HUMAN DECISION AUTHORITY
100%
01
The Stakes
The Oversight Board was created to independently review Meta’s most difficult content decisions. It launched under intense media attention and public scrutiny.
Requirements:
-
Institutional credibility
-
Global transparency
-
Governance integrity
-
Binding decisions
-
No perception of AI bias or automation overreach
Meta entrusted my team to design:
-
The public-facing platform
-
The internal reviewer system
-
A responsible AI layer integrated into decision workflows
Failure would be visible — globally.
02
The Problem
Board members were reviewing complex, precedent-setting cases involving:
-
Hate speech
-
Political content
-
Cultural nuance
-
Legal interpretation
Challenges:
-
Dense policy language
-
Heavy precedent analysis
-
Manual drafting of legal opinions
-
Cognitive overload
-
High case volume
Board members are not all lawyers — but rulings are binding.
03
Research & Foundation
We studied:
-
Case lifecycle timing
-
Precedent discovery friction
-
Policy misinterpretation patterns
-
Drafting iteration cycles
-
Escalation and reversal trends
We also leveraged prior AI workflow systems we built for Meta’s GRO (Global Resarch Org) teams:
-
Research citation surfacing
-
Knowledge graph retrieval
-
Context-aware recommendations
-
Document synthesis tooling
These patterns proved AI could augment expert reasoning safely.
Core Insight:
AI must support structured judgment — not automate it.
04
The Challenge
Three non-negotiables:
-
AI cannot make decisions.
-
All recommendations must be explainable.
-
Humans retain full accountability.
This required:
-
No automated rulings
-
Reviewer must actively choose outcome
-
Expandable explainability
-
Clear distinction between suggestion and authority
-
Logged AI usage for auditability
-
Legal-safe language generation guardrails
AI was structured support — not governance authority.
At the same time, the public platform had to communicate:
-
Independence (from Meta)
-
Legitimacy
-
Clarity
-
Design maturity
Designed for credibility and institutional trust:
-
Calm editorial tone
-
Accessible global design system
-
Structured information hierarchy
-
Motion used with restraint
-
Transparent decision publishing
05
AI Case Assistant
AI embedded at decision friction points:
-
Policy Interpretation
Surfaces relevant sections automatically. -
Precedent Discovery
Auto-suggests contextually similar cases. -
Explainable Mode
Confidence score + reasoning breakdown. -
Contextual Signals
Tone, political cues, cultural references surfaced. -
Draft Opinion Assistance
Structured draft language aligned to precedent. -
Predictive Assistance
Anticipates next reviewer needs.
All decisions require explicit human confirmation.
06
My Role
“I led experience design across both surfaces..”
Scope included:
-
Defining human-in-the-loop AI principles
-
Designing contextual AI assistance inside the reviewer workflow
-
Translating LLM capabilities into governance-safe interactions
-
Leading UX/UI direction for the public platform
-
Establishing scalable design system alignment
-
Aligning legal, policy, engineering, and executive stakeholders
This was platform-level product design leadership under public scrutiny.
07
Outcomes
-
Moved from manual review to AI-augmented
-
Context surfaced instantly
-
Structured drafting support
-
Reduced cognitive load
-
~30% faster turnaround
-
Increased case completion rates ~18%
-
More consistent precedent alignment
-
Reduced reviewer research time by ~80%
-
High-profile, successful launch under global scrutiny
It was AI embedded into a globally visible governance system — delivered successfully under pressure.
