Explainable AI for governance and public-sector decision support
# Overview
A decision the state cannot explain is a decision the state cannot defend.
Explainable AI is often pitched to the public sector as a nice-to-have — a
layer of transparency bolted on to improve trust. This brief argues the
opposite: in government, the ability to explain a decision is a precondition
of the authority to make it at all.
The state is not a recommender engine. When it allocates a benefit, flags a
fraud risk, or sets a sentence, it exercises coercive power over people who
did not opt in. Legitimacy there rests on an old bargain — reasons, a route
to contest, and a human who can be held responsible. An unexplainable model
breaks all three at once.
## What the full brief covers
- **Why the public sector is different** — coercion, due process, and the
slide from genuine decision *support* to *delegated authority*.
- **The case files** — the Dutch childcare benefits scandal, COMPAS, and
Robodebt, and why opacity (not complexity) caused the harm.
- **The explanation toolkit** — interpretable-by-design vs. post-hoc methods
(SHAP, LIME, counterfactuals), and the faithfulness problem.
- **What an explanation is for** — different audiences, automation bias, the
attribution gap, and contestability as the real test.
- **The legal architecture** — GDPR, the SCHUFA ruling, and the EU AI Act's
duties for high-risk public-sector systems.
- **Governance scaffolding** — algorithmic impact assessments, transparency
registers, audit, and public engagement.
- **The emerging-economy frontier** — leapfrogging without re-importing the
failures, with relevance to public-interest systems like road safety.
- **A builder's & policymaker's playbook** — eight commitments for decision
support that survives scrutiny.
## Read the full research
The complete brief — including all case studies, the legal analysis, and
28 cited sources — is available as a standalone document.
**→ [Read the full research brief](INSERT_LINK_HERE)**
---
*Prepared by the ECADEL LABS Research Team. Legal provisions are summarised
for orientation, not as legal advice.*
ECADEL LABS Research Team
AfricaGovernanceAI
Read The Mirror Has a Margin: AI Bias and Fairness in African Training Data Contexts
Bias in African AI is less about broken algorithms than absence: African faces, languages, bodies, and genomes are barely present in the data that trains modern systems — so models return a smaller, paler world and call it neutral.
This brief maps where African realities fall outside the frame, the mechanisms behind it, and documented harms across vision, language, and health — then sets out the data-sovereignty stance for building fairly from within the continent.
ECADEL LABS Research Team
AiResearch
Read Research Note
The Offline-First Imperative: Why African AI Must Work Without the Internet
This research note argues that the dominant AI deployment paradigm — cloud-first, connectivity-dependent — is structurally incompatible with the infrastructure reality of most African markets. We propose a set of design principles for offline-first AI systems and evaluate existing frameworks against these principles, with reference to ECADEL GROUP's Kiongozi AI deployment in Uganda as a case study.
ECADEL LABS Research Team
AIOffline-FirstAfricaMachine Learning
Read Position Paper
Mobile Money as Intelligence Infrastructure: A Framework for African Financial Data
This position paper proposes a conceptual framework for treating mobile money transaction data as a foundational financial intelligence layer for African economies. We argue that the structural characteristics of mobile money — ubiquity, SMS-based accessibility, and informal economy penetration — make it uniquely suited as the primary data substrate for African financial AI systems, distinct from the bank feed model that underpins Western fintech.
ECADEL LABS Research Team
Mobile MoneyFinancial DataAfricaFintech
Read 