Orionis GroupControl Erosion
Can the institution still understand, steer and reverse an AI-mediated process as automation becomes part of its everyday work?
Explore control capacity →Jean-Yves Djamen’s research examines how AI changes institutional decision-making, responsibility and oversight. A system can perform its assigned task well while the institution gradually loses the capacity to understand, redirect or reverse the process that uses it.
Institutional AI alignment asks how to recognize those changes and what evidence would justify a response.
Three research questions
Quiet Failure
Would oversight detect a departure from purpose when behavior remains stable and violation alerts stay quiet?
Explore accountable monitoring →Misalignment without misbehavior
How can locally adequate operation contribute to divergence from an institution’s broader objective?
Read the published research →From governance models to institutional dynamics
The programme builds on the 2008 book Gouvernance et Intelligence Artificielle — Bâtir une administration performante and its AuDiT model. The 2018 paper carries definitions, diagrams and formal operations into digital public administration. An ongoing continuation manuscript examines assignment, execution and accountability under algorithmic delegation.
Follow the intellectual trajectoryFrom research to tools
AuDiT prototype
Explore the relationships among administrations, actors and missions through assignments linked to legal source passages.
About the prototype →Proposed Observatory
A research initiative for studying changes in institutional control, responsibility and alignment over time.
About the proposal →