This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
Democratic institutions as the under-examined substrate of AI governance: a five-capacity framework (and a request for red-teaming)
By Alex Hakuzimana, Kayode Adekoya & Michal Kubiak (CORDA Fellowship, Project 19)
Summary: Most AI-governance attention focuses on mechanisms - evaluations, compute, standards. We focus on the institutional substrate that has to implement them: can democratic governments coordinate, learn, and adapt fast enough to retain meaningful oversight as capabilities accelerate? We propose the Democratic Coordination Framework (DCF) - five mutually reinforcing capacities - and, importantly, we're explicit that it is not yet validated. We'd value red-teaming of both the framework and our proposed validation agenda.
The claim, and its load-bearing assumption.
The argument rests on one assumption worth stating plainly: even well-designed technical governance fails if the institutions meant to implement it can't keep pace. If you reject that - if you think the binding constraint is purely technical, or that non-democratic actors will set the pace regardless - much of what follows matters less. We think the assumption holds, but we flag it because it's where the disagreement should start.
The problem.
Democratic institutions accrue authority slowly - deliberation, consultation, accountability. AI advances exponentially. van Kersbergen & Vis (2022) call the result a "temporal mismatch" producing an "exponential gap." Empirically, generative AI outran regulatory readiness, and the gap fills with regulatory arbitrage.
The recurring failure modes (from a qualitative review of OECD, UNDP, and academic sources): regulatory lag, institutional fragmentation, information silos, skills/capacity deficits, legacy infrastructure, and opacity that erodes legitimacy. Existing responses (anticipatory, experimentalist, deliberative, human-machine hybrid governance) each carry unresolved trade-offs - speed/accountability, expertise/participation, innovation/precaution, sovereignty/coordination.
The framework.
Five capacities, claimed to be mutually reinforcing:
Coordination across agencies, levels, borders.
Information integration - the foundational data layer.
Adaptability - converting learning into rapid, stable response.
Democratic legitimacy - the binding constraint on the other four.
The integration claim is relational: deficits propagate (silos starve learning → throttle adaptability); investment lifts the system.
This dependency structure is what distinguishes a framework from a taxonomy, and it's also the part most in need of testing.
What I'm not claiming.
I'm not claiming the five capacities are exhaustive or cleanly separable; some overlap (information integration and coordination especially).
I'm not claiming novelty over every existing taxonomy: the contribution is the integration structure plus the global-South vantage point, not the individual capacities.
I'm not claiming empirical support yet. This is a conceptual synthesis. That's the honest status.
The validation agenda (where I want red-teaming).
Retrospective case-study application: does the DCF explain real governance episodes better than alternatives, or does it just re-describe them? (Falsifiability worry: a framework that "explains" every outcome explains nothing. I'd welcome help specifying what would count as the DCF failing.)
Structured expert elicitation: 5–8 governance researchers/practitioners stress-testing completeness and the integration claim.
Longer term: an institutional self-assessment diagnostic.
The request.
Tell us where this is wrong. Specifically: Is the load-bearing assumption sound? Does the integration claim survive contact with cases? What would falsify it? Candidate cases and expert-panel volunteers also welcome.
Note on AI assistance: We used an LLM to help draft and structure this post. The research, framework, and arguments are ours; we have reviewed the content and are responsible for any errors.
Democratic institutions as the under-examined substrate of AI governance: a five-capacity framework (and a request for red-teaming)
By Alex Hakuzimana, Kayode Adekoya & Michal Kubiak (CORDA Fellowship, Project 19)
Summary: Most AI-governance attention focuses on mechanisms - evaluations, compute, standards. We focus on the institutional substrate that has to implement them: can democratic governments coordinate, learn, and adapt fast enough to retain meaningful oversight as capabilities accelerate? We propose the Democratic Coordination Framework (DCF) - five mutually reinforcing capacities - and, importantly, we're explicit that it is not yet validated. We'd value red-teaming of both the framework and our proposed validation agenda.
The claim, and its load-bearing assumption.
The argument rests on one assumption worth stating plainly: even well-designed technical governance fails if the institutions meant to implement it can't keep pace. If you reject that - if you think the binding constraint is purely technical, or that non-democratic actors will set the pace regardless - much of what follows matters less. We think the assumption holds, but we flag it because it's where the disagreement should start.
The problem.
Democratic institutions accrue authority slowly - deliberation, consultation, accountability. AI advances exponentially. van Kersbergen & Vis (2022) call the result a "temporal mismatch" producing an "exponential gap." Empirically, generative AI outran regulatory readiness, and the gap fills with regulatory arbitrage.
The recurring failure modes (from a qualitative review of OECD, UNDP, and academic sources): regulatory lag, institutional fragmentation, information silos, skills/capacity deficits, legacy infrastructure, and opacity that erodes legitimacy. Existing responses (anticipatory, experimentalist, deliberative, human-machine hybrid governance) each carry unresolved trade-offs - speed/accountability, expertise/participation, innovation/precaution, sovereignty/coordination.
The framework.
Five capacities, claimed to be mutually reinforcing:
The integration claim is relational: deficits propagate (silos starve learning → throttle adaptability); investment lifts the system.
This dependency structure is what distinguishes a framework from a taxonomy, and it's also the part most in need of testing.
What I'm not claiming.
The validation agenda (where I want red-teaming).
The request.
Tell us where this is wrong. Specifically: Is the load-bearing assumption sound? Does the integration claim survive contact with cases? What would falsify it? Candidate cases and expert-panel volunteers also welcome.
Full working paper:
Note on AI assistance: We used an LLM to help draft and structure this post. The research, framework, and arguments are ours; we have reviewed the content and are responsible for any errors.