Our Methodology
Why our ethical review is built differently.
Most AI ethics reviews ask the same handful of questions: is the model measurably biased, is it explainable, is there a data protection policy on file. These are necessary questions — but on their own, they’re not enough for AI systems operating inside or alongside public institutions: government agencies, courts and public benefits systems, humanitarian and NGO programmes increasingly run as quasi-public services.
For that category of AI, we believe the right standard already exists — not a new “AI ethics” checklist, but the body of thought public administration has used for over a century to ask exactly this question of human institutions: how does a large, rule-bound, discretion-exercising apparatus earn and keep legitimacy with the people subject to its decisions?
Four ideas from public administration theory, applied to AI
Legitimate authority. Following Max Weber’s classic work on authority, we ask whether a system’s decision-making power is formally delegated, bounded, and documented — or whether it’s simply borrowing the unearned credibility of “the algorithm said so.”
Real accountability. Using Mark Bovens’ accountability framework, we map every forum an AI system actually answers to, rather than accepting an internal audit log as proof that it’s accountable in any way that matters to the people it affects.
Bounded discretion. Michael Lipsky’s research on “street-level bureaucracy” showed that policy is really made case-by-case, in the gap between formal rules and frontline discretion. We ask where an AI system is exercising real discretion, and whether there’s a genuine path for a human to review and overrule a specific decision.
Public value. Drawing on Mark Moore’s “public value” framework, we ask whose value a system is actually optimising for — not just whether it’s efficient for whoever deployed it.
Why this is the right lens, not just a different one
These four ideas don’t replace the technical and legal checks every credible AI review already includes — bias testing, data protection compliance, documented limitations remain part of every review we do. What they add is a set of institutional questions technical checks can’t answer alone. For government agencies, courts, public benefits systems, and the humanitarian organisations we most often work with, this standard isn’t new or invented for the occasion — it’s the same one their institutions have always been expected to meet.
This page summarises our approach. The full methodology paper — with the complete assessment instrument and scoring framework — is treated as HCI intellectual property and is available to review clients and qualifying partners on request.