Our AI research falls into two streams. The first is methods and architecture: how language models behave under the conditions of substantive knowledge work, what structured reasoning frameworks do to outputs, how scaffolding systems can be built and tested against specific failure modes identified through deployment, and what technical tooling that work requires. The second is institutional: what these systems mean for public administration and state capacity, how they interact with the conditions of lower-resource governments, and where the policy and governance implications remain undertheorised.
A language model asked for a policy brief, an assessment, or a technical report produces something with the expected sections, the conventional framings, and recommendations in structurally sensible form. Read against what the document is actually for, a specific reader making a specific decision, a particular failure recurs: the structure is present and the reasoning that would justify it is not. We call this the intent deficit. It is the gap between what a model can produce when explicitly directed and what it delivers by default.
The deficit is not a capability failure. A model that introduces citation errors when generating freely verifies citations accurately when auditing is a required step; one that writes audience-agnostic prose recalibrates when asked to construct an audience model first. The capability is present. What varies is whether the conditions of the task elicit it. Two structural properties of generation explain why: pattern-completion pulls every decision toward the training distribution’s genre average, and attention attenuation outweighs early instructions as a long session fills. Prompt engineering improves starting conditions and does nothing to maintain them, which is why it does not close the gap.
Six failure modes recur often enough to name: register inheritance, over-literal instruction-latching, evidence substitution by evaluative language, plausible citation, audience-unaware design, and content loss during format conversion. They are domain-shaped and share one cause: decision without active reasoning from purpose, audience, or context.
The pattern is not new. Surface form adopted without the reasoning that connects form to function is what Andrews, Pritchett, and Woolcock call isomorphic mimicry in the state capability literature: governments adopting the organisational charts and procedures of capable states without the underlying function. A model producing a genre-correct brief without reasoning about why this reader needs this structure is doing the same thing at a different scale. That parallel is what makes the series matter beyond AI. If experienced judgement is the scarce input in both cases, then a tool designed to externalise expert reasoning during real work, visibly and contestably, is a candidate piece of capacity infrastructure for contexts where extended apprenticeship is unavailable. We put that forward as a proposition with a research agenda, not as a finding.
Part 1 diagnoses the deficit. Part 2 describes the reasoning architecture built to address it, deployable as a single context document. Part 3 develops the capacity-building case. Part 4 shows six paired cases of the failure and its correction.
A connected sequence on the structural failure modes of AI in professional knowledge work and the methods developed to address them.
We also publish a technical introduction to how language models work – covering architecture, behaviour, and the conditions under which these systems fail – written as a reference for policymakers. It is in the governance section: Introduction to language models for policymakers →