How Do Organizations Learn? The Diffusion of Scientific Evidence on Generative AI

From the wave-0 evidence ledger

Recorded claims

Each claim states what the inspected source says, at the recorded location—bounded by its scope and graded by its confidence. Nothing here is a synthesis across works.

All 511 World Bank headquarters divisions were assigned across treatment/control conditions; seeding evidence with a senior rather than junior employee increased stated sharing likelihood by 14.2 percentage points, multiple infographic-link clicks by 15.8 points, and colleague recall of at least one study feature by 5.0 points.

confidence: high within study scholarly-full-text

Scope: Evidence diffusion inside 511 headquarters divisions; hierarchy treatment, not AI-tool productivity.

Treatments changing beliefs about colleagues' AI adoption or emphasizing evidence credibility had no detectable effect on sharing; overall diffusion was limited, with only 3% of colleagues recalling a relevant feature two to three weeks later in round 1 and none measured in round 2.

confidence: high for reported null/recall; medium generalization adverse-or-corrective

Scope: Diffusion of two GenAI evidence summaries inside the World Bank; not all organizational learning.

Boundaries

Limitations & independence

Recorded at coding time, carried with the work forever. A claim without its limits is not evidence.

Recorded limitations

One large international bureaucracy
Seed response rates were 56% and 38% across rounds
One round had an article-link tracking glitch and affected data were dropped
Recall is a demanding but narrow diffusion outcome
Working paper

Source independence

Pre-registered academic/World Bank field experiment inside the World Bank; focal organization is also one author's institution, with ethics/privacy approvals and disclosed reproducibility package.