Between December 2023 and April 2024, one employee in each of the 511 divisions at the World Bank's Washington headquarters was invited to fill out what was billed as a brief study on AI use at the Bank. In three out of four divisions, partway through, the survey showed the respondent a recent research finding. In the first round it was a paper from Science reporting that professionals given ChatGPT finished writing tasks faster and turned in better work. Afterward each got an email with an infographic and the paper, built to be forwarded and carrying a note encouraging them to pass it along.1

The survey was cover for an experiment. Three researchers, one from the Bank's own research group and two from the University of Warwick, wanted to know what decides whether evidence travels once it is inside a large organization. Their main lever was rank. In some divisions the study went to someone at grade GG or above, a senior economist or a director, and in others to someone at GF or below, such as an analyst or an assistant. Two other levers ran alongside it. Some recipients saw survey figures suggesting most colleagues already used and liked ChatGPT while others saw the reverse, and in the first round some were told the paper had run in Science, that its authors were at MIT, and that researchers at Harvard Business School and Wharton had replicated it.1

The lever the experts ranked last

Beforehand, the team asked outside researchers, including economists at the Bank, to predict what would matter. They expected the peer information to do the most and hierarchy to be the least relevant of the three.1

Hierarchy was the only one of the three that measurably increased sharing. A senior recipient was 14.2 percentage points more likely than a junior one to say they would share the study, against the roughly 3 points the experts forecast, and the infographic link sent to a senior recipient was 15.8 points more likely to be clicked more than once. When colleagues in the same divisions were surveyed two to three weeks later, those in first round divisions where a senior person had received the evidence were 5 points more likely to name at least one correct detail of it.1 The peer figures did change beliefs, and people shown the favorable version guessed colleague adoption about 12 points higher. They made no detectable difference to whether anyone passed the study along, and the credibility details did not raise it either.1

How little moved at all

The number I keep coming back to is less flattering to everyone. Across both rounds, 55 percent of recipients said they were likely to share the study. In the first round, 3 percent of their colleagues could recall any detail of it a few weeks later, and in the second round, built on a different paper about ChatGPT and creativity, the researchers found no measurable recall. By four months, the difference in recall between senior and junior messengers was no longer statistically significant.1 That was evidence about the most talked about tool of the moment, inside an institution in the business of advising governments, so it is fair to wonder how far the average good idea gets inside yours.

What was not holding the junior staff back

The standard explanation for ideas that die low in an organization is fear, and it explains plenty, including the safety reporting gap where leaders blamed the systems while workers said speaking up did not feel safe. If junior people were holding back because they worried how an idea would be received, the favorable peer figures should have lifted their sharing more than anyone's, and so should the reassurance about the evidence. Neither did. The estimates are imprecise, and the authors say so, but they read the result as a sign that junior staff were likely not held back mainly by conformity or by doubts about the evidence. Asked in the second round what stopped them from sharing, people most often said their colleagues were short on time, and senior recipients were less likely than junior ones to name fear of conflict.1 The authors conclude that senior people were also "more likely to be heard and remembered," and that the results "point to potential bottlenecks when ideas originate lower in the hierarchy."1

If an idea travels about as far as the most senior person willing to repeat it, then the senior people in a building are running its distribution system, whether or not anyone gave them that job.

What a senior leader can do with it

The limits are real. This is a working paper, not yet peer reviewed, from a single organization, and one author works for the institution being studied. Not everyone who was invited took part, either: 285 people answered the survey in the first round and 176 in the second. The randomization chose which employee got the survey, and the senior group was more educated and more male and included more managers and existing ChatGPT users, so the authors allow that the gap may reflect who gets promoted as well as what holding a senior position does on its own. It held when they set aside everyone with a managerial title.1 How many layers a company should carry is a separate argument; I worked through what extra layers of corporate hierarchy buy and cost earlier this week.

For an individual leader, this is a question about habit. From a senior seat it is easy to treat repeating somebody else's point as a courtesy, something to do once the deciding is finished. This experiment suggests repetition by someone with rank is a large part of how an organization learns anything. When somebody two grades below you brings you something good, what happens to it next is mostly up to you, and my bet is that saying it out loud in the room where it counts, with their name attached, moves more ideas than another campaign encouraging people to speak up.

So think of the last useful thing you learned from somebody well below you on the org chart. Did you repeat it anywhere it could travel, and did their name go with it? And if nothing comes to mind, how would you ever find out which good ideas stopped with the person who had them first?