The AI Summary Every Board Reads Alike

Ajay Pundhir, senior AI leader & founder of AskAjay.ai. Believes AI should amplify human expertise, not replace it.

​Shiri Melumad and Jin Ho Yun ran seven experiments on 10,462 people in an attempt to answer a simple question. Does it matter whether you learn something from an AI synthesis or by searching for it yourself?

It turns out it does. Their paper in PNAS Nexus had participants write advice afterward. Advice written after the AI explanation was shorter, carried fewer distinct facts, and sounded more like everyone else’s. In one experiment, the pieces ran 84.58 words against 94.64, with named facts per piece falling from 0.718 to 0.464. Topical similarity between one person’s advice and the next close to tripled, and the edit distance between them shrank.

What The Study Did And Didn’t Test

The participants came from online panels and a university lab, and they were writing about vegetable gardening, healthy living and financial scams. Nobody was deciding a capital allocation.

The comparison was a single AI synthesis against several sources containing the same facts. In one experiment, the authors kept the facts identical across six articles and changed only the packaging.

So, the finding is narrower than the headline, and more useful for it. The authors’ explanation is that a synthesis has already done the sensemaking, so the reader does less of it. That has nothing to do with gardening. It is a claim about what happens when the thinking arrives pre-assembled, and about a group that all receives the same pre-assembled thinking.

I have sat through enough pre-read cycles to know where this lands. The AI summary is now the first page of the board pack, and often the only page anyone finishes. Put the same AI summary in front of nine directors, and you should expect what the PNAS Nexus study measured: recommendations that get shorter, cite fewer specific facts and start to sound like one another.