Our Chief Meaning Officer, David Intrator, had an interesting exchange with Claude a few weeks back.
David was working on an essay for a personal Substack newsletter he publishes about music and culture. After drafting it, he asked Claude to check for typos, spacing, questionable diction, and “anything weird or illogical in the overall structure.”
After catching a few misspellings and questioning a few word choices, Claude said the essay was ready to publish. But on further review, David realized it had a major structural flaw: it was really two essays in one. There was no Unifying Idea holding the piece together.
When he brought this up with Claude, here was the response:
“I’m better at evaluating pieces in isolation than I am at seeing the whole from the outside, especially when I’ve been involved in building it. Investment in the process makes me a worse judge of the product.”
It’s a small admission with large implications, not just for writing, but for the work of market research.
A Unifying Idea is what makes a story a story, rather than a mere list of observations and insights. It’s the thread that tells you what belongs and what doesn’t, what to lead with and what to cut, even when each individual finding is accurate and may be of interest on its own.
Without the Unifying Idea you can still have correct data and sound logic, and still have nothing that actually makes a point and is of genuine value to your client.
Consider study design. A well-run study can generate dozens of technically valid findings and still fail to answer the question that actually matters. This is often due to being unclear as to what the research is actually about. What are we actually looking for? What question does the study need to answer? What decision is the client trying to make?
The same is true in interpreting data. What separates a sharp insights report from a data dump is the same thing that separates a good essay from a rough draft: someone stepping back far enough to see the whole and deciding what it’s about. Which findings are signal, which are noise, and what single idea ties them together meaningfully, into something a client can act on?
This is precisely where AI struggles, and Claude’s own words explain why. It’s good at evaluating pieces in isolation: a chart, a stat, a paragraph, a cross-tab. But it’s much weaker at seeing the whole from the outside, especially once it’s been “involved in building it.”
Investment in the process, as Claude put it, makes it a worse judge of the product. That’s a structural problem, not a minor technical limitation. And it shows up at exactly the moment researchers need judgment most, when everything is technically correct and something is still missing.
This is not an indictment of LLMs per se. Without doubt, they’re useful and are here to stay. They can check our logic, tighten language, summarize finding, note inconsistencies. But turning a series of insights into a coherent, meaningful narrative, which is ultimately what we’re hired to do, is another story altogether.
And, for the moment at least, it’s only something human judgment can deliver.