Your best people know more than they can tell you

Telling a male chick from a female on the day it hatches is famously difficult. The differences are tiny, yet professional sexers sort around a thousand chicks an hour at over 98% accuracy. For a long time they couldn’t say how they did it. Trainees learned by guessing while an expert looked over their shoulder and said yes or no, for months, until they could do it too.

The philosopher Michael Polanyi had a phrase for this: “we can know more than we can tell.” A lot of what skilled people know lives in their judgment and never makes it into rules they could write down. Economists call it Polanyi’s paradox, and it is a big part of why some work has been so hard to automate.

Every brand has someone like this

Most brands depend on the judgment of one or two people. It might be a founder who can tell within a second that a line is wrong, or a head of brand who says “that’s not us” and is right, even though the guidelines don’t say why. Their taste is among the most valuable things the brand has, and almost none of it is written down.

When a team brings in AI, it hands over what can be told: the guidelines, the tone-of-voice document, the product facts. The AI does its best with them, and the results sound like a brand guideline.

What happened when we gave AI more to go on

We expected more information to help, so we tested it on 20 briefs. Some of it made the ideas worse.

A specific product fact cut the number of strong, original ideas nearly in half, because it pulled the writing toward the literal product. Brand rules hurt when they were written as things to include, since the AI dutifully illustrated them, and stopped hurting when they were written as limits to stay inside. Checking ideas against the written description of the brand’s voice told us nothing about which were good.

Guidelines still matter. They just don’t contain the part of a brand that makes its work good.

How taste gets passed on

The chicken sexers have one more lesson. In 1987 two psychologists, Irving Biederman and Margaret Shiffrar, studied an expert who had sorted 55 million chicks over 50 years. Working from what he did rather than what he said, they found one visual cue that settled most cases. Novices given about a minute of instruction on that cue went from barely better than chance to scoring as well as professionals.

So taste can be captured. It comes from watching what skilled people choose and reject and working out why, which is a different thing from asking them to write rules.

Much of what we build for brands works this way. We keep a record of what your best people choose and turn down, and why, and the AI learns from it until it starts choosing the way they would. When someone leaves, the guidelines stay behind. Their judgment stays too, if something has been learning from it.