What AI Can’t Decide

What the zero-clustering result was showing

Run a perception test with a human audience and something reliable happens: the responses cluster. People who have never met, responding individually to a set of visual elements, land in the same places. Every human test does this, without exception.

Point the same test at AI agents and, early on, nothing happened. Responses scattered. A brownout, with no pattern of preference anywhere in the results.

Figure 1 · The same test, three audiences

The straightforward read is that these models weren't making decisions about what things mean. They were producing plausible outputs one at a time, and there was no stable underlying position for the outputs to be consistent with. Human clustering isn't people agreeing with each other. It is separate people bringing a shared cultural read to the same stimulus, arriving at the same place independently. That shared read is exactly what the test captures, and it was absent.

The models improved, and the problem got harder to see

The picture has changed over the past year. Current models, ChatGPT and Claude in particular, will make preference choices now. A pattern shows up in their responses where before there was noise.

The pattern doesn't match the human one, and the gap isn't marginal. What a model produces on a set is dramatically different from what a real audience produces on that same set.

And these models have got very good at writing a compelling narrative about the decisions they made. The reasoning is articulate, internally consistent, and wrong. A sophisticated and convincing liar is more expensive than an obvious one, because the brownout was at least legible as a non-answer. A confident, coherent explanation of an incorrect preference is the kind of thing that survives a meeting.

Where AI genuinely earns its place in a brand project

None of this means keep it out of the work. It means point it at the job it is good at.

• Finding patterns in qualitative data you collected from people. This is the strong use. Take the results of a perception test and let a model work across them: what the elements that scored high have in common, what the rejected ones share, and where those two collections diverge in ways a person scanning a board would miss.
• Adding context to the data. Scraped comments from a social post about the brand, dropped into the mix alongside test results. Market-specific material and subject matter knowledge you hold that the raw responses don't contain. Layering that in produces better answers about how people are responding and why, which is the harder half of the question.
• Working on your own data rather than the world's. The useful analysis here doesn't need the largest general model available. It needs a model pointed at your material, with real context and real parameters around it.

Where it falls down is predictable: high-level, superficial prompts. Asking it to produce visuals or write copy with no human supplying depth of context, explicit guidance or constraints to work inside. AI is as good as the person running it and the method it is running inside. Without a methodical approach to design around it, it produces confident output with nothing underneath.

Contrast is what makes a pattern legible

There is a habit worth borrowing here, and it predates the AI question by years.

Most teams show a board and ask a general question, or they pull out the things people liked and look at those in isolation. The rejections get thrown away.

The more useful move is to capture what people don't like and hold it against what they do. Positives in isolation give you a pile of nice images and a weak signal. Positives set against a foil of negatives make the dividing line obvious, and the dividing line is the thing you design against. It is the same move that turns a brand audit from an inventory into a finding.

Figure 2 · One collection describes. Two explain.

This is also why the pattern-finding works. Give a model one collection of liked material and it will describe it. Give it two contrasting collections and ask what separates them, and you get the reason, which no single swatch can tell you on its own. The contrast produces the why.

What to tell the client who says “we’ll just ask ChatGPT”

It is a fair question to ask now, and it deserves a straight answer rather than a defensive one.

The model will answer. It will produce a preference, and it will explain that preference well. What it won't do is match what your audience does, and you have no way to tell from the output which parts are right. The explanation reads the same either way. You are not getting a read on your market, you are getting a well-written argument produced by something with no stake in your market and no membership in the culture your audience is reading your brand through.

Figure 3 · The order that works

Where those models help is on the other side of the data. Get the response from real people first. Then use AI to work through what came back, layer in the context you have, and find the patterns across it. That order works. Reversing it substitutes a persuasive narrative for evidence, which is the exact failure mode that fills a conference room with confident opinions and produces a rebrand nobody outside the building likes.

Frequently asked questions