In the late 1970s, researchers set up a table in a department store with four identical pairs of stockings and asked shoppers which pair they preferred. Most people chose the pair on the far right. When asked why, they talked about the knit, the sheerness and the quality. Nobody mentioned position, and when the researchers asked about it directly, people denied that it had made any difference.
The shoppers gave fluent, confident answers that sounded entirely reasonable. They just weren’t the real reasons. Researchers Richard Nisbett and Timothy Wilson used the study to argue that we have very little access to the processes behind our choices, and that what we report is a plausible story assembled afterwards from what we believe is generally true of people like us.
Later work pushed the idea further. In what became known as choice blindness experiments, Petter Johansson and Lars Hall showed people two photographs, asked which face they found more attractive, then secretly swapped the photos and handed back the one the person had not picked. Most people did not notice, and many went on to explain in detail why they preferred the face they had rejected. Other researchers have shown that we trust our own self inspection while judging other people by what they do, and that asking people to analyze their reasons for a choice could make the choice worse.
People do know plenty about themselves. Where we are unreliable is the why, and the gap shows up in everyday work. Someone calls themselves a detail person while skimming every document that crosses their desk. Someone says they want options and takes the first one offered. Someone insists they like thorough answers and then trims every one down to a paragraph. We believe these descriptions because they feel like reading a log of our own minds, when they are closer to writing a summary of it.
Behavior offers a way round this. What someone asks for, edits, accepts and sends back tells you more about how they work than anything they say about themselves. Most of us repeat the same small corrections to AI all week: shorter, less formal, answer first, give me an example. Those corrections add up to a pattern, and we rarely notice it ourselves.
When we do start to notice at least this friction, an obvious ‘fix’ is to tell AI about how you think. “I’m a x person”, “give me the answers more like this”. Such requests may be accurate – but they may not be. And the goal here should be to ensure that your own thinking is truly reflected and matched.
That is the thinking behind Frequency. It learns from how people work with AI, shows them what it found, and lets them change it. Since we started, we have watched it challenge some people’s notions of themselves and confirm others. Both reactions are useful, given that surprise shows where our own story and our behavior part ways, and confirmation helps prove some readings are on the mark.
Try it this week, notice what you change in the next three AI answers you get, and ask yourself what you would say that pattern is. Then see whether it matches what you would have told someone before you looked. It may – but it may not!