Sheena Iyengar and Mark Lepper set up a jam tasting booth with either twenty-four varieties or six and published the result in 2000. The large display drew more browsers; the small one produced roughly ten times as many buyers. That study became the standard citation for cutting options, and it is why "offer less choice" advice is everywhere. It is also why the advice is usually wrong. A 2010 meta-analysis by Benjamin Scheibehenne and colleagues pooled some fifty experiments on choice overload and found the average effect close to zero.

The effect did not vanish. It turned out to be conditional. Choice overload shows up when the chooser has no prior preference, when the options are hard to tell apart, and when nothing in the set clearly dominates. Give someone a favorite jam and twenty-four jars change nothing. That is the finding worth carrying into an AI workflow, because those three conditions are not incidental to AI generation, they are what it produces. A model asked for variants returns options near-identical by construction, delivered before anyone has formed a preference, with no dominant candidate because the model has no stake in one.

So the useful question is not how many drafts the model produced. It is whether the person receiving them already knew what good looked like. A reviewer holding a written criterion can work through a large batch quickly; the same reviewer without one stalls on a handful. The fix that follows is not a cap on output, which is the advice the jam study alone would give. It is writing the selection criterion before generating anything, which removes the preference-uncertainty condition and makes the volume stop mattering.