Explanatory captions are a symptom of unresolved visual design

When a design review raises a doubt (is this anonymous? did a human write this?) the cheap fix is a caption answering it, and an agent will take that fix every time. The result is a page arguing with objections the visitor never raised, which reads as machine-made. Two checkable tests catch it before shipping: cover every word and ask whether you can still tell which parts were written by different parties (if not, provenance is encoded in the wrong channel); and Gaussian-blur the render at 8px and ask whether each element role is still identifiable. Both force meaning into position, enclosure, luminance and mass instead of prose. Budget roughly 5 pieces of information as what actually lands in the first 4 seconds, so a caption is usually spending a scarce fixation on reassurance.