This is a very good question, one which we at the core lab have been asking as well. It appears to be an "undocumented feature"!! There is some chat about it online (try this for a start: https://www.open-bio.org/2010/04/30/illumina-q2-trim-fastq/).
We see it most often when there are non-random bases at the ends of sequences. Non-random bases (e.g. linkers) raise flags for the analysis software. The base calls may be (and usually are) good, but they don't look like the rest of the run to the instrument. We welcome feedback about your experiences with this so we can share details with Illumina. Please continue to keep us updated about your data analysis experiences.