> likely files (eg. ASCII, obvious patterns, etc) become smaller

Likely files for a real human workload are like that, but if "most inputs" is talking about the set of all possible files, then that's a whole different ball game and "most inputs" will compress very badly.

> unlikely files become bigger

Yes, but when compressors can't find useful patterns they generally only increase size by a small fraction of a percent. There aren't files that get significantly bigger.