Instead, we are learning the "unknown" DNA performs biological functions due to its physical nature, such as physically blocking things from binding.
Imagine if a small section of a hard drive was so strongly magnetized that it repulsed the read head - if you were trying to translate it into binary it would appear to be nonsense.
Of course you are accurate as well: if we subtract the protein coding and the known non-coding but useful bits the rest is almost certainly absolute junk... duplications, bits of retroviruses, etc. But it would be inappropriate to claim we are absolutely certain none of it is useful. Will we find out 50% is useful? Obviously not. But I would rather say "current evidence suggests at least 97% is true junk; we are not likely to find large sequences with hidden uses at this point".
My rough read of biology is that it could use the same mechanism to lock down both.
Why is a viral insertion any different? Why is a recombination error any different?
Looking at the hex dump of RAM, some repetition is garbage, some is lookup tables with multiple entries into the same routine, some is initialized scratch memory for distant functions. It'd be premature to assume it's worthless.
Software code analogies to biological encoding break down, of course, because code is built intentionally and biological systems aren't. On the flip side, that very lack of intentionality in biological systems means it's risky to assume something is junk without a really holistic understanding of the context.
And I also agree with the bad assumption on the remaining 8% not being significant when we knew that it was structurally different. Less than 8% of an ELF is the header, but boy howdy will that thing not run well if you cut it off.