I answered a related question here that may be helpful.
Some initial notes:
- Your first bullet point ("99%") is to my reading talking about not the total genome size but the data that they actually assembled- e.g. for 99% of nucleotides, we are 99.99% that the nucleotide at that position is not some other nucleotide, but there may be other positions not included in the denominator that are simply not in the assembly in the first place. In other words, the existence of the position in the assembly and our confidence of the accuracy of a nucleotide call for such a position are two different (though related) phenomena.
- Also, you are comparing a 2003 estimate to a 2017 estimate. We learned a lot in those 14 years.
To answer the last and most specific question about why this number is hard to know, there are a variety of reasons why:
There is no single number that is the "actual" true sequenced percentage of the human genome. In this I strongly disagree with commenter Konrad Rudolph (technological change is now giving us telomere-to-telomere assemblies of human chromosomes! The question is not how much it's possible to sequence and assemble but more whether it's the worth the time and expense to do so.). No two genomes are exactly alike, and that means not only that the nucleotides are different but that the positions are different. So even if we sequence 100% of your genome, that may only include 99% of my genome. These variable regions are, as suggested in the comments, largely repetitive regions but also just big chunks that are missing; to learn more about this google "structural variation" (here is an older review on the topic).
As mentioned already, technological advances have changed how we view the linear sequence of the genome. Earlier estimates were limited by available technology of the time; now that we have 3rd-generation sequencing tech like nanopore and pacbio, we are learning more about the extent of our ignorance. In that sense, it is unsurprising that the denominator has changed/increased.
As commenter Jeppe Nielsen states, all of these are statistical estimates based on available data. As available data increase and the statistical models that we use to estimate change, thus do the estimated outcomes change.
Related to (1): Different populations have different chunks of genome that are unique. The initial sequencing project selected a small number of humans, who were not representative of all human diversity. Efforts to fully characterize the human pangenome are ongoing.
Sometimes contaminants sneak into genome assemblies, which have to be removed. So for every sequence in the human genome assembly, especially if it's fragmentary or recently added, there's a nonzero chance that it's not actually part of any human genome.