The main reason is that we cannot do a full-fledged proper alignment, that
would take much too much time. So we must take short-cuts to save time. One
short-cut is to do only pair-wise alignments and to combine those all the time
pair-by-pair (this is how Clustal does it). This alignment order makes that
mistakes made early-on can never be corrected again.
Another problem is that near the ens of sequences you get artefacts because gap penalties my cause gaps to show up (or not) where you dont want/expect them.