The problem is that most web browsers submit multi-line contents of text areas teminated by \r\n characters (windows line terminators). These get exported properly to XML on backup. The problem is that according to XML spec upon parsing the XML file the line terminators are normalised to \n:
http://www.xml.com/axml/target.html#sec-line-ends

This means that next time an issue is edited the values appear different as the \r characters do not exists in the database but are sent by the web browser.

This problem surfaces when one exports JIRA data to XML and then reimports the data.