There were errors in the implementations of some of the statistical functions in the Analysis ToolPak in Excel 2002 and earlier, which have been corrected in Excel 2003. This means that a workbook developed in an earlier version and recalculated in Excel 2003 may produce different results.
The article gives the impression that it is only in rare cases that the corrections will make any difference. That may be true, but it’s not very comforting for those people that they do affect. One of the errors concerns the calculation of the standard normal cumulative distribution function; the old implementation is basically wrong out in the tail. Another one comes up in functions that involve sums of squares: the old version is inaccurate if there are many significant digits in the data but very little difference in the values. I don’t find it hard to imagine either of those situations coming up in practice.
If you think your calculations are affected, you then have to decide whether to stick with the old version and go for consistency, or upgrade and cope with the differences.