See perlsec.

In this case they were looking for a way to bring tainted data in contact with a system command in a way that was absolutely safe, but would still trigger the taint test. If you hit the taint test you bomb out and get a false value. If you don't then you survive to the true return. Flip the truth and voila!

Incidentally I disagree with them on their comment about warnings. Consider the following test:

sub is_tainted {
  eval { () = (join('',@_), kill 0) };
  if ($@) {
    if ($@ =~ /^Insecure dependency/) {
      return 1;
    }
    else {
      die $@;
    }
  }
  else {
    return 0;
  }
}
[download]

Not only does this pass warnings, but it at least tries to handle the possibility of other things going wrong. (eg a platform where kill is not implemented...)