Update: took out reverse in favour of a random shuffle

while (<DATA>)
{
 $_=~s/([.,\s]*?)(\w{1})(\w+)(\w{1})([,.\s]*?)(?{$j=sub{$x=shift;for(1
+..50){$n=int(rand(length($x)-1));$x=reverse(substr($x,0,$n)).substr($
+x,$n);}return $x };$x=$j->($3)})/$1$2$x$4$5/g;
  print $_;
}
__DATA__
Down that path lies madness. On the other hand, the road to hell is pa
+ved with melting snowballs.
Although the Perl Slogan is There's More Than One Way to Do It, I hesi
+tate to make 10 ways to do something.
And don't tell me there isn't one bit of difference between null and s
+pace, because that's exactly how much difference there is.
Randal said it would be tough to do in sed. He didn't say he didn't un
+derstand sed. Randal understands sed quite well. Which is why he uses
+ Perl.
As usual, I'm overstating the case to knock a few neurons loose, but t
+he truth is usually somewhere in the muddle, uh, middle.
Of course, I reserve the right to make wholly stupid changes to Perl i
+f I think they improve the language.
[download]

Update: Slightly OT for perl, but IIRC, this phenomenon is theorised to derive from 'orthographic word recognition', recognising the entire word by it's overall shape. As long as the outline of the word is roughly the same, the brain can make a pretty good stab at recognising it. It will probably also be helped by a priming effect, whereby the words most likely to follow a given word will be more 'activated'. There's some research (somewhere that is, I'm trying to track it down) suggesting that the more different the outline is from the original (especially through the placement of rising and falling letters like p and l) the more difficult it will be to recognise.