It is too bad that people posting "benchmark" code so rarely take the care to validate that the numbers that they are posting actually have any meaning behind them. For example, when you get a "Rate" like "3918166/s", then you should probably just throw away your benchmark results as having no practical meaning.

In this particular case, this was a hint that your $s = $s x 10_000_000­; line was completely pointless because of other aspects of your code. Fixing that problem (and reducing the string size significantly so it will finish before I need to go to bed), gives:

#!perl

use strict;
use warnings;

use Benchmark qw( cmpthese );
use English qw( -no_match_vars );

local $OUTPUT_RECORD_SEPARATOR = "\n";

print 'Perl version: ', $PERL_VERSION;

my $s = 'Aid bears out ' x 10_000;

print 'Length of string: ', length $s;

cmpthese( -1, {
   'lc'       => sub { () = lc($s) =~ m{ [aeiou][aeiou]           }xms
+g },
   '[Aa][Aa]' => sub { () =    $s  =~ m{ [AaEeIiOoUu][AaEeIiOoUu] }xms
+g },
   '[Aa]{2}'  => sub { () =    $s  =~ m{ [AaEeIiOoUu]{2}          }xms
+g },
   '/i'       => sub { () =    $s  =~ m{ (?i) [aeiou]{2}          }xms
+g },
} );
__END__
[download]

Perl version: v5.12.0
Length of string: 140000
           Rate  [Aa]{2}       lc [Aa][Aa]       /i
[Aa]{2}  50.7/s       --      -0%      -1%      -3%
lc       50.8/s       0%       --      -1%      -3%
[Aa][Aa] 51.5/s       1%       1%       --      -2%
/i       52.5/s       4%       3%       2%       --
[download]

which shows how the performance differences are minimal (and the actual difference you'll see in the performance of a real script will be even smaller than that).

So the "which will become noticable" assertion was wrong.

Update: Here, here! Learning++, (Good attitude)++.