Assuming each user will always encrypt to the same value, this still loses to statistical attacks.

As a simple example, users are e.g. more likely to vote on threads they comment in. With data reaching back far enough, people who exhibit “normal” behavior will be identified with high certainty.

In my experience, it is good at simple to medium complexity regex. For the harder ones it starts being quite useless though, at best providing a decent starting point to begin debugging from.