L33t-speak, Unicode homoglyphs, zero-width characters, spaced-out letters. Anyone determined to post something harmful gets past a word list in seconds. Supervisor reads what a message means, so there is nothing to spell around.
Word lists match exact strings. Evasion only has to change one character. AI that understands intent doesn't care how the words are spelled.
There is no normalization list to maintain and no arms race to lose. The model reads intent, so evasion tricks have nothing to trick.
Number and symbol substitutions read as the words they imitate. 5c4m, h8, k!ll: all understood.
Cyrillic, Greek, and fullwidth look-alike characters don't fool a model that reads meaning, not codepoints.
Zero-width spaces and joiners injected mid-word are irrelevant to semantic understanding.
The hardest bypass is using no banned words at all. Implicit moderation catches what a message means in context.
Evasion by switching language, transliteration, or mixed scripts is caught the same way.
One REST call per message. Keep your word filter for the obvious stuff; let Supervisor catch what slips through.
Paste your community's favorite filter bypass into the demo and watch it get caught.