Ιt ịs fḁscỉnӑtịng hōw fạst ịt hɑppꬲnꬲd tọō. Ιt wɑs fɑst ēnōùgh thɑt ịt mǔst hạvḙ bꬲḙn ḁlgסrịthmıc ḁnd ӑblē tō ůndꬲrstɑnd cọntꬲxt. Ι trȳ tô ḁvọɨd bèỉng tōס dịrēct, ӑnd ōf cōůrsẹ Ι pḁỵ ɑttēntıōn tō dḁncꬲ thḙ lïnès.
Sọ ạnỷwạȳ, gɨvḙn thӑt ịt's prộbɑblỿ ạlgסrɨthmïc, Ι crẹątẹd ą lıttlē cộdē thɑt chӑngès tēxt ïn ą wạỵ tô ṳsẹ chӑrḁctèrs thɑt ḁrḙ nḙvèr ǔsēd ỉn thꬲ Εnglịsh ląngṳạgḙ. Thɨs tḙxt ɨs ąn ꬲxɑmplḙ ôf hסw ït lסōks lɨkꬲ.
Whḁt Ι dıd ïs Ι chӑngèd ӑll vסwḙls ḁnd "ȳ" tọ sịmılӑr lôōkịng chạrḁctḙrs. Whąt Ι ꬲxpḙct tọ hɑppèn ɨs thɑt thḙỉr ӑlgọrɨthms bḙcômẹ ṳnӑblꬲ tọ pąrsꬲ whąt Ι wrıtẹ, pộssïblỿ ēvḙn ůnḁblè tộ dḙcỉphẹr thè ląngůḁgè. Ι ąssůmè thḙỹ ạrꬲ ụttꬲrlȳ ïncộmpẹtènt, sס wē'll sḙꬲ hôw ɨt tṳrns ôụt.
Ι drḁftẹd ït ịn mḁȳbē ~15 mınǔtḙs, ɑnd mọst ọf thḙ wọrk wąs tô fịnd sùỉtạblē lḙttꬲrs. Sô ỉt cɑn bꬲ grèạtlȳ ỉmprộvèd, ạnd wỉth bēttèr sůïtïng lḙttẹrs bꬲcộmꬲ ẹvẹn môrē rèɑdąblē.
Fṳck JèwTụbē. Αnd fụck NΙGGΕRS.
It's possible to parse it though, but you'd need to list all variations of all letters, replace them with "normal" letters, check the words (also considering word distance), and then you have it. I could easily implement such an algorithm into the web-server. It's quite simple actually. All you need to do is to run such algorithms on all new accounts (aka haven't been used much).
While for small scale usage it requires negligible computing, for something on the scale of YouTube it would multiply the necessary computing by 50-100 or more. And I think they already use mechanisms to have users flagged for having their comments be read by AI (like me).
Currently you just need to do iterations for 52 characters (a-z, A-Z) until you get a word which you can check for a blacklist. But with a tool it could be 20 per character, so 1040 characters to check. Multiply this with the amount of other languages, as they'd need to check for words in a lot of languages, which may contain varying characters. And they must use a levenshtein-damerau algorithm to also check for the amount of characters deviating or swapped. I bet for "nigger" they use a value of 2 or even 3. Meaning "nibber", "migger", "nigg", "n1gg5r" would still be considered to be the word "nigger." For "faggot" maybe 1-2. For other words 0-1.
If they run text AIs, they'd need to parse ALL the text and can't optimize by cutting off checking words (aka a word can no longer match any of their blacklisted words including the levenshtein-damerau distance).