A developer known online as "I" followed a tip from a tweet by Nick and opened a SQLite database sitting quietly at ~/Library/Keyboard/VulgarWordUsage.db on every Mac. The schema is revealing: it tracks vulgar words per app, per recipient, with usage counts and timestamps. The purpose is granular — your keyboard can learn that you say "fuck" in Messages to one friend but not in Mail to your boss. The unique index is on (app, recipient, word, reading). Craig Federighi announced the underlying feature at WWDC 2023 with the now-famous "ducking" line, but nobody outside Apple had mapped the full machinery until now. The database itself turned out to be a dead end on the author's machine — empty, with no rows despite deliberate attempts to trigger recording via TextEdit. But the database is only the counting layer. The real payload is upstream: Apple's system dictionary, readable through private C functions in Lexicon.framework under /System/Library/LinguisticData/. Every word in the enUS dictionary — 199,019 entries — carries two 32-bit flag fields. Of those, 4,271 words are tagged with one of three tiers: vulgar (1,050 words like fuck, shit, cunt, wtf), offensive (663 words like kike, fag, retard, Oriental), and sensitive (2,558 words like damn, LSD, abortion, gun, suicide, pee, poo). The tier system maps to three adjacent bits in the meta flags: 0x00800000 for vulgar, 0x01000000 for offensive, 0x02000000 for sensitive. On top of the tiers, up to eleven topic flags further classify words by domain: sex, drugs, religion, politics, violence, and more. In enUS, tiers never overlap — a word gets at most one. The author inferred all meanings by grouping words by bit pattern and reading samples until each cluster's theme became obvious, aided by an enum case called offensiveVulgarSensitive found near the vulgar-word code in the shared library cache. The editorial choices are where things get interesting. "Oedipus" is flagged vulgar. "Tom" and "Mick" are offensive (presumably as slurs for specific groups). "Pee," "poo," "niggle," and "underling" are sensitive. Most striking: "Trump" is the only living politician in the English sensitive tier, sitting alongside Hitler, Mao, Osama bin Laden, and the Taliban. Biden, Obama, and Putin carry no flags at all. These are not algorithmic classifications — someone at Apple made these calls, and the decisions propagate to every keyboard on every device. There is no public API for any of this. NSSpellChecker, UITextChecker, NaturalLanguage, and UILexicon expose nothing about vulgarity flags. SensitiveContentAnalysis covers images and video, not text. The only places "profanity" appears on developer.apple.com are App Store age ratings and review guidelines. The author published working C and Swift code to read the flags — both read-only, using either direct linking against the private framework or runtime dlopen/dlsym — and dumped the full flagged word lists for 15 languages in a public gist. The infrastructure reveals a design philosophy: Apple wants autocorrect to be context-aware about profanity without ever making the classification system visible or auditable. Symbol names in the shared library cache spell out the flow — setcurrentappid, setcurrentrecipientids, incrementusagecount, canautocorrect, filtercandidates — and a synchronization symbol suggests the usage database syncs across devices. The keyboard engine asks whether every vulgar word in a candidate is allowlisted for the current context; if not, the clean alternative wins. The 15-language dump means this isn't just an English-language editorial problem. Every language Apple supports has its own curated list of words deemed too dangerous for autocomplete to suggest unprompted. The classification decisions — which words, which tiers, which topics — are made by unknown people inside Apple, with no external review, no documentation, and no mechanism for users to inspect or contest the flags. The system works exactly as designed: invisibly.
