An independent researcher named Derik Kauffman has unearthed a hard-coded table buried in Apple's iOS containing more than 100,000 words and phrases that the company's predictive text system will never suggest. The list, which Kauffman discovered after noticing inconsistent behavior in swipe-typing, is encoded directly into the operating system — not learned, not contextual, but deliberately suppressed by corporate decision. The expected entries are there: drugs, sexual acts, and — in a moment of corporate self-preservation — "iMessage isn't working." But the list extends well beyond profanity filtering. "Trayvon Martin," "Kashmir dispute," "Muslim extremist," "male chauvinist pig," and "Dictator Trump" are all suppressed. In Hebrew, "the occupied territories" won't autocomplete. In Arabic, "the Hamas movement" is blocked. These aren't obscenities. They're political speech. Kauffman described the system to Reuters as "a light-touch form of censorship," and the characterization is precise. Users can still type every suppressed word manually — nothing is truly blocked. But predictive text shapes what is easy to say, and what is easy to say shapes what gets said. A grey list doesn't ban speech; it adds friction to specific speech, selectively, across languages and geopolitics. The discovery arrives in the context of Apple's evolving autocorrect architecture. The company moved last year from n-gram statistical prediction to a transformer-based system more akin to an LLM. Reddit users noticed a spike in bizarre suggestions. But the grey list predates the new system — it's a hard-coded override sitting on top of whatever model Apple deploys, a manual layer of editorial judgment baked into billions of devices. Apple's statement was boilerplate: the keyboard is designed to avoid "inappropriate or offensive" suggestions, and users can always type words manually. This frames the issue as content moderation — keeping the keyboard polite. But moderation implies a moderator, and the moderator here is a proprietary list that no user can inspect, appeal, or modify. The company decides which political phrases in which languages deserve friction, with no transparency about criteria. The deeper structural question is about default power. Predictive text is infrastructure — it mediates billions of daily communications. When the default behavior of that infrastructure is shaped by an undisclosed editorial list covering geopolitical terms in multiple languages, the company operating the infrastructure is exercising a form of governance over expression. Not censorship in the legal sense, but something that functions like editorial control at planetary scale. A brief 2024 glitch — where Apple's dictation feature briefly displayed "Trump" when users spoke the word "racist" — hints at the messy interaction between hard-coded rules and learned models. Apple fixed it quickly. But the grey list itself, running silently across every iPhone on Earth, was not a glitch. It was a policy choice, and until Kauffman found it, nobody outside Apple knew the policy existed.
