iOS Keyboard Blocks More Than 100,000 Words and Phrases, Researcher Finds

Researcher Derik Kauffman has found more than 100,000 words and phrases that iOS blocks from predictive text and swipe-typing. Reuters published its exclusive report on the findings on November 6, 2026. The report was written by Raphael Satter. According to The Verge, the blocklist covers every major language.

The block operates at the suggestion layer, not the input layer. A blocked word can still be typed one letter at a time, but it never appears as a predictive suggestion. Swipe-typing is less predictable. Swiping a blocked phrase can cause iOS to alter it, so “left wing conspiracy” can come out as “left wrong conspiracy.”

The list includes “illegal immigrant,” “dictator Trump,” “sleepy Joe,” “covid was a hoax,” and “left wing conspiracy.” It also blocks “bad girl” but not “bad boy.” Kauffman has published the full list online, with separate lists for different languages.

Kauffman used OpenAI’s Codex and Anthropic’s Claude Code, AI coding agents that can read and search software, to help locate the phrase list inside iOS. The Verge reported that the list was stored as a hash table, a lookup structure that scrambles entries into coded keys, and that Kauffman converted it back into readable text. A second researcher independently obtained the same iOS blocklist.

Apple said its keyboard is designed to avoid potentially inappropriate or offensive suggestions. It also said iPhone users can still type what they want or override predictive text suggestions.

In my view, the distinction that matters most to users is the one between suggestion and input. The blocklist does not stop a word from reaching a document, a message or a search field. It removes that word from the menu the keyboard offers while the user is typing. That is a narrower intervention than a filter on what gets sent, but it is still an editorial decision made inside the operating system, and until now its contents were not readable by the people it affects.

The broader context here is the tooling. Locating and decoding a phrase table inside a shipped binary used to require considerable reverse-engineering time. Kauffman’s account describes AI coding agents as part of that work. The same capability that lets a developer audit a codebase can now let an individual audit a vendor’s shipped behaviour, and that changes who can check how a platform handles language.

The asymmetry between “bad girl” and “bad boy” is the kind of result that a list of this size, whether automated or hand-curated, tends to produce. Whoever builds such a list inherits its gaps and its judgments. A published list makes those judgments available to be argued over in public. Readers, developers and linguists can now examine the entries directly, rather than inferring them from the keyboard’s behaviour.

Predictive text was built to save keystrokes and reduce friction. The override remains in place, and the list is now visible. Whether a platform should decide which phrases its suggestion engine will offer, and how it should explain those decisions, is the question the published lists have put on the table.