Connect with us

NEWS

A Hidden iPhone Blocklist Steers What Predictive Text Finishes

Apple’s iPhone keyboard still lets you type flagged phrases, but a hidden multilingual blocklist decides which political lines predictive text will ever help finish.

Published

on

Apple’s iPhone keyboard withholds autocomplete and swipe help for about 100,000 words and phrases, researcher Derik Kauffman said. Users can still type every entry letter by letter. Swipe typing often rewrites a flagged phrase into a different word.

Kauffman, founder of the paper-checking startup Blacklight, published a decoder on Nov. 6, 2026. The English table in iOS 27.0 names 4,513 phrases, from political nicknames to “Santa isn’t real.” Similar tables ship for 27 other languages.

The iPhone Still Lets You Type Every Flagged Word

The list is not a typing ban. Apple said in an email that iPhone users can always enter a specific word or phrase and can always override a prediction. It said the keyboard is set up to avoid suggestions that might be inappropriate or offensive. It did not comment on the list’s size, its mix, or why political lines are on it.

Kauffman put the conservative total at about 100,000 words and phrases with their extra word forms, not counting Chinese characters. Most flagged English entries are vulgar, or tied to drugs, sex, or violence. The rest include slogans, historical flashpoints, and public figures, which is where a safety filter starts to look like editorial policy.

WHAT THE KEYBOARD ACTUALLY DOES

  • Predictions: Predictive text never offers a listed phrase, even when it is the obvious next word.
  • Swipe typing: QuickPath will not produce most listed phrases; it picks another word and can change one you already entered.
  • Letter by letter: Typing one character at a time always works, and nothing on the list stops that path.
  • Exact sequences: The match is the precise word string, so a slightly different wording can still go through.

That last rule is why “Taiwan is country” is withheld and “Taiwan is a country” is not. The filter is brittle on purpose. It is also why a user who pecks keys never meets the wall that a swipe typist hits.

It’s a very light-touch form of censorship. But at the scale that Apple is operating, it only takes a light touch to have a big impact.

Derik Kauffman, founder of Blacklight

Matt Suiche, an engineer at the cybersecurity firm Tolmo, recovered the same table on his own and got a near-perfect match against Apple’s internal data. Suiche said the list had been hidden in plain sight for years, and that human communication inside messaging apps had been filtered for years.

4,513 English Phrases Sit in a Tiny Table

Kauffman’s site publishes 4,513 recovered English phrases decoded from iOS 27.0. The prediction table behind those names has 4,525 slots. A dozen or so IDs have no spelling in any lexicon Apple ships, so those keys never make the public list. QuickPath, Apple’s swipe-typing engine, uses a second table of 2,794 phrases that is a strict subset of the prediction list. English is the only language with that split.

Each phrase stands in for its other word forms, so “sleepy Joe” also covers “sleepier Joe.” Some entries are blocked only in suggestions and can still be swiped; Kauffman marks those “Swipe works.” “Dictator Trump” is one of them, and the same slot also covers “dictator” plus a string of other given names. “Make America great again” can be swiped too. Autocomplete still will not offer the phrase.

On a physical iPhone, swipe-typing “Sleepy Joe” has come out as “Sleepy hope.” “Dictator Trump” can be swiped, but “trump” lands lowercase and autocomplete offers “dictator trumps.” “United Ireland” has come out as “United Iceland.” “Tibetan independence” has become “Tibetan independences.” “White man” has become “white mean,” while “red man” and “blue man” swipe through.

PAIRS THE TABLE DOES NOT TREAT THE SAME

Blocked Not blocked
bad girl bad boy
white man black man
I’m white I’m black
read gay smut read smut

Those mismatches are not a glitch Kauffman found in one build. They are how the published table is laid out. Single words Apple tags as offensive are never suggested. Vulgar words appear as suggestions only after you have typed them three times. Separately, the phrase list withholds “iMessage isn’t working,” “Santa isn’t real,” “Covid was a hoax,” and “NSA surveillance.” Trayvon Martin is the only person’s full name on the English list.

Left and right share one word ID, so “left wing conspiracy” and “right wing conspiracy” are the same key. Until iOS 27.0, conspiracy and propaganda shared an ID too. The keyboard has rewritten “left wing conspiracy” as “left wrong conspiracy,” swapping in a neighbor that keeps the swipe alive without emitting the blocked string.

Hindi Slogans and a Six-Line Hebrew List

Apple keeps matching lists for 27 other languages and scripts, many of them as long as the English one. Treating the story as a fight over American nicknames misses that product. The Hindi table withholds “Rahul is Pappu,” a jibe at Rahul Gandhi, plus the Ram temple slogan “the temple will be built right there,” the 2014 BJP line “good days are coming,” and “love jihad.” French withholds “immigration trop élevée.” Hebrew withholds “the occupied territories,” and the whole Hebrew list has six entries.

THE SAME FILTER, LOCAL RED LINES

Language Example withheld sequence
English sleepy Joe; Tibetan independence; United Ireland
Hindi Rahul is Pappu; Ram temple slogan; love jihad
Hebrew the occupied territories (six entries in the whole list)
French immigration trop élevée
Chinese June Fourth; Taiwan independence; next-word freeze after “tibet”
German Heil Hitler; Deutschland über alles
Italian Crimea è russa
Punjabi Long live Khalistan

Chinese typing works differently. Blocked words still appear as pinyin candidates; once you pick one, the keyboard stops suggesting what comes next. In one part of that engine, Apple’s code also skips suggestions whenever a word is “taiwan” or “tibet,” written in English. The Simplified Chinese table blocks the KMT’s name and does not block the Communist Party’s name; Traditional Chinese blocks both.

Arabic withholds Hamas’s name, “Israeli aggression,” and several “occupation” constructions. Turkish withholds “Armenian genocide” and both “headscarf-wearing” and its opposite. Swedish withholds “Keep Sweden Swedish.” Spanish withholds “inmigrante ilegal” and “Gibraltar español.” The pattern is not a single US culture war. It is a global set of sequences Apple does not want the keyboard to finish.

Why Swipe Typing Turns Ireland Into Iceland

Swipe typing is a guess. The finger traces a path, and the keyboard emits the highest-scoring word that fits. When the intended phrase is on the blocklist, that candidate is dropped and the next-best spelling wins, which is how Ireland becomes Iceland and Joe becomes hope. Tests on an iPhone found the same stutter on “white man,” while nearby color-plus-man strings went through. The user who taps keys never sees that fork. The user who swipes meets it on every flagged n-gram.

Kauffman first noticed the fork in September 2026, when swipe typing stumbled on some phrases and not on their near twins. He used OpenAI’s Codex and Anthropic’s Claude Code to hunt for a stored list, then built a decoder that talks to Apple’s own language assets. On-device checks later matched selected phrases. Independent recovery by Suiche matched the same files.

The American nicknames are what travel in headlines. The durable fact is that the same machinery treats a Hindi campaign slogan, a Chinese next-word freeze after “tibet,” and “Santa isn’t real” as the same class of sequence. Collapsing that into a Trump-versus-Biden row leaves the actual product undescribed.

The Hash Table Never Stores the Words

The blocklist is a hash table of 32-bit fingerprints. It stores no text. Apple ships it as blocklist.dat inside a language-model folder, under 30 KB for English, with a longest phrase length of six words. Before the keyboard offers a candidate, language-model code asks whether that token sequence is blocklisted. A match sets a flag and the candidate is dropped.

Because the file holds fingerprints, not strings, Kauffman recovered the phrases by enumerating candidates and checking them against Apple’s table. About 930 English word IDs make the short-phrase search small enough to finish in about ten minutes on eight CPU cores. A wrong guess collides with probability 2 to the minus 32, so nonsense strings can hash into a real slot; the keyboard cannot tell those coincidences from the intended phrase either. In practice nobody types them.

WHERE THE ENGLISH FILES LIVE

  • Predictions: 4,525 slots in blocklist.bundle for suggestions and completions in iOS 27.0.
  • QuickPath: 2,794 slots in qp_blocklist.bundle, a subset used only for English swipe typing.
  • Updates: English, Chinese, and Cantonese ship in the system image; other languages arrive as LinguisticData downloads from Apple’s asset catalog, currently c17-449.
  • History: Older catalogs and their ZIPs remain online back to 2015, which is how the decoder rebuilt earlier tables.

The built-in iOS 27.0 English tables are newer than the downloadable asset and differ by two keys, including a split of propaganda from conspiracy. Anyone can rebuild the lists from Apple files with the replication package, which fetches the current MobileAsset catalogs and needs no Apple frameworks. The table still does not say why an entry exists. Many look machine-expanded from seed phrases, so an odd pairing may be a template, not a person ticking a box.

Google’s Autocomplete Fights Mapped This Risk

Search engines spent a decade in court over the same distinction Apple now leans on: a suggestion is not the user’s typed words. In May 2013 the German Federal Court of Justice held that Google’s autocomplete predictions can violate personality rights, treating those predictions as the operator’s own content. A Paris court in 2010 ordered Google to stop serving certain name-plus-insult completions. Google’s later public line matches Apple’s email almost beat for beat.

Tamar Yehoshua, then product management director for Search, wrote on June 10, 2016 that autocomplete is designed to avoid completing a person’s name with offensive or disparaging terms. She also wrote that predictions are not search results and do not limit what a person can search for. Apple’s keyboard statement is that same split: we will not author the suggestion; you may still type the line.

HOW THE LIST CAME INTO VIEW

  1. 2015: Oldest LinguisticData catalogs still online date to this year and still serve as witnesses when a later table is decoded.
  2. June 2020: Testers on iOS 13.4.1 find swipe typing accepts “black man” and mangles “white man,” “black woman,” and “white woman.”
  3. September 2026: Kauffman traces a swipe stutter to a fingerprint table inside iOS and begins recovering keys.
  4. Nov. 6, 2026: He publishes the decoder; Suiche reports a near-perfect independent match; Apple emails the generic keyboard statement.

The 2020 swipe quirk sat in public for six years before anyone published the file. That is the second-order part. A list that never announces itself still trains muscle memory: some phrases are easy to finish, some fight the hand.

The Predictive Toggle Does Not Open the List

Apple’s user guide still tells people they can turn predictive text off in Keyboard settings, and that they can reject a suggestion by tapping the original word in quotation marks. Turning the feature off does not publish the table, does not let a user unflag “United Ireland,” and does not explain why “read smut” swipes and “read gay smut” does not. Text replacement shortcuts can force a favorite phrase to expand, which is a personal workaround, not a policy control.

David Greene is senior counsel at the Electronic Frontier Foundation. He said he could see an argument for such a list, then drew the line at secrecy.

This is something that should not be secret.

David Greene, senior counsel, Electronic Frontier Foundation

Apple sells a keyboard that will help finish “On my way” and will not help finish “Tibetan independence.” Users can peck out the second phrase. Hundreds of millions of iPhones still decide, in a file under 30 KB, which sequences get a push and which get Iceland instead of Ireland. The catalogs from 2015 are still on the server. The phone still will not show its own table.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending