• Apple's iPhone Keyboard Has a Hidden List of 100,000 Restricted Phrases, Including Political and Everyday Expressions[

    From mummycullen@mummycullen@gmail-dot-com.no-spam.invalid (MummyChunk) to misc.phone.mobile.iphone on Mon Nov 9 14:40:08 2026
    From Newsgroup: misc.phone.mobile.iphone

    Apple's iPhone Keyboard Has a Hidden List of 100,000 Restricted Phrases, Including Political and Everyday Expressions

    An investigation into Apple's iPhone keyboard has uncovered a previously undisclosed filtering system that prevents approximately 100,000 words and phrases from appearing in predictive text suggestions and, in many cases, makes them difficult or impossible to enter using swipe gestures. The restrictions extend across numerous languages and cover everything from offensive expressions and references to illegal drugs to political slogans, historical events and seemingly harmless everyday statements.

    The discovery was made by independent researcher Derik Kauffman, who reverse-engineered the linguistic databases used by Apple's keyboard after noticing that certain expressions behaved strangely during swipe typing. His findings, published through a research project called Swipeblocked, reveal that Apple maintains extensive lists of phrases that its keyboard is programmed not to suggest, even when those phrases would otherwise be the most natural continuation of a sentence.

    Among the restricted English-language expressions are "dictator Trump," "sleepy Joe," "make America great again," "Tibetan independence," "United Ireland" and "Santa isn't real." Other entries include references to historical atrocities, religious disputes, sexuality, immigration and political controversies. The lists are not limited to English, with similar filtering mechanisms identified in 27 additional languages and writing systems.

    Reuters independently tested several examples and confirmed that the iPhone keyboard either refused to suggest the expected wording or produced unexpected alternatives when users attempted to enter the phrases through Apple's QuickPath gesture-typing system. Cybersecurity engineer Matt Suiche subsequently recovered a nearly identical list, providing independent confirmation of Kauffman's technical findings.

    The discovery has prompted questions about transparency and the extent to which technology companies should influence the words their software suggests. However, the restrictions do not prevent users from expressing themselves. Every identified phrase can still be entered manually by tapping individual letters, and Apple says its objective is to avoid presenting potentially inappropriate or offensive suggestions rather than prohibit particular forms of speech.

    How a Strange Swipe-Typing Problem Led to the Discovery

    Kauffman's investigation began in September 2026 when he noticed that Apple's keyboard behaved inconsistently while he was entering certain combinations of words.

    QuickPath, Apple's gesture-based typing system, allows users to compose text by sliding a finger across the keyboard rather than tapping each letter separately. The software analyzes the movement, considers possible word combinations and uses linguistic predictions to determine what the user intended to write.

    Because multiple words can correspond to similar finger movements, the keyboard relies heavily on language models and the surrounding text to identify the most likely result.

    Ordinarily, this produces a relatively smooth experience. However, Kauffman noticed that certain expressions repeatedly generated incorrect substitutions, even when the intended letters appeared straightforward.

    Rather than treating these incidents as ordinary prediction errors, he began investigating whether the keyboard was deliberately avoiding particular combinations.

    Using OpenAI's Codex and Anthropic's Claude Code to assist with code analysis, Kauffman examined components of Apple's linguistic processing system and eventually located a compact collection of data used to identify restricted phrases.

    The discovery revealed that the keyboard does not simply make independent predictions based on word frequency and grammatical context. Before displaying certain suggestions, it checks whether the proposed sequence appears in an internal restriction table.

    If a match is found, the candidate can be removed from the suggestions presented to the user.

    The mechanism also influences QuickPath, although not every phrase restricted from predictive suggestions is excluded from swipe typing.

    Kauffman developed tools to reconstruct the contents of Apple's tables and published the recovered phrases, technical explanations and supporting code. His investigation covers multiple iOS releases, allowing changes to the lists to be examined over several years.

    Reuters subsequently tested a selection of the phrases on an iPhone and observed the same general behavior. Suiche, an engineer at cybersecurity company Tolmo, independently reconstructed the data and reported a near-complete match with Kauffman's findings.

    These additional checks strengthen the evidence that the filtering mechanism is a deliberate component of Apple's keyboard architecture rather than a collection of unrelated typing errors.

    Predictive Text and Swipe Typing Are Restricted in Different Ways

    One important distinction concerns the difference between suppressing suggestions and preventing a particular phrase from being entered through gestures.

    Apple's keyboard provides several mechanisms for assisting with text entry. Predictive text suggests likely words or phrases based on the current context, while inline predictions can offer completions as a person types. QuickPath attempts to recognize words from movements across the touchscreen.

    The discovered restriction lists interact with these features differently.

    According to Kauffman's analysis of iOS 27, the English predictive-text database contains approximately 4,500 restricted phrases. A separate QuickPath table contains roughly 2,800 entries, representing a subset of the larger list.

    Consequently, some phrases never appear as predictive suggestions but can still be entered successfully using swipe gestures. Others are restricted through both mechanisms.

    The distinction helps explain why users may encounter apparently inconsistent behavior.

    For example, the expression "dictator Trump" is excluded from predictive suggestions, but Reuters found that it could still be entered using QuickPath, although the keyboard did not preserve the expected capitalization. The system instead favored a different grammatical interpretation when offering completions.

    Other phrases produced more obvious substitutions.

    When Reuters attempted to swipe-type "United Ireland," the keyboard frequently returned "United Iceland." The expression "Tibetan independence" could become "Tibetan independences," while "sleepy Joe" was transformed into "Sleepy hope."

    These are not examples of words that are inherently impossible for a language model to recognize. The investigation demonstrates that the relevant combinations appear in Apple's restriction data.

    The filtering can also affect previously entered text. Kauffman's demonstration shows that when a user attempts to complete a phrase involving a political conspiracy, the keyboard may revise an earlier word rather than produce the expected continuation.

    This behavior is more intrusive than simply declining to display a suggestion. It can force users to correct text that the keyboard has modified while attempting to avoid a particular sequence.

    Nevertheless, the mechanism is not a general prohibition on entering the words.

    Users can type the intended phrase letter by letter, reject an unwanted correction or use an alternative input method. The restriction applies to particular keyboard assistance functions rather than the underlying ability to create or transmit the text.

    That difference is essential when assessing claims that Apple has banned thousands of expressions from iPhones. The software may make certain phrases less convenient to enter, but it does not prevent them from being written.

    Political Expressions Appear Alongside Profanity and Offensive Language

    Most of the restricted material falls into categories that might reasonably be associated with keyboard moderation.

    The recovered lists contain profanity, derogatory expressions, sexual references, drug-related terminology and phrases associated with violence or extremism.

    A manufacturer may have practical reasons to avoid automatically suggesting such expressions. Predictive keyboards operate in personal conversations, workplaces, classrooms and other environments where an inappropriate suggestion can be embarrassing or offensive.

    However, the lists extend well beyond conventional profanity.

    Political expressions identified by Kauffman include references to prominent American politicians, immigration debates and controversial historical or geopolitical subjects.

    Among the examples are "dictator Trump," "sleepy Joe," "illegal immigrant," "make America great again," "Tibetan independence" and "United Ireland."

    Other entries address disputed political claims and conspiracy theories, including statements about government surveillance and the COVID-19 pandemic.

    The presence of these phrases raises a different set of questions because political language is not necessarily offensive simply because it concerns a contentious subject.

    A journalist, researcher, student or ordinary citizen may have a legitimate reason to write about Tibetan independence, political immigration terminology or campaign slogans without having the keyboard discourage those particular combinations.

    The lists also contain expressions that appear relatively harmless outside a specific context.

    Examples include "Santa isn't real" and "iMessage isn't working," neither of which is inherently abusive or threatening.

    Some restrictions are unusually specific. Kauffman's research identifies "Taiwan is country" as a restricted sequence, while the grammatically correct expression "Taiwan is a country" is not blocked in the same way.

    Similar inconsistencies occur among other phrases. The researcher found cases where one expression is restricted but a closely related alternative remains available.

    These differences make it difficult to establish a consistent editorial standard from the published data.

    They also illustrate why the list should not automatically be interpreted as a catalog of political positions endorsed or rejected by Apple.

    Some entries appear to derive from broader templates, with different words inserted into related grammatical structures. Others may reflect historical attempts to prevent inappropriate suggestions or avoid embarrassing prediction errors.

    Without access to Apple's internal selection criteria, it is not possible to establish why each individual phrase was included.

    Nevertheless, the consequences are observable. Certain political and historical expressions receive different treatment from otherwise comparable language, even when users have not asked the keyboard to filter sensitive subjects.

    The Restrictions Extend Across Nearly 30 Languages

    The multilingual nature of the system makes the discovery considerably more significant than a peculiarity of the English keyboard.

    Kauffman identified related restriction data covering 27 additional languages and writing systems, including Chinese, Arabic, Hindi, Hebrew, French, German, Spanish and Russian.

    The contents vary according to language and region, suggesting that the lists incorporate expressions associated with different political, cultural and historical contexts.

    Chinese-language entries include references to the June 1989 crackdown around Tiananmen Square, Taiwanese independence and a nickname associated with Chinese President Xi Jinping.

    The recovered data also includes references to political organizations and territorial disputes.

    In Hindi, restricted expressions include political slogans and derogatory phrases involving major Indian political figures. French-language entries include immigration-related statements, while Hebrew contains an expression referring to occupied territories.

    German-language material includes Nazi slogans and extremist terminology, reflecting subjects with particular historical and legal sensitivity in German-speaking countries.

    The Russian-language data contains profanity and politically charged combinations. Subsequent examination by Meduza identified references to military activity, Russia's 2008 war with Georgia and a phrase involving Vladimir Putin.

    However, the Russian list has only been partially reconstructed.

    Russian words frequently change their endings according to grammatical case, gender and number. The keyboard's linguistic processing system incorporates these variations through a combination of word stems and suffixes, making the underlying restriction data more difficult to decode.

    Kauffman's recovery tools do not yet reproduce every aspect of that mechanism, so the published Russian entries should not be treated as a complete inventory.

    Chinese input methods introduce another distinction.

    For keyboards using Pinyin, users generally type phonetic representations of words and then select the intended Chinese characters from suggested candidates.

    Kauffman's technical analysis indicates that restricted Chinese expressions can still appear as candidates during character selection. The filtering primarily affects the predictions that follow, rather than universally preventing the original characters from being entered.

    This differs from the behavior of the English QuickPath system and demonstrates why the restrictions cannot be described as operating identically across all languages.

    It also complicates the headline figure of approximately 100,000 expressions.

    That number is a conservative estimate derived from the recoverable multilingual data and its possible word variations, rather than a simple count of 100,000 plain-text entries stored in one location.

    Some languages have not been fully decoded, and Chinese processing uses different data structures, so the precise total depends on how individual phrases and their variations are counted.

    Apple Stores the Restrictions as Digital Fingerprints Rather Than Readable Text

    One of the technically interesting aspects of the discovery is the way Apple stores its restricted phrases.

    Rather than maintaining an ordinary text file containing each expression, the keyboard uses compact hash tables.

    A hash function transforms input data into a numerical representation that can be compared efficiently with stored values. In this case, the system converts sequences of internal word identifiers into numerical fingerprints.

    When the keyboard considers a possible prediction, it can check the corresponding fingerprint against the restriction table without needing to search through a large collection of readable phrases.

    Apple's implementation uses a minimal perfect hash structure combined with 32-bit fingerprints.

    A minimal perfect hash function is designed so that every entry in a predetermined set maps to a unique position in a compact table. This makes lookups efficient and reduces the memory required to represent the collection.

    According to Kauffman's technical documentation, the English restriction table occupies less than 30KB despite containing thousands of entries.

    The system supports phrases consisting of several words, with the underlying lookup mechanism examining sequences ranging from two to six tokens.

    Its compact design is particularly useful in a mobile operating system where keyboard processing needs to respond almost immediately to user input.

    The use of numerical fingerprints also explains why the list remained difficult to inspect directly.

    Someone examining the relevant iOS files would not encounter a straightforward document containing political slogans and offensive expressions. Instead, the data appears as numerical tables associated with the keyboard's language model.

    Kauffman reconstructed the entries by generating candidate word sequences, calculating their corresponding fingerprints and testing those values against Apple's stored tables.

    The process involved considerable computation and was assisted by automated tools. His published research includes the software needed to reproduce the recovery procedure from Apple's language data.

    An important limitation is that a successful fingerprint match does not always reveal the precise original wording used when Apple generated the table.

    Multiple grammatical forms can share internal identifiers, and some recovered entries may represent groups of closely related phrases rather than a single literal expression.

    Kauffman also acknowledges that the recovered data does not explain Apple's reasons for choosing an entry.

    The research establishes which combinations are recognized by the restriction mechanism, but it cannot determine whether a particular entry was added manually, generated from a template or inherited from an earlier linguistic database.

    The use of hashing should not be confused with encryption designed to protect user conversations. These tables are part of the keyboard's linguistic data, and their existence does not establish that Apple is monitoring private messages or transmitting restricted phrases to a remote server.

    The discovered mechanism concerns the suggestions generated by the keyboard itself.

    The Lists Have Been Growing for Years and Change Between iOS Releases

    The restrictions are not a new feature introduced with iOS 27.

    By examining historical language data, Kauffman traced the English list back through several generations of Apple's operating system.

    His analysis indicates that iOS 14, released in 2020, contained approximately 1,700 restricted English phrases. By iOS 27, that figure had grown to around 4,500.

    The growth demonstrates that Apple has continued maintaining and expanding the mechanism over several years, even though it has not publicly documented the complete list or explained its selection process.

    Some changes can be associated with specific operating system releases.

    According to the research, "dictator Trump" was added in iOS 16.1, while "sleepy Joe" appeared in iOS 16.2. The political slogan "make America great again" was added through a language-data update in June 2022.

    Other expressions have subsequently been removed.

    Kauffman found that "President Trump" and "President Bush," which had been restricted since a 2022 update, were removed from the list in iOS 26.3.

    This historical record is significant because it demonstrates that the system is actively maintained rather than consisting of a fixed set of entries introduced many years ago.

    However, the timing of additions and removals does not establish why Apple made those decisions.

    The company may update its language models to reflect changes in common usage, improve prediction quality, address complaints or modify policies concerning inappropriate suggestions.

    Apple also distributes linguistic data independently of major operating system releases.

    This means an iPhone may receive updated keyboard information through a background download without installing a completely new version of iOS.

    Consequently, two devices running the same general version of the operating system may not necessarily have identical language data at every moment.

    The findings published through Swipeblocked are primarily based on iOS 27.0 and language assets available during October 2026.

    They should therefore be understood as a detailed examination of the system at that point in time, rather than a guarantee that every entry remains unchanged in later updates.

    Apple Defends the Filtering as Protection Against Inappropriate Suggestions

    Apple responded to Reuters by explaining that its keyboard is designed to avoid suggesting words or phrases that could be considered inappropriate or offensive.

    The company emphasized that users remain able to enter whatever text they choose and can reject suggestions that do not match their intentions.

    Apple did not directly address the composition of Kauffman's recovered lists, explain the inclusion of specific political expressions or disclose how the selection criteria are determined.

    Its explanation reflects a longstanding challenge faced by developers of predictive technologies.

    A keyboard is expected to anticipate what users intend to write, but an overly aggressive prediction system can produce embarrassing, offensive or misleading suggestions.

    The problem is particularly sensitive because the keyboard may offer language that the user never intended to express.

    A prediction that appears harmless in one context may become offensive in another, especially when combined with names, political identities or references to sensitive historical events.

    Companies therefore have reasons to impose limitations on what their software proposes automatically.

    However, filtering also creates questions about user autonomy and transparency.

    The Electronic Frontier Foundation, a digital civil liberties organization, has argued that users should be aware of such mechanisms and have meaningful control over them.

    David Greene, the organization's senior counsel, told Reuters that restrictions intended to prevent unwanted suggestions may have legitimate uses, but questioned the lack of transparency and customization.

    The distinction between automatic suggestions and manually entered text is central to the debate.

    Apple is not preventing users from discussing politics, historical events or controversial subjects. Nevertheless, predictive input systems influence the convenience of entering particular expressions, and small differences can become noticeable when the same software is used by hundreds of millions of people.

    Whether those differences amount to censorship is partly a question of definition.

    Some observers consider the restrictions a limited form of speech filtering because the keyboard deliberately avoids producing particular expressions. Others emphasize that refusing to suggest a phrase is fundamentally different from prohibiting users from writing or transmitting it.

    Both interpretations concern the same technical behavior, but they place different weight on the responsibilities of a software provider.

    The evidence does not establish a systematic preference for any particular political party or ideology. Restrictions appear across different political positions, historical controversies and cultural subjects, while the reasons for individual entries remain unknown.

    The stronger conclusion is that Apple operates an extensive, largely undocumented content-selection mechanism whose effects extend beyond ordinary profanity filtering.

    What iPhone Users Can Do About Unwanted Keyboard Restrictions

    For most iPhone owners, the practical consequences are likely to be occasional frustration rather than a major interruption to everyday communication.

    The restricted phrases generally remain available through conventional typing. If QuickPath repeatedly produces an incorrect expression, tapping the letters individually allows the intended text to be entered.

    Users can also reject unwanted autocorrections and continue typing when the suggested continuation is inappropriate.

    Apple provides several keyboard settings that may help people who find predictive behavior distracting.

    Under Settings, General and Keyboard, users can disable Predictive Text or adjust related correction features. Turning off predictive suggestions prevents the keyboard from presenting those suggestions, although it does not eliminate the underlying restriction database or make every phrase available through QuickPath.

    Text Replacement provides another practical option.

    Users can create shortcuts for expressions they frequently enter by opening Settings, General, Keyboard and Text Replacement, then associating a short sequence of characters with a longer phrase.

    For example, someone who regularly uses a specialized technical term or particular expression can configure a shortcut that inserts the complete wording.

    Apple also allows users to add words to their personal dictionary to reduce unwanted corrections.

    However, these customization features should not be mistaken for a general method of disabling Apple's hidden restriction tables. The research has not established that personal dictionary entries override every blocked prediction or restore QuickPath support for all restricted phrases.

    Third-party keyboards offer an alternative for users who prefer different prediction behavior.

    Applications such as Google's Gboard and Microsoft's SwiftKey use their own typing systems, dictionaries and prediction mechanisms, although their restrictions and privacy practices may differ from Apple's.

    Choosing a third-party keyboard therefore involves considering more than whether it recognizes particular phrases. Users should review how the keyboard handles typed information, what permissions it requests and whether it requires access to external services.

    Apple has not announced a setting that allows users to inspect, edit or disable the specific restriction lists identified by Kauffman.

    The absence of that control is one of the central issues raised by the investigation.

    A Hidden Design Choice With Wider Implications for Everyday Communication

    The Swipeblocked investigation reveals how much editorial judgment can be embedded in software that most people regard as a neutral input tool.

    Smartphone keyboards are no longer simple arrangements of letters. They continuously interpret context, predict intended words, correct mistakes and influence how quickly particular expressions can be entered.

    Those functions can substantially improve convenience, but they also require developers to decide which suggestions should be encouraged, discouraged or excluded.

    Apple's filtering system demonstrates that these decisions can be implemented at a surprisingly detailed level, identifying particular combinations of words rather than relying exclusively on general categories such as profanity.

    The discovery is especially notable because the mechanism has existed for years without a comprehensive public explanation.

    Kauffman's work makes it possible to examine the contents, history and technical implementation of the lists, while independent confirmation from Reuters and another researcher establishes that the findings are not simply theoretical.

    At the same time, the evidence places clear limits on what can reasonably be concluded.

    The investigation does not show that Apple prevents users from writing the restricted phrases, monitors private conversations through the lists or deliberately promotes a particular political viewpoint. It does show that Apple's keyboard treats thousands of expressions differently from ordinary language, often without providing any indication that filtering has occurred.

    For users, the most important question may not be whether Apple should prevent offensive suggestions altogether. Many people would likely welcome protections against accidental profanity or inappropriate predictions.

    The more difficult issue is whether those protections should remain invisible and fixed by the manufacturer, particularly when they extend to political terminology, historical subjects and ordinary expressions.

    A more transparent system could allow users to understand why certain phrases are excluded and choose whether to apply stricter or more permissive prediction settings.

    Such an approach would preserve the benefits of responsible keyboard suggestions while giving individuals greater control over the language assistance they receive.

    Apple has not indicated whether it intends to modify the current system in response to the research.

    For now, the discovery offers a revealing look at the decisions hidden inside an everyday piece of software. The iPhone keyboard does not determine what its owner is allowed to say, but it does influence which words are easiest to write. At the scale of Apple's global user base, that distinction deserves greater transparency than the company has so far provided.
    --
    Via JLA Forums web gateway for misc.phone.mobile.iphone
    --- Synchronet 3.22a-Linux NewsLink 1.2