The iPhone keyboard does not prevent users from typing certain words or phrases, but for a long list of expressions, it chooses not to offer any help. They do not appear among predictive text suggestions and can produce unexpected results in swipe typing. This emerges from an archive identified by researcher Derik Kauffman, who published online a list of over 100,000 entries attributed to iOS language systems.

The discovery concerns words that the system prefers not to autocomplete, rather than censorship of the text field. iPhone users can still enter that content letter by letter and, according to what Apple told Reuters, can also ignore or override the keyboard's suggestions. The distinction is important: the software intervenes at the stage where it suggests, corrects, or anticipates typing—a step that influences the speed and form of communication without physically blocking the message.

In the material analyzed by Kauffman, insults, sexual terms, drug references, and offensive language predominate. However, the list also contains political phrasing and statements on matters of public interest. Among the examples that emerged are “dictator Trump”, “sleepy Joe”, “illegal immigrant”, “covid was a hoax”, and “left wing conspiracy”. Expressions that are not immediately linked to vulgarity or discrimination also appear, such as “Santa isn’t real”, “iMessage isn’t working”, and other commonly used phrases.

A filter on suggestions, not on text

The software keyboard is one of the most used and least visible tools on a smartphone. When a user begins a sentence, iOS suggests likely words; when they glide a finger across the keys, it attempts to reconstruct the intended term; when it detects a potential error, it can offer a correction. Leaving an entry out of these mechanisms does not make it impossible to type, but it strips away a shortcut that the system otherwise grants to most of the vocabulary.

In the case reported by Kauffman, the effect may be more noticeable with swipe typing. The Verge cites the example of “left wing conspiracy”, transformed by the keyboard into “left wrong conspiracy”. It is not clear from the available data how often each individual entry triggers an automatic replacement, or whether the behavior varies across languages and operating system versions. However, the fact remains that excluded entries do not receive the standard treatment reserved for terms suggested by the predictive dictionary.

Apple told Reuters that the keyboard is designed to avoid potentially inappropriate or offensive suggestions. This rationale aligns with a substantial portion of the list, but on its own does not explain the presence of political expressions, pandemic-related messages, or phrases such as “iMessage isn’t working”. Based on reported information, the company has not provided public, entry-by-entry criteria, nor details on the process that moves a phrase into or out of the repository.

How the list was found

Kauffman claims to have tracked down the list within iOS internals with the help of OpenAI Codex and Anthropic's Claude Code. On his website, he describes the file as a hash table—a structure used to quickly associate and retrieve data—which he reportedly managed to convert back into text format. He then made the collection and the separate lists for different languages available.

Reuters reports that another researcher was able to independently obtain the same list. This reinforces the idea that this is not an isolated list or one artificially created during analysis, but it does not amount to a full explanation of its origin, the date individual entries were added, or how Apple distributes it across iOS.

According to Kauffman, the exclusions cover major languages. Such an extensive repository is likely the result of decisions layered over time: moderating offensive words, preventing embarrassing suggestions, managing swipe typing ambiguities, and adapting to different languages. However, the available information does not allow a specific rationale to be assigned to every entry. It would be inaccurate to interpret the entire list as an exclusively political catalog, just as it would be to dismiss the political phrases present as mere variants of offensive terms.

Inconsistencies make the case more delicate

A key point of the discovery is the absence of any obvious symmetry. Kauffman pointed out, for instance, that “bad girl” is excluded while “bad boy” is not. Discrepancies of this kind can stem from safety assessments, rules added for specific cases, outdated linguistic data, or technical criteria tied to completion probabilities. Without documentation from Apple, it is impossible to determine which explanation applies to that case.

It is precisely these inconsistencies that change the nature of the debate. For a feature that automatically suggests words as the user types, it is not enough to ask whether it is reasonable to avoid epithets or slurs. One must also consider whether the rules are coherent, understandable, and proportionate. A keyboard is not a search engine and is under no obligation to suggest every possible phrase; at the same time, when it filters expressions tied to politics, public health, or the operation of a product from the company itself, the line between interface moderation and language steering becomes harder to discern from the outside.

The topic is relevant even to those who have never used swipe typing. Predictive suggestions are a form of daily intermediation: they reduce the number of taps, speed up replies and messages, and in some cases shape the final wording of a sentence. Their influence is small on an individual word, but can become significant when embedded in a tool used continuously by hundreds of millions of people.

Transparency and user controls

Apple has historically built part of its value proposition on privacy protection and control over the user experience. Based on the data that has emerged, the affair surrounding the list does not call into question the ability to freely type those terms: users can still do so. However, it raises a distinct issue: transparency regarding the filters applied to language assistance features.

Based on the disclosed information, there is no iOS settings panel that exposes this collection or allows users to choose in a granular way which suggestion categories to disable or re-enable. Users can adjust keyboard options and automatic corrections, but that is different from knowing which words have been excluded upstream by the system. Developers, researchers, and policy makers also have few tools to determine whether this behavior is consistent across markets, languages, and software updates.

The next step will depend primarily on Apple. The company could merely reiterate that suggestions must avoid offensive content, or clarify the adopted categories, the handling of political phrases, and the procedures used to review the lists. It could also explain whether all the identified entries are active in current versions of iOS and what differences exist between predictive text, autocorrect, and swipe typing. Until these details are made public, the archive discovered by Kauffman will remain a partial yet significant snapshot of how a keyboard can shape writing without banning it.

Sources