“Australia is a very bad place to question your sexuality on MDMA with NSA spies” might be a text you would wish to send to your friends, if you have had an extraordinarily complicated night out in Sydney.
But your iPhone won’t help you spell any of it if you try to swipe-type it – that is, by gliding a finger across a keyboard as opposed to pecking out a word.
New research shows that Apple has a “grey list” of more than 100,000 words and phrases that it has restricted from being completed by its predictive text.
So if a user wishes to type phrases such as “Dictator Trump”, “NSA spies” or, surprisingly, “Australia is a very bad place”, “question your sexuality” or “on MDMA”, they will have to spell them out.
Apple’s predictive text will not suggest these words. And, if a user attempts to swipe-type them, they will often stumble, changing “United Ireland” to “United Iceland”, for example, or “Tibetan independence” to “Tibetan independences”.
The “grey list” was discovered by the independent researcher Derik Kauffman, whose other projects include using AI to find errors in scientific papers. After noticing that he had problems typing certain phrases – but not others – on his iPhone, Kauffman was able to unearth a table containing the list, encoded in Apple’s operating system.
The list he found includes some unsurprising words and phrases. Apple’s iPhone will not autocomplete references to drugs, sexual acts and, perhaps in the spirit of corporate protection, “iMessage isn’t working”.
But it also includes possibly politically charged phrases, which Kauffman described to Reuters as a “light-touch form of censorship”. Among these are “Trayvon Martin”, the name of the Black teenager shot in Florida in 2012, “Kashmir dispute”, “Muslim extremist” and “male chauvinist pig”.
In Hebrew, Apple will not autocomplete the phrase “the occupied territories”. In Arabic, it will not autocomplete the phrase “the Hamas movement”.
The list is one more glimpse into the dark bowels of how autocorrect – the chirpy keyboard partner that helps you look more educated to your loved ones – actually works.
The technology has taken many forms. In its earliest version, Unix, an operating system, would look up each word a user typed against a table of words, to determine if this word was spelled correctly.
This evolved into n-grams, a tool which allows an operating system to statistically predict what words might come next, based on the words before it. These form the basis of Apple’s familiar autocorrect, which was known to change expletives into prudish alternatives: “fork” and “ducking”, among other gaffes.
Last year, Apple launched a new autocorrect – a “transformer-based” system that was more akin to an LLM. Reddit sleuths saw this as a possible reason behind a spate of avant-garde spelling suggestions their iPhones made last year.
This hard-coded list illuminates one longstanding mystery behind the world’s spelling coach. Others remain, such as a brief glitch in which users of the dictation feature saw the name “Trump” flash across their screen when they spoke the word “racist” into their iPhones – which Apple quickly remedied.
Apple said that it designed its keyboard to avoid offering inappropriate or offensive predictive suggestions. It said iPhone users could always type out specific words or phrases, and override predictive text suggestions to achieve their desired meaning.
