Why iOS QuickType won’t predict ‘abortion’

The poetics and politics of predictive text

Editor's Note: This is the sixth installment of the Living With Data series exploring how our online data is tracked, collected and used. Do you have questions about how your personal data is being used? Curious to learn more about your daily encounters with algorithms? Email the Decoder at thedecoder@aljazeera.net or submit your question via the form here. Screen shots and links are helpful clues! 

Predictive text technology makes typing with our thumbs a lot easier.

Predictive text in mobile phones is not new; Android phones have had some version of it with Swype keyboards for a while. And before smartphones, T9 made keypad typing a little less arduous.

With the introduction of iOS 8’s new QuickType feature, iPhone and iPad users are experimenting with the creative constraints of the new form. And the feature is even available in the latest desktop OS, Yosemite (by pressing Esc). The autopilot dream of Silicon Valley April Fools’ jokes is close to reality. But does this predictive, personalized algorithm also have the potential to limit our speech?

QuickType avoids certain sensitive words
Screen shot, The Decoder

How predictive text works

Apple calls QuickType “our smartest keyboard ever.”

In many ways, it is very smart. From the moment you update iOS, QuickType takes into account all your previous conversations and writing style to tailor its predictions for you.

It’s also aware of context, meaning that it will match your writing style when you are texting, and it will adjust to be more formal in email mode, even depending on whom you address in the conversation.

As soon as you open a blank message, QuickType offers three plausible choices, for me, often “I | The | I'm.” In the span of a few minutes, I experimented hitting the middle prediction button every time in an email, text or tweet. I got all these different and often surprisingly appropriate, if cryptic, results:

blank text message: “the fact I can get it right away with the same thing”

blank email: “the fact I have a good time with the same thing to say it is not the same”

email responding to a colleague about a draft: “the fact I can see the new version is the only one that is a good time”

text to my brother: “the only thing that would have to go home”

text to my husband: “the only one that is the only one that has a great way for the next day”

new tweet: “the fact I can get a follow back on my way home from work”

QuickType takes into account only the conversation data that’s on your devices, so the predictive training is happening locally, rather than in some Apple cloud server somewhere. That means Apple hasn’t copied all your email into a profile in order to process how you write.

If the QuickType predictions are frustrating for you, you can temporarily hide them by clicking anywhere on the prediction bar and sliding down. You can also turn off the feature permanently by holding down the keyboard selector and turning off the predictive keyboard. You can also turn it off in settings/general/keyboard, where you have more options.

Creative constraints

After a silly experiment, I was intrigued with the potential creative possibilities. I tried variations on the theme — clicking only middle, only first and only last prediction options. I started from scratch using only predictions. I started with one word and went from there.

I sent these nonsensical Mad Libs to friends, who were rightly confused until I revealed my messages as QuickType poetry. My friend Laura thought it sounded “like a depressed chatbot” and “it’s like if Eliza the chatbot therapist grew up to become an emo poetry slam artist.”

A Twitter search for “predictive text” and “QuickType” suggests I’m not the only one experimenting with the form. Twitter user @alyannapadilla shared her version of the game, writing, “Predictive Text With Friends, brought to you by iOS 8. No app purchase required! Try it out today!” alongside the hilarious screen shot output “I have zero tolerance for your support and love.”

A Gizmodo writer used QuickType to chat with matches on the dating app Tinder. The Verge used it to autocomplete famous Bartleby’s quotes. Web comic xkcd brilliantly played with iOS-mangled movie quotes. And YouTuber Jonathan Mann wrote a pop song completely from predictive text. QuickText even offers some predictive Easter eggs like “hide yo” that leads to “kids | wife,” a reference to the Bed Intruder meme.

Meme predictions
Screen shot, The Decoder

Others have started to share how predictive text knows them well. My friend @gregory on Twitter shared QuickType’s completing “coffee” as “CoffeeIsForClosers,” stating “My iPhone knows me well.”

We see QuickType personalization at work when these strange predictions surface. In response to “hahaha,” I noticed QuickType suggested two smilies, perhaps based on my predilection for emojis.

Emoji response prediction, QuickType FTW
Screen shot, The Decoder

Both the creative and the highly personalized encounters illustrate just how much power these predictions have in shaping our words and our experience. We’re testing the constraints of the form, often to poetic ends. These examples are amusing, in an almost Dada-data absurdist kind of way.

But they can also be unsettling. One QuickType finished “Hey sorry, lost all my” with “friends | life | heart.”

These experiments all play with the constraints of what QuickType suggests, but what missing from QuickType’s vocabulary? And how else might the constraints of this input system constrain our speech?

Limiting speech

How will we know when our conversation partners are autocompleting our conversations?

“Talk later?”

“OK”

“No”

Last year Michael Keller (now on Al Jazeera’s interactive team) investigated a massive corpus of misspelled words to determine whether and how iOS would spell-check them. He uncovered a list of sensitive words — kill words — like “abortion” and “bullet,” for which iOS offered no suggestions when misspelled.

Curious to see how these words fared in the new world of QuickType, I employed some of Keller’s tactics at a small scale. Since spell-check is essentially another flavor of natural language processing or predictive text, I expected Keller’s surfaced flag list to uncover similar biases in the new system.

When you type “abortio,” for example, QuickType suggests completing it as “abortive” or “absorption” instead of the more sensitive “abortion.” “Murde” offers “murde | nurse | nurses” as options instead of the more obvious “murder.” “Bulle” suggests “bulls | bulletproof.” Mistyped “rapw” suggests “tape | taped” instead of “rape.”

QuickType won’t suggest “bigot,” “cuckold,” “deflower,” “homoerotic,” “marijuana,” “pornography,” “prostitute” or “suicide” — all words Keller’s spell-check investigation flagged. The Cupertino effect goes beyond improbable autocorrect to ignoring these kill words completely.

In Keller’s investigation, he started with a clean install of the iOS because spell-check would add frequently used words to the device dictionary. With QuickType, that scale of computational investigation would be hard to do, since so much of QuickType magic lies in its adaption to your personal style. All my examples are already filtered by existing writing patterns with my phone.

Could I train my personal dictionary to learn to complete “abortion?” Or would it never learn and default to the absurd “abortifacients?”

Of course, we can always take the time to type the word ourselves. This is not complete algorithmic censorship just yet. Yet it is worth considering how these new input paradigms potentially limit and constrain our speech, especially when they purport to take on our voice.

Seems to be “the only way I can get the hang of this.”

Beautiful gibberish
Screen shot, The Decoder

What strange predictions is QuickType making about what you might say next? Share in the comments or on Twitter with #TheDecoder and #QuickType hashtags. The Living With Data series explores how our data is tracked, collected and used online. Do you have any questions about how your personal data is being used? Curious to learn more about your daily encounters with algorithms? Email The Decoder at thedecoder@aljazeera.net, tweet me @smwat, or submit your issue or question via the form here. Screen shots and links are helpful clues!

Living with Data

Find Al Jazeera America on your TV

Get email updates from Al Jazeera America

Sign up for our weekly newsletter

Get email updates from Al Jazeera America

Sign up for our weekly newsletter