You’re writing a text message. You write a word with a typo, but thanks to autocorrect, as soon as you press the spacebar it’s fixed. Another word, which contains more typos, is underlined. All you have to do is click. You can turn on the autocomplete function, and then after a few letters, the software adds the rest. Gmail has an even more advanced option.
I write: ‘Hi’ – and the machine adds a comma.
I write: ‘We haven’ and the machine adds, in grey: ‘’t seen each other in a while, I hope you’re doing well’.
The algorithm indicated that that phrase fits best with the sequence of letters at the beginning of the sentence in a message addressed to a person I don’t keep in touch with regularly. Google’s engineers decided their program knows all of this, because the algorithm learnt it from the content of emails written by 1.4 billion users. But do I really want to say, ‘I hope you’re doing well’? A space is enough to make the automatically generated pleasantry disappear. Or I can hit the tab key to get to the end of the sentence, which becomes mine.
The newest invention is the GPT-3 model. All you have to do is give it a few words or sentences, and it writes the rest of the text. Its makers have allowed it to teach itself on petabytes of data, of which more than 6 million English-language Wikipedia articles account for barely 0.6% of the material provided to it. The people who tested GPT-3 say the text it generates is as good as it would be if it were written in a particular style by a particular person. It doesn’t matter whether it’s a 16-syllable monster, a software code or an advertising slogan.
In the 1940s, Lin Yutang built a typewriter in which suggestions for the following characters were found next to each key. Associative clusters, the precursor of today’s predictive text, were invented less than a decade later by the Chinese typesetter Zhang Jiying. And if things had stopped with these conveniences in writing Chinese, millions of people around the world wouldn’t feel embarrassed at the mere recollection of the mortifying mishaps they owe to autocorrect. Or only Chinese people would feel this embarrassment.
In the early 1990s, immediately after finishing his studies at Harvard, Dean Hachamovitch was hired at Microsoft. Work was already underway on integrating a dictionary with the text editor in Word, but Hachamovitch had the idea that after pressing the space bar, if the preceding word contained an error, the program could correct it by itself. The young intern was dragooned into the tedious work of creating a dictionary of the most common errors.
Thirty years later, this same work, though on an incomparably greater scale, happens in the cloud. The algorithms independently sort through petabytes of data, weighing up the phonetic similarities or proximity of letters on the keyboard but also the linguistic contexts and frequency of words, changing autocorrect into an index that’s less like a dictionary and more like a popularity contest.
You want to kill yourself? Go ahead!
