Start typing a reply in Google’s Gmail and it might feel like it is reading your mind. Tap out the letters “tha” and up pops the rest of a sentence in greyed out letters. Press tab and your sentence completes itself.
In simpler times, you would need to go to the hassle of using valuable seconds to type out “thank you very much” or “that sounds great” all by yourself. But now, email systems like Gmail can finish sentences for you. This feature is powered by a form of artificial intelligence known as natural language processing, which aims to understand and use language in ways that are more human-like than computers have managed before.
But while it can help to reduce the amount of time spent composing an email, many people have mixed feelings about the autocomplete feature, known as Smart Compose, with some describing its ability to scan our messages and come up with a suitable response as downright “creepy”.
With machines starting to take over some of the art of composing messages to our colleagues, friends and loved ones, does this also risk robbing us of something more important? Are they sucking away our individuality and the joy of human interaction? And could these autocomplete technologies even be changing the way our brains work?
“Prediction is fundamental to our perception and our relation with the world,” says Maria Geffen, who studies neuronal circuits for auditory perception and learning at the University of Pennsylvania. “Our brain is constantly making predictions. For example, when we're listening to someone talk against a loud background noise, we're making predictions for what the phrase was that they uttered even if we only heard a subset of sounds. This is also true for more complex cognitive tasks, including producing a sentence.”
Writing – both by hand and on a computer – involves a complex coordination of cognitive processes, including the use of long-term memory, the semantic system, working memory and planning.
But if we subcontract the work of composing our sentences to a machine, Geffen argues that it could have some profound implications for the way our brains work.
“We are doing experiments now in which we are tracking the activity of neurons that represent sounds in the brain over days, and finding that the same ensembles of neurons exhibit varying patterns of activity from day to day,” she says. “So, it is very interesting to think about what happens, when, on one hand, prediction is performed for us not by our brain but by a computer algorithm, and how this repeated experience affects our interaction with the world.”
An estimated 281 billion emails are sent every day by the world’s 3.8 billion email users. It means that each of us are on average receiving around 74 emails a day. Back in the days before the internet, we would be lucky to receive more than two or three letters a day, and most of them would be bills.
So, it is hardly surprising we might need a little help to lighten the load when responding to this barrage of messages. Technology that analyses our writing habits to predict the words we are most likely to use next can reduce the need to type out dozens of individual letters to just a single tap.
Smart Compose is just one of a number of technologies that use AI to predict what we might be trying to say. Predictive text algorithms, which use what we have typed in the past to suggest the next words in a sentence as we write, already feature on most smartphones. Applications that can finish our sentences on desktop word processing programmes are also now beginning to appear.
智能作文造句只是众多人工智能其中一种，正式称之为文本预测算法(Predictive text algorithm)，已经被大多数智能手机所采用。这个技术经计算我们过去在电邮输入的文字内容，因而能在我们写作一个句子时弹出预计我们会使用的下一个单词。这个文本预测算法应用程序也开始出现在桌面文字处理上。
These technologies follow on the heels of other attempts to take some of the heavy lifting out of the typing we have to do.
Before it launched Smart Compose, Gmail introduced Smart Reply, which uses a similar approach to offer users three short potential responses as “shortcuts” that can be selected with the click of a button. Search engines now commonly use autocomplete to automatically suggest questions we might be asking. Our smartphones and web browsers also feature autofill, which will complete online forms for us.
在推出智能作文造句（Smart Compose）之前，Gmail先引入了智能回复程式（Smart Reply），使用类似的方法向用户提供三个简短的潜在回复供快捷选择，用户只需点击一个按钮就可以作答。搜索引擎现在也普遍使用自动完成输入技术，自动提出我们可能会问到的问题。我们的智能手机和网络浏览器也有自动填表功能，可以帮我们快速填写在线表格。
There are certainly benefits. Autocomplete, for example, can reduce the cognitive load and the time it takes to fill in online forms. It can even help you get lower air fares, according to one study.
But even relatively simple features such as autocorrect – first introduced on Microsoft Word in the 1990s and later a ubiquitous texting aid on mobile phones – led to concerns about the impact it was having on children’s writing skills.
Although it may be too early for there to be a great deal of research on the effects of more advanced technologies that can automatically complete our sentences for us, there are some indications that they could alter the way we use language.
One year-long study found that secondary school children who used predictive text on their mobile phones made more spelling errors than non-users, but university students who used the technological writing aid made fewer grammatical errors.
There is even some evidence that predictive text technologies may have a positive impact for those who use them, says Clare Wood, a Nottingham Trent University psychologist who led the study.
负责这项研究的英国诺丁汉特伦特大学(Nottingham Trent University)心理学家克丽尔‧伍德(Clare Wood)说，甚至有一些证据表明，智能预测文本技术可能对使用者有正面的影响。
“We know that for adults, exposure to misspellings can sometimes interfere with their memory for correct spellings,” she says. “So autosuggest functions may be beneficial in minimising the possible negative impact of seeing misspelled words in texts and other online communications in adult users.
“The autosuggest function also has the potential to positively impact the grammatical construction of online communications.”
But Wood warns that as these AI-powered systems rely on learning from what we have typed in the past, it could also introduce errors.
“If it detects that particular ungrammatical word combinations frequently co-occur then these will be reinforced,” she adds.
But autosuggest could also affect what we want to say as well as how we say it. Computer scientists at Harvard University and the Draper Laboratory in Cambridge, Massachusetts, have found that using AI-powered predictive text systems could introduce biases into what people write.
This is because machine learning algorithms, which are used to train AI systems on large sets of data, can pick up and amplify biases contained within the data. So, a predictive text system that has been trained using text from positive online reviews might tend to suggest words that are more positive as a result.
When people writing about a restaurant were presented with text shortcuts that were skewed to be more positive, the resulting reviews tended to be more positive than if they were presented with negative skewed shortcut suggestions, the researchers found.
"Predictive text systems are starting to offer suggestions that are longer, more coherent, and more contextual than ever before,” says Ken Arnold, a researcher at at Harvard’s school of engineering and applied sciences who was involved in the study. “It's exciting to think about how predictive text systems of the future might help people become far more effective writers, but we also need transparency and accountability to protect against suggestions that may be biased or manipulated."
With children as young as eight years old now using mobile phones, and presumably sending text messages and emails with them too, it raises questions about what impact this could be having on the developing minds of the next generation.
“Most of the data on neuroplasticity comes from paediatric populations, where neuronal connections are being formed at a rapid rate,” says Chinwe Dryer, a medical doctor practicing in Kansas. “As more and more young people have access to cellphones, it is plausible that predictive text also impacts vocabulary formation.”
There have been rumblings from the residents of Silicon Valley – the very place where much of the technology we use today was developed – who do not want their children using screens often. The Waldorf School of the Peninsula in Mountain View, California, which has hefty tuition fees of $38,000 (£29,000) and is attended by children of senior Silicon Valley technology executives, avoids the use of computers by its younger students. The school points to surveys of teachers that have suggested the use of computers and texting have hurt student’s writing skills.
我们今天使用的许多高科技技术是在硅谷开发出来的，但硅谷的居民却一直在抱怨，不希望自己的孩子经常面对电脑和手机屏幕。加州山景城(Mountain View)的半岛华德福学校(Waldorf School of The Peninsula)，学费高达3.8万美元，学生都是硅谷科技公司高管的子女，但学校却尽力避免其年轻的学生使用电脑。校方指出，教师的调查显示，使用电脑和发短信有害于学生的写作技能。
It is a view that many other teachers agree with.
“Frequently, students who use autocorrect rely so much on it to know what word they have spelled that they don’t stop to see if the word is the correct word they wanted,” says Kate Heitkamp, who teaches children in Ohio. “If students don’t have basic spelling skills, the autocorrect doesn’t seem to help since it will give them an incorrect word.”
美国俄亥俄州小学教师凯特•海特坎普 (Kate Heitkamp)表示，“经常会有这样的事，使用拼写自动更正的学生太过依赖这个技术，以至于他们不会停下来动脑筋看看这个自动系统提供的单词是不是他们想要的正确单词。如果学生没有基本的拼写技能，自动更正似乎没有帮助，因为会给他们不正确的单词。”
Over-reliance on this technological “crutch” could have longer term impacts on today’s children. A child’s vocabulary, even from a young age, can be an indicator of how “successful” they will be later in life.
Woods, however, says there is no indication that autosuggest and predictive text technologies has any impact on literacy development in children.
“Ultimately, the main benefit of such technology is to increase the speed with which we can compose online messages,” she says. “For children whose reading may be stronger than their spelling ability, autosuggest will facilitate their ability to communicate effectively online, thereby opening up texting to a younger age group, or to children who may be struggling with more conventional literacy.”
But there could be other lasting impacts on the way we communicate with each other.
Google recently revealed that it has built filters into its Smart Compose technology that prevent it from suggesting gender-based pronouns. A research scientist at the firm discovered that when he typed a sentence including the word “investor”, the AI tool assumed that the accompanying pronoun should be “him.” Similarly, the AI technology assumed that “doctor” was male and “nurse” was female. To avoid embarrassment, Google opted to remove gender pronouns entirely from the system.
Regardless, our use of technology that can write for us is likely to grow. Entire sports reports are now written by artificial intelligence from strings of data. Others are using the technology which underpins predictive text to write new forms of fiction.
Is there a risk that the languages we love and use everyday will lose something as a result?
“By taking prediction away from our language system, more advanced predictive text fills it in with something that regresses to the mean, reducing variability,” says Geffen. The result, she warns, risks making the words we write ever so slightly less human.