How Smart Reply attempts to mimic the way we talk.
Last month, researchers at Google unveiled Smart Reply, a piece of artificial intelligence that scans the e-mail you’re reading on your phone and suggests three possible responses. Why bother composing an answer yourself? Now you can choose one of Smart Reply’s with a quick tap. “Do you have any vacation plans set yet?” asks the sample e-mail. “No plans yet,” you might choose; or “I just sent them to you”; or “I’m working on them.”
Smart Reply uses neural networks to calibrate its future suggestions, meaning it learns from how we use it. But Greg Corrado, a senior research scientist on the project, observed a “bizarre feature of our early prototype”: “its propensity to respond with ‘I love you’ to seemingly anything.” Analysis suggested “that the system was doing exactly what we’d trained it to do, generate likely responses—and it turns out that responses like Thanks, Sounds good, and I love you are super common.”
Thanks, sounds good, I love you: it might be that our culture is glutted with gratitude, approval, and love, and that our e-mail responses reflect this. But at the heart of this apparent abundance of amity and tenderness, I suspect, is something more slippery. Think of Saussure’s distinction between langue and parole—between a system and its actual use; or Austin’s distinction between semantics and pragmatics—between the meaning of words and how we actually use them in speech-acts, in using words to do things.
“In practice, any engineer’s ability to invent ‘rules’ would be quickly outstripped by the tremendous diversity with which real people communicate,” Corrado explained. Thus Smart Reply has been built on a map of “thought vectors”—concepts related to other concepts—instead of a map of word vectors—words and their meanings. The AI scans our correspondence to learn how we relate thoughts to other thoughts—it learns language as we actually use it, rather than learning words as we formally define them: parole, and not langue.
Therein lies the rub. Thanks, by definition, expresses appreciation and gratitude. But we know there’s more to the word than that. In parole linguists have found that the scope of thanks extends to cover “conversational opening, changing, stopping, closing, leave-taking … dissatisfaction, and discomfort.” “Where all men take bribes and give them, nobody does,” wrote William Golding in The Double Tongue: when a word can mean anything, it means nothing. And this near-emptiness allows its meaning to be determined by the user’s intent. Sounds good, when you think about it, is not much clearer in practice than thanks: its uses can range from approval to indifference to a very particular kind of skepticism. (“I predict your failure, but I will not intervene to help.”)
I love you, in its formal semantic meaning, is at once fetishized and sacrosanct; our familiarity with it as a speech-act is equally uneasy. Type “using I love you” into Google and the first autocomplete result is “too much”; the second is “as a weapon.” “We are hard on each other / and call it honesty,” writes Margaret Atwood in “We Are Hard on Each Other.” “The things we say are / true; it is our crooked / aims, our choices / turn them criminal … If I love you / is that a fact or a weapon?” Potentially a weapon, Iris Murdoch seemed to think, potentially a trap. “Nobody loves you as I love you,” Mitzi tells Charlotte in An Accidental Man. “Nothing follows it,” decides Charlotte, and leaves. “The most important aspect of the matter as far as I’m concerned is that I love you,” Tom tells Emma in The Philosopher’s Pupil. “I love you too,” Emma replies, “but nothing follows from that.”
If, as Corrado puts it, Smart Reply resorts to I love you “as a safe bet if it [is] unsure,” then its fragile reasoning is simply a mirror of our own—of our means for dealing with the bewildering uncertainty of intimacy. Without a sense of the meaning of the phrase in langue, and taking its use only in parole, Smart Reply can’t discern the dissonance between the two; how could it compute that saying I love you seemingly at random is inappropriate, when this is the Pollock-like pattern of our use of the phrase?
But to regard this as a failure of Smart Reply, or a transgression against language on our part, is to misunderstand how contradictions and ambiguities, misdirection and dissembling, may be necessary to us. There’s a tradition in Ethiopian culture called “wax and gold” that has its origin in the nation’s Orthodox Church: the “gold” of one’s meaning is hidden beneath the “wax” of the surface. For example: “Greetings to you,” said the notorious priest and poet Aleka Gebre-Hana, using the plural form of you, when he encountered a man and his mule on the path. Was the holy man denoting respect, or was he implying that he considered the man and his livestock to be social equals? The theologian Mohammed Girma has drawn a contrast between an American anthropological analysis of wax and gold and an Ethiopian philosophical analysis: for Donald Levine, the wax-and-gold trope allows one to create expressive and critical space in the strict setting of Ethiopia’s Orthodox Church; for Messay Kebede, wax and gold is a hermeneutic practice within the Ethiopian Orthodox Church rather than against it—a way of drawing attention to the dualism of metaphysics, the dissonance between appearance and reality. “That is wonderful,” your partner says, and her disappointment is amplified and made poignant because, of course, it is not wonderful, not wonderful at all (you can see that now).
Smart Reply “can take care of the thinking and save precious time spent typing,” promises the software engineer Bálint Miklós. Perhaps it might save the sender time in typing out a reply, but it won’t save the recipient the effort of interpreting it. Smart Reply may never be able to iron out language, because we’re constantly abusing our lexicon, using it to express ourselves beyond the range formally expected or permitted. Google researchers have reduced the frequency of I love you as a suggested reply by finding a way for Smart Reply to calibrate the probability of its being a suitable response given the message it’s responding to. This makes it unlikely, but not impossible, that I love you would be suggested for work correspondence: it depends on what you are doing at work, and with whom. Fortunately or not, I love you, in all its nuclear potency, remains an option for non-work replies, for which you can deploy it to whatever end you’d like. Sounds good—thanks.
MG Zimeta is a writer and academic philosopher, and an Honorary Research Associate at the Department of Science and Technology Studies, UCL.