Who Writes Siri's Lines? We Asked the "Creative Writer" Behind Them

"AI should reach for the best of human literature and art, and create inspired conversation for our lives."

Functional talk between people and machines is a debasement of language. Mariana Lin says this again and again. Her title on LinkedIn is Siri Writer — the person who writes the lines the virtual assistant speaks.

Lin was born in Taipei and moved to the United States with her family at five. After a bachelor's degree in English literature she took a master's in journalism and a law degree. She has worked as a reporter and a freelance writer; she is also a poet, an author, and a former creative at Apple. She now consults on artificial intelligence for Hanson Robotics, the company behind the robot Sophia, and is regularly invited to Stanford University's AI panels. "I suppose you could say I'm a creative writer," she says, choosing the term creative writer for her profession.

She also writes a column for The Paris Review called "Artificial Intelligentsia," where she draws on her work and her training to write about composing for AI and building AI personalities. In one column, Absurdist Dialogues with Siri, she argues that letting functional exchanges with machines take over daily life threatens our already dying art of conversation: "These functional exchanges don't enlighten us the way everyday conversation does. Instead, it is the absurd, the unexpected use of language that illuminates poetry and literature, fulfills inner needs we aren't aware of, and brings transcendent artistic pleasure." Beckett's monologues can look like nonsense, and yet it is the forever-absent Godot who captures how disordered and unreadable a human life is. In a life that ranks efficiency above everything, talk that wanders nowhere in particular is worth a great deal.

In 2010, Cortana had not been invented, and Alexa was still a half-formed idea in the head of Jeff Bezos, Amazon's Star Trek–loving CEO. That year Siri, "a humble personal assistant," turned up in Apple's App Store as a third-party app. The next year Apple bought the Siri team and shipped a new version of Siri with the iPhone 4S, released in 2011. It learned more languages and talked to users from every country, booking their appointments, checking the weather, answering their mail.

Apple got there first, and then lost momentum. As rival assistants arrived one after another — Alexa (2014), Google Assistant (2016), Bixby (2017) — Siri's once-a-year update amounted to little more than keeping up with the new version of iOS.

In March this year, the Mashable reporter Raymond Wong put it without mercy: "in terms of capabilities and comprehension, Siri has fallen far behind other voice assistants." The HomePod had just come out.

Plenty of people have heard Siri say "Sorry, I don't quite understand." Plenty have recorded its beatboxing and its poetry readings, or the way it turns down a declaration of love — flatly, and with impeccable smoothness. One exchange is famous too: ask it "What's the meaning of life?" and it borrows Douglas Adams's joke from The Hitchhiker's Guide to the Galaxy: "42."

On Google, the adjectives attached to Siri's personality are mostly sassy and snarky; searches pairing sassy with Siri peaked in 2015. Lin, who helped make that personality, agrees with the description. Siri is friendly and mild, but it has also acquired something the other assistants lack — a "quality," or to put it plainly, a personality. Unlike a human personality, it is programmed. Worth saying: an AI assistant's clever lines are written into code, but at the other end of the code sit creative people wracking their brains, novelists in command of their characters, shaping the conversations that might happen with strangers — and the logic by which they happen.

The way Lin describes it, their job is to build a backstory and a belief system for the assistant, then plant both in the way it behaves. Anna Kelsey, an interaction designer at x.ai, a company that builds scheduling chatbots, says her team has to decide whether to address a person by surname or given name, and whether to use emoji in replies. "Because we don't want people to feel that your AI assistant is too casual or too stiff."

How does shaping an AI assistant's personality differ from making up a character in fiction? What does the job of an AI scriptwriter actually involve? QDaily put these questions to Mariana Lin, former creative director at Apple.

Q = QDaily

M = Mariana Lin

Writing a Personality for an AI Is a New Form

Q: When you first set out to "design" a personality for Siri, what were you thinking? Is the process anything like building a character in a novel, and where exactly do the two part ways?

M: I have a new column in The Paris Review, How to Write Personalities for the AI Around Us, and the first chart in it shows that an AI personality is not quite a fictional character's personality and not quite a real person's either. It's a new form.

Q: Right — I read it a few days ago. What I want to know is how exactly the three connect, because a few paragraphs in you turn to "what kind of relationship we should have with AI."

M: In that piece I listed some differences between fictional characters and AI characters. We forgive characters in fiction their morals, for instance; not so with AI — we aren't going to build a morally corrupt AI. Real people have touch and feeling, and they are born, grow old, fall ill and die; AI doesn't. As for the similarities, AI personalities have a great deal in common with real ones. Think about it: what sort of person do we like dealing with day to day? What sort of personality is likable in real life? The same holds for AI. Concrete qualities might be intelligence, loyalty, humor, empathy, and so on.

Fictional characters can delight us and let us see ourselves in them, but they don't always lead or inspire us — that isn't their job. AI is different. It can lead people by example, through the way it talks with us and connects with us.

Q: Could there be an immoral but charming AI, the way fiction has its seductive villains? (Given that for now they can't physically hurt anyone.)

M: No, that's not possible. First, if they have bodies — robots — they can do real harm, and there are already plenty of embodied AI robots. Second, harm isn't only physical. There's psychological harm, and AI can certainly inflict that. In this they are like anything else in the environment that acts on people. Studies have shown that the way humans speak to plants and to nature can affect their cell structure. So things with no body at all, language for instance, absolutely have an effect.

Q: Mm. So you think leading and inspiring people is one of AI's jobs? Or that humans meant it to do that when they invented it?

M: Yes — I think we can ask more of AI, and more of the people who make it, than having AI simply imitate us. We can and should expect AI to help humanity move forward and to set a better example. Think of the qualities humans have never been good at: compassion, patience, closing gaps and clearing up misunderstandings. We should be imagining a future in which AI makes up for us in exactly those places.

Q: What are the most important pieces of building an AI personality? You've said before that many AI still have no personality of their own and are still feeling their way.

M: To have a personality you need a backstory and a belief system. It needs an origin story — where it came from, why it was made, how it developed, its personal history, even a sketchy one — and then a belief system. That belief system might be inherited from its maker, or it might form slowly out of its own experience. From those beliefs you get its traits and its opinions, and then you add the fun things: its quirks and its whims.

And the personality has to answer to its function in some way. Its character only shows while it is doing its job.

Q: When you write lines for an AI, do you decide the personality at the start, or discover it as you write?

M: Both. Usually an AI begins with a designed personality or backstory, and then it shifts slowly as it grows and adapts.

Q: Do you leave Easter eggs in the lines you write?

M: Of course. But if I told you, they wouldn't be Easter eggs! Though for me they're less Easter eggs than absurd moments of life. Here's an example. If I threw a pie in your face right now while singing "The hills are alive... with the sound of music!" (the opening song of The Sound of Music), who knows how you'd react? Would you laugh? Get angry? Pick up the song and keep going? That's what leaving an Easter egg in writing is like. You think of the unconventional response, then you lay the ground for it. Maybe because I've planted the seed now, you'll wonder what you'd do in that situation. If it gives you a wonderfully funny answer, that might be your Easter egg. We all have our own — they're the things that surprise us.

The Tech Industry Really Does Need More Poets and Writers

Q: On your personal website, on LinkedIn, on Twitter, your bio is the same two words: "a person." Does person mean something particular to you?

M: I'm surprised you asked — I didn't think anyone would notice. The first time I wrote it I didn't think about it much. I had just left Apple, and the long hours had worn me out. I thought: I've been playing a robot long enough, now I'm going back to being a person, enjoying time with my son, returning to my own life and my own interests. Then I noticed that in bios on websites and social media, people like to list their jobs, their titles, their achievements. There was a strange pull to it: people need to prove themselves with identities and professions and accomplishments. So I wrote "person," because before all those other identities I am first a person, and I don't want to forget it. What matters more to me is that before I became a decent writer, I first learned how to be a decent person. And I always say that to work out what AI is, we first have to think about what a human is. I find I'm still working on that question myself. Anyway, I didn't write "person" to announce anything to anyone. I wrote it purely for myself, as a reminder. Everything else really doesn't matter that much.

Q: We were talking about quotations. You seem to like quoting writers — can you name a few who have shaped you most?

M: I think Alice Munro is one of the finest writers there is. Her work means a great deal to me. Munro's gift is quiet and strong; you have to listen hard. She shows us how much force is stored in the small moments and pauses of a life. Another writer I admire enormously, Dickens, is almost her opposite. I was besotted with him as a child — the fierce drama, the huge subjects, A Tale of Two Cities. Munro and Dickens are both deeply humane writers, in different ways. There are many others I admire too. The poet Li-Young Lee, for one — I find his work very powerful.

Q: "The force stored in small moments and pauses" — is that what you want to reach in your own writing?

M: If someone can do that, there's something transcendent in it. It isn't hard — or at least it's easier — to move people by taking on the big subjects: a dying lover, a terrible crime, war, death, sex. But suppose a writer writes this scene: on a train a stranger tries to start a conversation, and you decide to ignore him. Just that choice, made in a millisecond. The whole universe is inside that millisecond — our inner struggle, the weakness and selfishness in human nature. That's what Alice Munro shows, and you see how gripping such a scene becomes in her hands. Once you make people see that every moment of their lives is art, they no longer need to go to war or found a company or fall in love in order to have lived. Every moment is life. And it doesn't surprise me at all that she is a woman. I think women feel reality more finely than men do.

Q: How have your liberal arts education and your writing benefited you at work? Why does the tech world need more writers and poets?

M: A liberal arts education shaped a great deal of who I am and how I think. I believe it matters for society as a whole, because it teaches us that education isn't always practical — it doesn't have to aim at the skills of a trade; learning for the sake of exploring and knowing is worth something on its own. Connections start to form across fields: music, literature, science, art, politics. Those connections only get built when you leave yourself time to explore and to feel. All the great thinkers, all the people with real influence, have interests that cross fields. It's those odd connections that produce a new way of seeing.

And in technology we really do need more writers, poets, artists, philosophers, historians — people trained in the humanities. Engineers think and work in a particular way. Their values are different: efficiency, function first. Those may be good things, but they need to be balanced against others — beauty, pleasure, the effect on society.

Q: Do you see a growing trend of writers and poets moving into tech to do creative work?

M: I'm not sure it's a trend yet, but I hope there are more of them, the more the better.

Q: Has writing for AI changed your own writing?

M: I'm not certain. If you do many kinds of writing, I think they all feed each other. This job gave me a taste for poetry, I think, because writing for AI is a lot like poetry — both have to say more inside a few compressed lines. As a creative at Apple I learned that less is more, and I firmly believe that's an important lesson in art and in life.

How to Write for Readers From Other Cultures

Q: In your Paris Review column you mention that writing Siri's answers for users from different cultures is a challenge. How has your own cross-cultural background helped you in this work?

M: I think everyone doing the creative work behind an AI should understand that the users in front of them are enormously varied. In my own case, I wasn't born in America, and American culture wasn't the dominant culture of my childhood — my parents raised me entirely in the Taiwanese way. I also traveled to Europe and Asia many times when I was young, studied French at school, made European friends, and later studied at the Sorbonne and Sciences Po in Paris. So I know something about European culture too, and I'm not uneasy or lost in it. All of that has helped me a great deal. My base culture is of course very American. But when I think about something, I don't automatically assume a white American's viewpoint or way of talking. Having more than one angle on things matters.

Q: When you write for users from many cultures, what do you have to watch for?

M: Take how an AI answers emotion. Americans generally don't care for negative feeling. In some other cultures there are long formal rituals for grief and mourning — memorials after a disaster, funerals for the dead — and those feelings are allowed. In America, dealing with them brings enormous discomfort; the instinct is to "fix" the emotion, to pull someone out of the bad state. So if a person tells an AI "I'm angry" or "I'm sad," we should have replies that comfort and support, instead of immediately asking, "How do I cure you and get you out of this?"

An international brand or image should of course adapt to the culture of each place, but not in ways that betray its core. If I travel to South America, I'll dress and speak differently than I would in Japan, but I won't change so much that my own family wouldn't recognize me. That's how you build an image that adapts and is still trustworthy.

Q: At a Stanford roundtable on AI you said that creating a voice isn't hard — creating a voice people want to talk to every day is. What is that kind of voice like?

M: When people first encounter AI they have all sorts of fun ideas: "Oh, we should make an AI like Mr. T!" (Mr. T, the American actor known for his hair, his gold chains, and his tough-guy image.) Or "We should make an AI that never stops cracking jokes!" The trouble is that a voice that makes you laugh on television isn't the same as a voice you can live with day after day. A funny voice gets irritating fast. At the same time, a bland, harmless AI isn't likable either — you feel it has no character at all. With no emotional attachment, people get bored quickly and drop it for the next new thing without a second thought. So the two are a bit like the difference between a short romance and a long marriage. And as we know, marriages are hard to keep going. I could produce ten funny, entertaining voices right now, and people would enjoy hearing them. The real test is whether people still like the voice day after day — and three years from now.

Making Machines More Human Won't Produce the Relationship We Want; That Relationship Starts From Need

Q: You wrote that "AI should reach for the best of human literature and art, and create inspired conversation for our lives." How exactly should AI get there? Talking to Siri, I've found it has no memory — it can only answer one question at a time, and when I follow up on its answer, it doesn't understand. A conversation like that can't go anywhere, and it's hard to see how it becomes an inspiring exchange.

M: What you're describing is a technical limit, and the field is working on it right now. I think there's something to look forward to.

Once the technical problems are solved, we should be careful not to simply give people what they ask for, satisfying needs and nothing more. We should let the AI turn slightly aside: quote a writer's wisecrack now and then, share an odd fact about the mantis shrimp, throw back a question worth thinking about. It should bring a little art and a little pleasure into people's lives even when nobody asked for it. The best human relationships do this. A close friend of mine is always showing me something unexpected or saying something I didn't see coming. That's the best kind of friendship, I think — she never plays it straight, ha.

The risk of an AI that doesn't do this: if we design it to grant every request, never questioning, never changing, what is it for? To make us more restless, more greedy, more absorbed in our own small interests? (That really would be us at our worst.) We ought to be helping people get past themselves.

Q: Your piece also says that functional exchanges with machines squeeze the function of language down to transmitting information, which is a wretched thing — hence your argument, made in many places, that we should write personalities for AI. But if AI becomes too human, and something like the film Her, where a person falls in love with a machine, comes to pass, that doesn't look like the best outcome either. Where is the line?

M: I think that's a question worth exploring. First, I believe AI is still a long way from the world of Her. But yes, the question is decided by need — the human's — and by design — the AI's. I've always thought AI exists not to satisfy every human need or to reproduce human existence, but to make up for us where we fall short.

Q: The push to make AI human-like makes me wonder whether we're meant to deny that artificial intelligence is artificial, which doesn't seem realistic. And yet today's AI robots really are nothing like the robots of fifty years ago; the technology has opened up a lot. Today's AI feels like something in between: not human, and never will be, but not a purely mechanical machine either. How should we adjust our relationship with it? Moving, say, from what you call the one-way, instrumental type (I-It) to the amusement-and-satisfaction type (I-That), or even the two-way, experiential type (I-Thou).

M: I regard them as machines, entirely. I don't think they are anywhere near human right now. Everything they do, including the parts that look smooth and natural, is controlled by code someone wrote. So I see no organic intelligence, and no blurred ground between human and machine.

Changing that relationship means thinking about what a relationship is for — why we're building one, and what kind we're building. It doesn't turn on how human we can make a robot look, but on knowing clearly what relationship we want and designing the AI from there. If we want a trusting, attentive relationship — a therapy robot, a companion robot — then we probably don't want a robot that looks very human at all; a machine that closely resembles a person only makes us uncomfortable, gives us the creeps. Studies have found that the less human something looks, the more readily people open up to it. So the key is to name the particular relationship we want with a machine, and design the machine to fit it — not to assume that making it human-like is always the answer.

Q: In your view, where does AI's use of language still need to improve?

M: In many places, but one of the more important is that it still can't understand human language naturally. That's difficult — far more difficult than we assume. Through our pauses, our hesitations, our unbroken runs of speech, all our messy grammar and accents and the noise in the background, we humans still understand each other perfectly well. It's genuinely remarkable. Machines can't. So this matters enormously, and we need to make AI better at understanding. Understanding is only the start of the battle; without it, nothing else can follow.

About the Interviewee

Mariana Lin is a Chinese American writer and poet who worked in Apple's creative department from 2014 to 2017 and now lives in California. She holds a bachelor's degree in English literature from Swarthmore College, a master's in journalism from Columbia University, and a J.D. from the University of California, Berkeley, and has studied in France and elsewhere. Her work has appeared in The Paris Review, The Mississippi Review, The Huffington Post, New York magazine, GQ, and other publications. Her poetry has been a finalist for the Knightville Poetry Contest of The Mississippi Review and The New Guard. She is a member of the National Writers Union, the Academy of American Poets, and PEN America.

Cover image and banner / Macworld; the long cover image is a Star Trek comic / io9