Something That Is Ours

I have made a few albums with AI and put them on Spotify. I don’t think anyone but me listens to them. But I like them. It’s music I think is great, and when a song hits what I was looking for, the feeling is quite simply: Fuck, that was good.

It’s a slightly strange experience to bring into the debate about artificial intelligence and culture. We talk about whether machines will take over musicians’ work, fill services with cheap content or write books nobody asked for. These are important questions. My records just point to another possibility. There doesn’t have to be a market for what I make. It can be enough that I want to listen to it myself.

I have also used AI to make stories for my children. With ElevenLabs, I made audiobooks with a narrator’s voice and a separate voice for each character. It was really wonderful to be able to bring the story to life that way. Several people I know have also made stories for their children.

A story for your children doesn’t necessarily have to compete with the best children’s books. It can do something different because it was made for them. Imagine a dragon living beneath the local supermarket. The child knows the place. The dragon can have a name they invent together. It’s an imagined example, but it’s easy to see how the work of making the story could become part of what happens between an adult and a child.

Of course, we were making things for one another long before AI. Birthday songs, bedtime stories, photo albums and home movies have never needed a large audience. Now we can get help with parts of the work that previously required skills, equipment or other people. A story can have several voices. A musical idea can become something you can actually listen to.

That opens up more than cheaper production. More of us may get the chance to make something that previously remained a thought. And some of it will interest only a very small number of people.

Where do we meet?

Yuval Noah Harari returns to the importance of stories for human beings’ ability to cooperate. We can organise ourselves in large groups of strangers because we share ideas about money, nations, laws and institutions. Stories aren’t just entertainment after the day’s work. They help hold society together. In Nexus, he also examines how information can create vast networks without necessarily bringing them closer to the truth.

A banknote and a bedtime story don’t do the same thing. But on a smaller scale, we also know the importance of shared cultural references. A song can get a whole room singing along. A film gives two people who don’t otherwise know each other something to talk about. We can quote a line and expect someone to know where it comes from.

If each of us can get precisely the songs, stories and films we want, it becomes less necessary to seek out what others experience too. My own record suits me. Yours suits you. We no longer have to meet in either of them.

In Schrödinger’s AI, I wrote about AI being able to tell millions of people their own version of a story. That possibility is both fascinating and unsettling. When a story is tailored to me, it can also spare me what I don’t already understand or like.

The debate about social media echo chambers concerns, among other things, the fear that we encounter only our own views. AI could, in principle, take that tailoring further: the story is written for me from the outset. If that also applies to the story about society, we may find it harder even to agree on what we disagree about.

But that would be a strange conclusion to draw from my records. Liking my own music doesn’t mean I’ve acquired my own reality. Nor does a story for a child automatically shut the family off from the rest of the world. There is a difference between having your taste satisfied and having all your assumptions confirmed.

A story for those who belong

When we talked about that risk, I thought of When the Dead Speak. That essay began with letters from the artist Bertha Wegmann and her friend, which my mother had collected along with knowledge of their time. I used AI to try to create a coherent narrative in Bertha’s voice. I couldn’t know whether it really sounded like her. But the attempt showed how a large body of material could become something you could read and experience.

Why should that possibility be reserved for people many others are interested in?

You could imagine bringing together your grandparents’ story. Letters, photographs and recollections from those who knew them could form the basis of a book or an audiobook. You could tell the story of the town you grew up in and the places that mattered to the people who lived there. The family or the town still has work to do finding and examining the material. The machine cannot know what a photograph means if nobody remembers who is in it.

Even so, the help can make a difference. A family history that could never be sold to a publisher can be worth spending time on for those who belong to the family. It can give a child access to people the child never got to meet. A place can acquire a story its new residents can step into.

So the small stories needn’t only compete with what we share. They can create it too. A story almost nobody cares about can bind together precisely the people who do.

That doesn’t replace our need to talk across families and towns. But there’s also something impoverished about measuring a story’s value by how many strangers it can attract. Grandma’s life can matter without being interesting to the whole of Denmark.

If everyone tells stories in the same voice

There is another risk in letting AI help us with all these small stories. The content can become more personal while the way of telling it becomes more alike.

Imagine a thousand families having their histories written. The names and places are different, but all the books have the same calm narrative voice. Adversity becomes learning. An old photograph opens a chapter, and in the end we are told what the family has always understood deep down. It’s an imagined picture of something the models may tempt us towards: a story that hangs together so well we don’t notice what has been smoothed away.

The risk also has a concrete line of research behind it. In an experiment by Anil Doshi and Oliver Hauser, some participants were given access to AI-generated ideas while writing short stories. The stories were rated as better written and more entertaining, but the AI-assisted stories also resembled one another more. It was a limited writing experiment, not a verdict on the future of culture. Even so, it shows how a gain for an individual text can come with less variation between texts.

Nor is it a law of nature that a language model makes everything alike. The result depends on the model, the material and what we ask it to do. But familiar narrative patterns can be easy to fall back on, both for the machine and for those of us who accept its suggestions.

Grandma’s detours may be part of her story. She uses an expression nobody else uses, repeats something or tells an episode that doesn’t end in a point. A machine can make it more elegant. It can also remove what the family would recognise her in.

Then there is the difference between bringing a story together and inventing what is missing. Plausible dialogue is still invented. It didn’t become true because it sounded good. If a family history is to be something future generations can build their memories on, it must be possible to see what comes from the letters and accounts and what has been added.

Who works out what we want to say?

That difference cannot be resolved simply by having the human write every word. A large part of the possibility would disappear again. I don’t need to be able to play every instrument to have a sense of what music I’d like to hear.

In the AI debate, people talk about automation and augmentation: letting the machine take over a task or using it to extend what we can do ourselves. When we create culture, the boundary is hard to locate in the finished product. AI may have written almost every word while a human has still cared deeply about what the story should be. Conversely, you can correct ten commas without having taken a position on anything at all.

In The Spotter, Not the Crane, I explored the difference between help that removes a challenge and help that makes us better able to meet it. In storytelling, the aim may not be to train ourselves to do without AI one day. The aim may be to give greater reach to something we want to express. But then we must still notice when the help takes us away from it.

My excitement about the music comes when it hits what I wanted. That feeling doesn’t require me to have done all the work myself. There is still a difference between hearing that something works and hearing that it works for me.

This essay was also made with AI. We went back and forth about the idea. The AI suggested stories people could make for one another. I brought in the risk of losing what we share, talked about my records and audiobooks and moved on to grandparents and local history. Afterwards, we talked about how all those different stories might end up being told alike.

That is not proof that this text is free of that risk. The AI can suggest an elegant connection I hadn’t seen myself. It can also make a thought seem finished before I am finished with it. A smooth formulation is pleasant to accept. It takes attention to feel that something is still missing.

The records I’ve put on Spotify don’t need a large audience to give me something. In the same way, a story about grandparents could have value for those who share their history. With AI, more of these things can take a form we can hear, read and pass on.

If a child one day listens to such a family history, it is worth hoping the child also encounters something they wouldn’t have thought to order. An unusual turn of phrase. A detour. Something that is there because Grandma was another person — and that we let remain.