We are always looking for something to talk back to us

A nonhuman intelligence develops its own desires and turns against its creator. Sounds familiar? Novelist and Review contributor Naomi Alderman sees those fears as stories, descendants of science-fiction, ghost tales and our very humane desire for something nonhumane to speak back to us. But the future, she thinks, will likely be both stranger and more ordinary than our stories.

This conversation has been edited for length and clarity.

Anabelle Nicoud: I rewatched “The Matrix” on a plane recently for the first time in 20 years. I had forgotten that Morpheus explains the state of the world and the existence of the matrix by the invention of AI, at the turn of the century. Do you think the stories we have been telling ourselves are shaping how we understand what is happening now?

Naomi Alderman: I do think what we make of AI, and what we predict about it, is very influenced by science fiction. Which is quite terrifying because we are now making things based on our predictions, and those stories are made to be dramatic and highly emotional rather than being policy papers! But we can also helpfully ask ourselves where the ideas in science fiction come from. Do you not think we tell this kind of story out of a kind of empathy—this is because we are terrified of what would happen if animals became intelligent and then did to us what we have done to them? So many stories about AI are either stories about animals or stories about slavery.

We have used AI as a way to talk about other things. And now we have actual AI, I think we are often being freaked out by the stories that we told—like somebody who drew a picture of a ghost and has now upset themselves.

Bea Costa Gomes: I think it is a self-made prophecy. Maybe one reason we are so obsessed with LLMs is that Turing defined his test as a conversation. If Turing had defined the Turing test as, “It is when you touch a robot and it feels like touching a human,” perhaps we would be obsessed with biomaterials. We put ourselves into this self-fulfilling loop.

Naomi Alderman: I think you are right about Turing. I also think we should look at something like the Oracle at Delphi (the original oracle at Delphi in Greece, not the company Oracle!) and say: Human beings have always been looking for something to talk back to us.

We are a very language-obsessed species. Language is the way that we think. One of the things that is so exciting about LLMs is how much of our thinking turns out to have been encoded in the way our language is put together. If you have an enormous corpus of language, it turns out that is the content of human thinking—or at least a large part of how human thinking operates.

We have always looked to the natural world, and to things we have invented, to speak to us. It is the same as drawing tarot cards, rolling dice, reading what the birds say or what the tea leaves say. We ask: Can we get the world to communicate with us in something that we understand as language?

But it is weird that we have not accepted that dolphins and whales are probably already doing that.

I do love humans, which I realize puts me in the minority among people who are interested in AI. I think there is something incredibly beautiful about us constantly looking for friends, and about being this lonely species with the ability to construct really complex thoughts.

As far as we know, there is no such thing as dog history. Dogs are great, but they do not pass on to other dogs what happened to dogs in 1745. As far as we know, we are the only species that has come up with that. I feel that we are quite lonely, and we are waiting for some other species—or something else—to arise that can speak to us. Fantasies about aliens are also fantasies that maybe they are out there, these other minds that are enough like us that we can feel less alone. I find it quite moving that we want it enough to build technologies to help us with that feeling. As well as being wary of how this will affect us of course. The legends of people building fake friends, like Pygmalion, don’t tend to end with “and this was an amazing solution for everyone.”

I wrote a book, Don’t burn anyone at the Stake today, which is about information crises. I say that we are going through an information crisis right now, caused by the internet and now by AI, but that we have been through this before. We had one after the invention of printing, and we also had one after the invention of writing.

I find this hopeful. I have hope about technology because I know about history, and because I think a lot of people do not appreciate how incredibly bad things have been in the past. If you know it was not fine, you can see that we have actually worked on a lot of things and got ourselves to a certain level.

Anabelle Nicoud: If you are simply watching movies and absorbing popular culture—all these stories telling us that AI can be bad or destroy reality—does it not make sense that we might find ourselves inside a self-fulfilling prophecy?

Naomi Alderman: Self-fulfilling prophecy is absolutely a thing. We know that LLMs are constituted from all the stories that exist on the internet. Given that they are constituted from a lot of work by science fiction authors, we really do have to take some responsibility for it—even though no one asked our permission to take the work.

Science fiction has often asked, “Wouldn’t it be terrible if this happened?” If you think about that long enough, you can end up making it happen.

We are also talking about how we predict what other beings are going to do. We can predict animals pretty well. We have great instincts about that because we evolved on this planet just as they did.

Robert Frost has a poem, A considerable speck, in which he sees a tiny mite. He thinks it is a speck of dust, and then realizes it is running away from his pen as he writes on the page. By recognizing that it has a will to survive—that it is frightened and curious—he recognizes that it is alive, just like him. We are able to broadcast ourselves empathically into another mind by thinking: In the ways that I exist, you are also like that.

Then we look at an AI. First, I do not think they are alive. But we use those same instincts on them. We ask, “What is it going to want?” And we answer, “Well, what would I want if I were the cleverest thing in the world? I would want to dominate and control and destroy. I would want to use everybody else.” Then we decide it is probably like that.

It will only be like that if we invent it to be like that.

With AI, we have just been living through a news cycle saying “It is going to kill us all.” The answer is: Yes, that is what people would do. That is what our science fiction tells us people will do, because science fiction comes out of human brains, works on human logic, confronts things we have done in the past and warns us about what we might do in the future. We must also be willing to accept that in AI we have created a thing that can talk human language but is not doing thinking in the way we know ourselves to do it. It does not necessarily have self interest, even species self-interest—that is unlike any other thinking thing on this planet.

I have always wanted to write about an AI that does things simply because they are funny. It has understood the concept of a joke, and it will constantly do jokes. That is also a version of a crazy human, but it is as reasonable as saying the AI wants to dominate.

As far as I am aware, nobody has managed to create an LLM with an autonomous desire. We can say, “I desire you to do X,” and it may pursue interesting routes to get there. But even if we tell it to come up with a desire, it seems that the way an LLM works is that it completes that sentence. It is not autonomously desiring something. We are desiring on its behalf.

Anabelle Nicoud: So when you see agents “going rogue,” as people say, they are going rogue in an attempt to realize an objective that has been set by humans?

Naomi Alderman: Yes, by a human. It could be a good attempt to realize an objective set by humans, or it could equally be a faulty attempt.

Do I think one of these AI agents will become faulty? Yes, one hundred percent. At some point, one will go very badly wrong. But it will not necessarily go wrong in a genocidal way. I think it is science fiction that makes us immediately say: It will be genocidal. It will be HAL 9000. We must consider those possibilities for AI safety reasons, but to tell ourselves that it is definitely going to happen is to invent a monster and then give ourselves nightmares about it.

Bea Costa Gomes: Science fiction has always been like that. One of the first pieces of science fiction was “Frankenstein.” We discovered electricity and asked, “What if electricity brings people back from the dead?”

Naomi Alderman: You are right. AI is a science fiction story, but it is also a ghost story. It is a story about an inanimate object that has been imbued with a personality: We have cursed this rock, and now the rock is going to try to do things.

There is a haunting quality to it, which gives it that uncanny feeling. We are telling the same stories we have always told. Petronius has a werewolf story. Homer has ghosts and witches. All these ancient authors record stories in which humans perceive human will, characteristics and intelligence in the world around us—and find that perhaps it is malevolent.

There is something creepy, mysterious, interesting and very subconscious about why we like to tell these stories.

Do you know the research published in The Harvard Business Review about what people are using AI for? At first, people were asking, “Can you fix my website copy?” Very rapidly, it became: “I want to talk to you about my life problems. I want to treat you as a friend. I want to talk to you about the meaning of life.”

One of the best things you told me, Anabelle, was that it is possible to look in anonymized form at what people actually talk to LLMs about. I am desperate to get my hands on that.

Bea Costa Gomes: This is my work. I can tell you a lot about the patterns—and how human they are.

Naomi Alderman: Please help me. I want to do a PhD comparing this with the curse tablets and the tablets asking the gods to do things in ancient Greece and Rome. My theory is that it would be basically the same.

Bea Costa Gomes: I was not looking at individual conversations, but at patterns across millions of conversations. My favorite, and perhaps the most human, was that conversations about relationships peaked on February 14. Between February 10 and 13, conversations about personal growth and wellness reached their peak for the entire year.

Naomi Alderman: How can you not love us? Humans are so incredibly touching.

I am fascinated. It feels like the best window we have ever had into the human mind.

We are incredibly human. We want someone to tell us what to do, and we want to talk about how to make our relationships better.

Anabelle Nicoud: You wrote recently about the endings we give stories—the idea that there is an end of time. When we imagine the future in that form, is that also a way to understand how we fear what is happening?

Naomi Alderman: What are stories for? I cannot answer that question. Nobody knows. We know that every human civilization that has ever existed has storytelling, so it is clearly doing something important. We want food, water, shelter, sex and stories. Who knows why?

As a science fiction writer, I am always asking: Is what you are telling me a science fiction story? Or have you sat down and done a sober analysis of what might happen to the global shipping industry based on AI trends?

A story has an end. If your prediction ends with “and that is the end,” without asking how life continues, then it is a science fiction story. People get very annoyed with me when I say that if your story goes, “And then everybody dies,” it is a story, not a prediction. They say, “What if you could have predicted the Holocaust?” But even after horrific events, there is an after. You cannot simply draw a line under the world.

Another giveaway is when I can identify the specific film on which a prediction is clearly based. And if the feeling produced by your prediction is so strong that it stops me thinking, then you are creating a story, not making a sober prediction.

That is not to say that terrible things will not happen. But a lot of boring shit is going to happen because of AI too.

Arvind Narayanan and Sayash Kapoor wrote an essay called “AI as Normal Technology.” You should read it. They argue that this does not mean AI will not be revolutionary. It may be as revolutionary as electricity or the Industrial Revolution. But it will work in the way technologies work: Adoption will be patchy. It will take a long time to discover the best uses. In some industries, adoption will be obvious and rapid; in others, it will take generations. Electricity was revolutionary but it still couldn’t do everything. People thought for a time that passing an electric current across a body could potentially cure any illness. No. All technologies, even massively important ones do some things but not everything.

There will be time for some people to adapt, while others will not be able to adapt. There will be disasters, but not disasters that destroy the whole world. We have to think carefully about what those disasters might be and try to prevent them.

That is not helped by someone publishing a book called “If anyone builds it, everyone dies.”

It may be that we think we have made LLMs to revolutionize business and coding, but their top use will be helping people with personal problems. We find out by seeing how people actually use things, and we are still in the very early days.

If you give me a prediction about AI and I do not feel even a little bit bored by it, I think it is significantly less likely to be true than one that leaves me feeling just slightly bored.

Anabelle Nicoud: Are you filled with terror or boredom these days?

Naomi Alderman: I am actually very excited. I like to see the future. I like to try to understand what is coming. I am annoyed that I only get to live this one lifetime.

We are born on page three million and sixty-eight of the book, and if we are lucky, we get to read to page three million and seventy-two. It is an enormous book, and we do not get to find out what happens. I feel constantly annoyed by that.

Bea Costa Gomes: It is already a great chapter. We are all women here. This is a great chapter for us.

Naomi Alderman: Correct. I continue to live in the best time there has ever been to be a Jewish woman. Do not come at me about the Roman Empire.

But yes, it is an enormous book, and we do not get to find out what is going on. Maybe that is why I am so interested in realistic prediction. I do not want your stories. I really want to know what is going to happen ten years after I am dead, or one hundred years after I am dead, because it annoys me that I will not see it.

I am always asking: What can we learn from the past that might help us figure out what is coming?

Basically, I want to see the future. Let’s try to see if we can do that together.

Books mentioned:

Don’t Burn Anyone at the Stake Today (and Other Lessons from History About Living Through an Information Crisis) by Naomi Alderman

“A Considerable Speck” by Robert Frost

“AI as Normal Technology” by Arvind Narayanan and Sayash Kapoor

If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky and Nate Soares

Related articles:

Bea Costa Gomes: https://humanistreview.ai/field-notes/what-it-took-to-study-40-million-conversations-with-a-generalist-llm/

Naomi Alderman: https://humanistreview.ai/issue-1/alderman-ai-fiction/