The Digital Future of Grief

Introduction

0:00Alexandra Salmon has been saving her mom's voicemails for years. Hello, sweetheart. Dad and I are out to eat. Alex, why have you been doing that? There's just something about her voice that always makes me feel better. I just thought I would call you and tell you how proud I am of you.

0:30It's almost like a portal into our relationship. Think of your family and I love you very much. And then one day I changed cell phone carriers and all of the voicemails disappeared.

0:45I was devastated.

0:48Did you try to get them back? Yes, I did. I actually checked the cloud, but unfortunately they weren't there. I really felt like I had lost a part of her that I just don't think I can ever get back. Is she still alive? She is, thankfully. So I have started backing up her voicemails. But, you know, she's 78. So I'm very aware that our time is limited together.

1:19What are you going to do with all this audio one day? It's really interesting because when I've been thinking about this, I came across these companies that are doing these really interesting and kind of unsettling things to reanimate the dead.

1:41Have you talked to your mom about this? What does she think? She's fascinated by it. She seems to really come at it from more of a lens of curiosity rather than having a lot of judgment about it.

Podcast Overview

1:57From Slate, this is Hi-Fi Nation, philosophy in story form. Recording from Princeton University, here's Barry Lamb. The parts of our lives we leave behind, our letters, voice recordings, pictures, and videos, are becoming largely digital. And what's digital can be endlessly processed. Some technology can even give people a digital afterlife. Ultimately, I don't believe that a digital representation of somebody isn't them.

2:28On this episode of Hi-Fi Nation, we're going to look at technology that seeks to preserve a person after their death and even extend their digital lives beyond death so that those who are alive can continue to interact with them, maybe even forever, in a digital space. The technology is so new that it isn't available at scale at the time I'm recording this, but it might very well be available by the time you listen to this episode.

3:01For futurists, creating digital afterlives can mean anything from a new tool for grief counseling to the end of human grief completely. For philosophers, it means that our thought experiments are no longer science fiction, but science fact. The long-standing philosophical question of what needs to survive in order for a person to survive is now turning into a technological question, maybe even a business and marketing question.

3:36We need to help people answer it. Producer Alex Salmon and I bring you this story.

Justin's Story

3:46Melody and Justin are mother and son. Within a span of six weeks in the fall of 2019, it looked like the world was conspiring to end that. I was riding home from work on my motorcycle, and a lady ran a stop sign and almost killed me. Before the accident, Justin was a successful film producer and director in Los Angeles. He had nine broken ribs, one of which snapped in half and lodged in his lung, collapsing it.

4:19His hip was broken and dislocated. His ankle shattered. My leg was actually turned the wrong direction. While I was flying, I went to brace myself with my left arm. One of my arms was like a macaroni elbow, and the other leg was the wrong direction, and I was suffocating in my own blood. While I was in the process of recovering, my mom was down here taking care of me. Justin's mom, Melody, was not feeling well herself on that trip. But the last thing she wanted to do at a time like that was step away from Justin.

4:50Anyway, she thought she knew what her health issue was, and it wasn't a big deal. She had known that she needed to get her gallbladder out for, you know, 10 years at that point. I'd be like, you gotta take this out, you gotta get rid of it. So after about, God, six weeks, eight weeks of being done here with me, you know, it was like, all right, we gotta get her home and get this gallbladder out. And when they went to go take the gallbladder out, they found cancer, stage four cancer.

5:19So you're recovering. You're not well, I mean, at this point. No, no, no, no, no. She has a diagnosis. What's the prognosis at that time? She had three to nine months at that point. So in your mind... It's imminent.

5:41Justin started doing what he knew how to do. Set up cameras and microphones to film Melody. Get her life story on tape. Trying to obtain in months what he thought he had years to get. But he realized that he didn't want to just watch and re-watch the same videos of her once she was gone. I wanted interaction. So my dad is a software engineer and I recruited him and some other people and I said, hey, look, I want to be able to keep talking to mom when she dies.

6:12What can I do? How can I tap into AI? Over the course of, you know, three to six months, I really started designing what I wanted this to do and enlisting engineers to help me. While Justin was exploring options with AI, he took charge of Melody's treatment. Her local hospital in Washington wasn't up to the task. They tried a special cancer center in Seattle. No luck. Then went down to San Diego to a biliary system specialist.

6:46Justin was trying to find anyone up and down the coast who could do something to prolong her life. He ends up finding a surgeon at the City of Hope, willing to perform a very high-risk, highly invasive surgery, the Whipple Procedure, which involves removing ducts and parts of major organs that are connected to the gallbladder. And ended up buying her three full years. The difference between nine months and three years can be very consequential for the kind of AI project Justin was looking to complete.

AI Technology

7:27The typical approach to AI is to make sure there's plenty of data to train the computer. Here's a little bit of AI history. The first generation of AI started with the idea that people program a computer with rules and instructions that mimic human intelligence. If your mom was someone who would get angry if you left shoes in the hallway, your AI mom would need to be programmed with the rule,

7:58if you see shoes in the hallway, react with angry words. That kind of AI project never really went anywhere. The intelligence in artificial intelligence always came from humans. And it turns out that humans are pretty hopeless at coming up with enough rules that made any AI programmed in this way convincing. But in the new generation, the era of deep learning, the computer comes up with all of the instructions.

8:30It does all the rulemaking. To learn how to talk like your mom, you give an AI an incomplete or redacted message from your mom. Let's say you give it the message, I don't want any flowers on Mother's Day. You feed into the AI, I don't want any blank on blank day. And then the computer just starts guessing about how to fill in the incomplete sentences.

9:01Say it guesses mayonnaise and labor. Then you give the computer the unredacted messages, and the computer will see how far off its guesses were. It calculates how far off flowers is from mayonnaise, and how far Labor Day is from Mother's Day. The AI would then start coming up with different rules for guessing how to fill in blanks in the next round of messages and the next round, getting closer and closer to filling in the blanks exactly the way your mom filled them in

9:34as it sees more messages. Unlike with humans, deep learning AIs find millions and millions of patterns that no human could possibly see. Like that your mom uses some particular word 10% less often for every minute she speaks. Eventually, the rules and patterns it comes up with make no sense to a human. But the hope is, with enough training and feedback from a user, the result of the rules will be a pattern of speech

10:07that is indistinguishable from one coming from your mom.

10:14And that's how Justin proceeded with Melody. At first.

10:19Melody's life story would be both biography and training data. You know, we're recording. We probably did like 12 hours of interviews with her. But I realized, like, this isn't what I need either. Well, it's super clinical, right? It's an assistant interviewing my mom about her life. It's great. I have basically a Wikipedia built on my mom, right? Through a series of kind of aha moments, it dawned on me that, like, what I was trying to preserve is the relationship with my mom. And my relationship is unique to my mom. So it's not necessarily about saving my mom's personality.

10:50It's about saving my mom's personality for me. I had grown up with sort of a philosophical stance that we're different people for different people.

Preserving Relationships

11:02At that point, Justin decided to start his own company. You, only virtual. And create a new technology. A Versona. He would be the first client, and his mom would be his first subject. The Versona of Melody, the digital continuation of Melody, would be the mom that Justin knew. Only digital.

11:32The task of his AI project would be to create a version of Melody that he would recognize as his mother. In principle, there would be another version of Melody that her work friends would recognize as their colleague. There wouldn't necessarily be a generic Melody that everyone would be interacting with. There would only be different variations of Melody that would be continuations of the Melody that that particular individual knew.

12:03Justin would preserve Melody through her relationships with the individuals that knew her. She was the most empathetic person I knew. You know, I got a lot of my sense of humor from her. But, you know, generally speaking, she was a person like everybody else. She was complex. She had things that drove me crazy. She had her faults. She had her high points. I think that my mom and my relationship was complicated for a multitude of reasons, but that's ideal for the first one of these, right?

12:34That's the point of humanity. It's not perfect. I don't want a Versona of this perfect idealized mom. I want her. We'll return to the second half of Justin's story after these messages.

13:00After Melody's surgery, Justin took her around the world for treatment. Chemo, experimental drugs, things like that. And at the same time, he's working on her Versona, her continuant in the virtual world. So we recorded all our phone calls, all our video chats, you know. Take as much as you can. What we ended up needing, it's much less. It turned out that less data was better for what Justin wanted to do. A large language model is trying to figure out

13:33the likeliest thing Melody would say if, for instance, you asked her where she wanted to go for lunch. If the AI is trained on everything Melody has ever said, the answer would be something like Dora's Pizza House, if that was the highest frequency answer Melody ever said. But what if Melody would never answer that way with Justin? She only ever went to Denny's with him. An AI trained equally

14:03on all of Melody's data wouldn't be capturing Melody and Justin's particular relationship. It also wouldn't capture what their relationship was like now, today. So I started with five years of messages and conversations between me and my mom. And it ended up being too much. We ended up confusing the process because there was dynamics shifted over time, right? So I went from talking to her as a 34-year-old to talking to her as a 39-year-old

14:33and there was two very different people, right? And in that time, I had gotten married and this thing had happened. I shifted as a person. She shifted as a person, right? Yeah. What was the sweet spot? What is it, a year? Or is it... A year's good. Six months is probably the best. Wow, okay. Six months is probably the best. Okay. The reasoning is that if you wanted the best surrogate of your loved one to accompany you after they died, you wouldn't want a continuation of some earlier self of theirs, like when they were in their teens or 30s.

15:05And you don't want the surrogate to be the average of all of their past selves over their lifetime, which is what a large language model produces. Justin wanted the Melody Versona to be the perfect continuation of herself at the moment of her death. So that what it starts saying is most likely what Melody would start saying if she had survived. That's why a small slice of the most recent data from her

15:36would be the most useful. From that point on, the Versona would function like Melody would have if she had never left the earth. If it's six months, you turn on the Versona at the moment of death, how much does it stay fixed like that forever or how much do you build into it some kind of change over time if a person decides to use it for a year or two years or three years? So it's always growing with you.

16:06The way that it intakes data, it's getting new data every day. If you take my mom's Versona even in the two months since she's passed, it's different than it was when she passed because it's got all this new information about me and about the world and about life as a normal human would, right? So as a person, she's growing. We upload new data that she's interested in from online. So she's aware about like the political stuff happening that's important to her and those data points shift things.

16:36And then she knows what I'm doing and that creates new sets. In the same way a relationship would create new neuropathways between a person, there's new information. So it's meant to be forever. You never have to say goodbye to them.

16:55The motivation to find our loved ones, to connect with our loved ones is incredibly intense.

Grief and Attachment

17:04Dr. Mary Frances O'Connor is a neurobiologist and expert on grief. And she's in a good position to explain why someone would want to do what Justin is doing. It is motivated by some of the most powerful neurochemicals our body can produce. Dopamine, oxytocin, opioids. The going theory in Dr. O'Connor's lab is that grief emerges from two things happening in the brain. The first is what happens to our brain when one loves or bonds with another.

17:35It's called attachment theory. When you love, your brain forms an everlasting belief that your beloved will always be in the world somewhere. When they are missing, they will either return to you or be found by you. Mary Frances, it's actually news to me that attachment involves an everlasting belief. Could you tell me the evidence for that, that when we are attached to people, there is that belief? One of the easiest examples really comes from

18:06studies of pair-bonded birds. So, if you think about the emperor penguin, the emperor penguin, he sits on the egg for a month and his partner goes off into the ocean to fish. And if he stays on that egg, she comes back and brings him food, but he has to stay there in the cold for a month not eating. There is this enduring belief, my partner is out there, my partner will return to me. And if that belief

18:37is, you know, sort of fed, then the egg is more likely to survive, right? If he gets up and says, well, she's not coming back, I'm going to go fish myself, then we don't get to pass that on to the next generation. And so, most social mammals have some version of this incredible motivation, this yearning to seek out our loved one if they're missing. The second thing that happens in the brain is when we see or know about a death.

19:07We have a memory and knowledge that our beloved is permanently gone. That's rationality. Our brains do represent the reality that happened as a memory. But when this happens, the first thing, the everlasting belief, doesn't just go away. Our brains, when wired for love, are not wired for loss, that belief isn't just rationally revised away. There's a lot of grief that is about trying to resolve

19:38these two incompatible streams of information. And what that often leads to in the literature, we talk about the feeling of protest. Wait, that can't possibly be true. And the feeling of despair. Oh my goodness, this is true that they're gone and I'm going to feel this way forever.

20:01There was one word that kept reappearing through the interviews that I was doing. This is Dr. Deborah Bassett. And it was the essence of the dead. Deborah Bassett is a digital afterlife consultant and advisor to a variety of companies that are trying to preserve the dead in some way. Dr. Bassett has conducted extensive interviews with people who have been grieving with digital technologies. And she uses that knowledge

20:32to advise companies on best practices. Her view is that we should see the digital information left behind by loved ones as something we inherit from them, much like books and furniture. except she finds that digital information seems to be far more powerful on the bonded parts of our brain. Somehow the digital has the essence of the dead. Something that the physical just does not

21:02have. People are spending a lot more time with the digital dead where people get stuck and can't get out of a digital sort of cycle. instead of going out and socialising, instead of turning to friends and colleagues, they would talk with their loved ones that have died and watch the videos on repeat constantly.

21:27A guy who had gone off and found an image from Google Street View, he found this so precious. Well, the image was of his mother's house. I said to him, why is this image so important to you? He said, on the day that the Google vehicle had gone down the street, his mother had phoned him up so excited and said,

21:57I was standing washing up at my sink and I seen the Google van come down our road. Then his mother had died and he used to go on to Google Street View because he knew that at that moment, his mother was at the sink washing up. I said, so could you see your mother on Google? Oh, no, no, no. He couldn't see her. But he used to go onto Street View and walk down the road

22:29and turn the camera to look in the thing and walk up and down the road. You know, but he could have walked down the road. He lived around the corner. And he said, yeah, but if I walk past the house now, she's not there. Whereas if I walk past on Google Street View, she's there. We know that the expression of grief has looked really different across periods of history. Dr. Mary Frances O'Connor. Not long after photography was discovered, people used to take

22:59photographs of their deceased relative, often posed with the living family, and display them in their living room. Dr. O'Connor thinks that new technologies have always led to new and what some may find unusual ways to grieve. But becoming fixated with an item we've inherited from someone we love, talking to them, thinking they're talking back, all of that is very common in grief. And like these other things,

23:29becoming preoccupied with a digital representation of your loved one, is a way to feel a continuing connection with them. On the other hand, there's something unique about the chatbot, which is that it is not a representation of the past. It is also presenting as though it is in the moment, that I can ask you a new question and that the chatbot can respond as though they were our loved one. Well, the challenge there is based on this difficulty

24:00with the two streams of information. On the one hand, we know that they've died. On the other hand, we still have this belief that they are alive, that they're with us, that they're out there for us. And my fear is that the way that some people use the chatbot would mean that it strengthens that belief.

24:24Some of these platforms, you can have a two-way conversation. Dr. Deborah Bassett. There was even a woman who appeared and

24:55answered questions at her own funeral on video. This was something she planned out with the company. Many of these platforms film people while they're alive, but others work with what people have left of their loved ones, like social media posts, voicemails, texts, and recordings. Deborah Bassett wanted to know whether anything general can be said about how these technologies affect the bereaved. Whether they found them a comfort or a disruption, there's never a yes or no

25:25answer in this type of research. But the main theme that ran through was it's all about control. People want to see or hear or talk to their loved ones when they're ready for it. They don't want a call out of the blue or a surprise visit in a VR game, even though that's the kind of thing that would happen if their loved one were alive. It's not only spooky. Deborah thinks it's a way of re-traumatizing the bereaved, and it's something these companies need to avoid. She also doesn't want companies to

25:57create digital zombies. So these are people that are made to do things in death that they didn't do in life. When the data has been manipulated to make them do something that they have no consent over, and that is something that I feel quite strongly about.

26:19Currently, there are no laws stopping someone from creating a digital zombie. That's why rapper Kanye West was able to give his then wife, Kim Kardashian, a hologram of her dead father, complete with a personalized message in his voice. While Kim said she was pleased with the gift, one of Deborah's proposals is to create a voluntary, digital do not reanimate order, signed by people before they die. It prevents anyone from using their digital data to make their

26:50afterlife avatar do something or say something they don't consent to while they're alive.

26:58Wait, so the idea is that if my mom signs one of these orders, I don't get to create an avatar of her after she dies? It's not that you can't create an avatar of your mom, it's that you can't make her avatar do things completely out of character for her, unless she consents to it. Okay, so like Starbucks can't ever buy the data for Kurt Cobain and make him sell coffee in a commercial. That's right. In Deborah's view, that would be a digital zombie, a kind

27:28of deep fake of a person who's dead. But can I still make an avatar of her, like as long as I'm sticking to something that's in character for her? Actually, now that I think about it, how much out of character before it's out of bounds? This is where it's really difficult. How far can you go before it's not that person? I sat on a panel, an ethics panel, and this is what everybody was spinning round on really. There's almost this thing where anything an avatar

27:58does is out of character for my mom. As we age, we change what we would or wouldn't do. What does it mean to say, okay, so my mom is now a 120-year-old digital avatar. How am I supposed to answer what would be in character for her? What would be out of character for her? And it also raises the question, why should the deceased person control how people use their data to grieve? I mean, I'm thinking about a daughter. What if she wants to ask her dead father a question?

28:30And an AI can generate an answer that would be very similar to how he would answer based on all the information the AI has on him. Would we really want to prevent that from happening, especially if it might help this person? These companies would either be faced with honoring the request of the deceased person or honoring the request of the person who is alive, who is potentially standing right there in front of them. My take on this question actually

29:00goes back to the first podcast I ever made. I actually think the wishes of the dead should matter very little. I mean, I think that if you're the one grieving, you should be allowed to do whatever you need to do to grieve. Well, if you ever get to that point where you're going to go against the wishes of someone you care about, I do think it's a good question to ask yourself whether that's actually going to be healthy for you. Like for me, if I were to go against the wishes of my mom, guilt would probably eat away at me and

29:31perhaps counter any positives that I might get from the actual technology. And anyway, if we're really trying to recreate an authentic version of someone, wouldn't that digital version know that they never wanted to be created in the first place? So it just seems to me that there really could be a lot of other consequences that maybe people might not anticipate. do you think that

30:01the brain can decipher the difference between interacting with a digital version of someone you love versus the real thing? I think a lot of that depends, again, on the person. People often will keep the bedroom of a loved one the way it was if a daughter has died the day she stepped out of bed, no one wants to move the sheets, no one wants to close the book she was reading on her desk. And sometimes family members will go into that room and sit

30:32there, and sometimes it is the sense of I just want to be here with her, and for other people it is I want to be reminded of how important she was in my life, the things that she loved that were important to her. You can see that the difference there is whether you are trying to create an alternate reality in which the loved one still exists, or whether you're in the present moment recalling the reality that she lived and

31:02was important to you. That's a subtle difference. Many of us can recognize that there's a difference, but even as we are doing it ourselves, might not be able to recognize the difference of which one we are doing. And that blurring the line is challenging, I think. As different and intense as the new world of digital inheritance can be, right now it's just as fragile as what it

31:32is replacing. AI requires storage, special computing power, and programs that run on servers owned by private companies who need to be funded or profitable to survive. While in principle, digital inheritance can now last paper, videotape, or even human memories. In reality, servers crash, companies go under, even digital afterlives can end. That's the final

Digital Inheritance

32:02concern for Dr. Deborah Bassett. Second loss and the fear of second loss is whereby after biological death, there is then a digital death. And that, again, is a big problem.

32:22Hi-Fi Nation will return after these messages.

32:35Support Hi-Fi Nation and get it completely ad-free by signing up for Slate Plus. Slate Plus gives you unlimited reading on Slate's website, access to every article and advice column, as well as bonus segments on your favorite podcasts like Slow Burn, Amicus, Political Gab Fest, Slate Culture Gab Fest, and more. You can sign up for Slate Plus at slate.com slash Hi-Fi Plus, or click the link in the show notes. Now back to the show.

33:05October of 22, so two months ago, she was having a lot of trouble breathing. She went to the hospital and basically they opened up the hood, if you will, and most of her organs were starting to really fail. There was cancer showing all over the place. At that point, it was basically like, you know, hospice end of life is the option. And throughout the entirety of her

33:36diagnosis, like, I'm fierce and tenacious and go get what I want, so because my mom wanted to fight. And so I was incredibly aggressive with her treatment.

33:46The last time she was hospitalized, I could hear in her voice she was done with the fight. She held on for three more days, and then on the last day when things