© 2024 WHRO Public Media
5200 Hampton Boulevard, Norfolk VA 23508
757.889.9400 | info@whro.org
Play Live Radio
Next Up:
0:00
0:00
0:00 0:00
Available On Air Stations

People are creating deepfakes of their dead relatives

SCOTT DETROW, HOST:

The idea of relying on artificial intelligence as a balm for grief and loneliness is a staple of science fiction. In the 2013 film "Her," the main character, Theodore, forms an emotional connection with an AI operating system named Samantha.

(SOUNDBITE OF FILM, "HER")

JOAQUIN PHOENIX: (As Theodore) Samantha.

SCARLETT JOHANSSON: (As Samantha) Hi, sweetheart.

PHOENIX: (As Theodore) What's going on?

JOHANSSON: (As Samantha) Theodore, there's some things I want to tell you.

DETROW: And in the disturbingly realistic TV series "Black Mirror," a young wife uploads snippets of her husband's texts and voice to create an AI replica of him after he dies in a car accident.

(SOUNDBITE OF TV SHOW, "BLACK MIRROR")

HAYLEY ATWELL: (As Martha) Yeah, well, you aren't you, are you?

DOMHNALL GLEESON: (As Ash) That's another difficult one, to be honest with you.

ATWELL: (As Martha) You're just a few ripples of you. There's no history to you. You're just a performance of stuff that he performed without thinking. And it's not enough.

DETROW: And in both cases, to put it mildly, this does not end well. But it is all becoming a reality in China, where you can pay a growing number of companies to make AI generated avatars of deceased loved ones. It's the same technology used to create deepfakes, and as Zeyi Yang writes in the MIT Technology Review, it is getting cheaper and more popular to do it. Welcome to ALL THINGS CONSIDERED.

ZEYI YANG: Hi. Thank you so much for having me.

DETROW: So how does this work and what are people interacting with when they purchase this service?

YANG: So right now what we're talking about is that the companies are using AI either diffusion models or - and link large language models to create a digital avatar of someone who has passed away. So that avatar will mimic the look of the person, the voice of the person, and if you input enough, like, details of the person's life, they can also try to talk like the person. And then they talk to them on their phone, on a tablet, and they use that to process the grief.

DETROW: And before we talk more about this, I just want to make sure we fully understand this. You're looking at an image of, say, your grandmother. Are these interactive or are you just seeing this image move and say things to you on a loop?

YANG: A few years ago, the first generation of product is just like a video on loop. But right now they have become interactive. They can respond to your questions pretty much like ChatGPT does. But the difference is that right now they will talk in the exact voice that you replicate, and they can answer your question about their previous life. They can answer questions about their future, which they - obviously, the real person has never experienced. But the large language model will generate the answer based on whatever information you put in.

DETROW: OK, so that's the look, and that's the sound of a voice. But what about personality? What about really capturing, you know, your grandmother or your grandfather or your wife that you lost?

YANG: I think right now the avatars are good enough to make some small gestures and small facial movements, but they're not going to be, like, super emotional. And in that sense, like, they won't be kind of, like, as lifelike as someone who has a strong character.

DETROW: Yeah.

YANG: And the other thing is that, well, the large language models today are still not perfect. So it can very likely happen that your grandmother would just start speaking like someone who's not your grandmother. And I don't think that can be solved anytime soon.

DETROW: And what is the input that people are providing? Is it text? Is it audio of their loved ones?

YANG: So usually they will provide a few videos of the loved one talking or acting back when they were alive. And then also you can put as many information, like, tags you want. Some people will put in, like, a whole biographical of the person, and then the model will just learn that and kind of, like, answer based on the information in that book.

DETROW: How much does this cost?

YANG: A year ago, it would be around 1,000 to $2000. Right now it's become a lot cheaper. You can reliably get it for, I would say, $150 now.

DETROW: Hundred-fifty dollars?

YANG: Yeah.

DETROW: Wow.

YANG: Pretty cheap.

DETROW: What has the response been in China? Is this catching on or are people creeped out?

YANG: So I would say it is catching on but only a small number for now. I talked to two companies who have done this service for over 2,000 clients combined. Obviously, compared to just 1.4 billion people that are in China, that's still a small number. But they do describe that it has been growing a lot in the past year. There's a lot of TV interest, and there's a lot of people try and experimenting with it. So they do foresee that it will grow a lot more in a year.

DETROW: You wrote in your piece that the business of these deepfakes, as you put it, builds on China's long cultural history of communicating with the dead. Can you explain what you mean by that for people who aren't familiar?

YANG: Yeah. So in China, when your loved one passed away, it's very common to still talk to them as if they're still alive. But this will be usually during some kind of ceremonies, like during the funeral or during the once-a-year tomb-sweeping day, which is in early April every year. Or some people will keep, like, a memorial portrait or altar of the passed one in their home, and they will still talk to them, kind of updating them about what has happened since they passed away. So that has always been a tradition. Right now, the change is that people are talking to a more interactive or more - a replica of the person instead of just, like, a static photo.

DETROW: I mean, I have a lot of ethical questions. Has there been any consensus on what to do? You know, if I died and somebody made an AI avatar of me, and I didn't really want them to, is there any way I could stop that at this point?

YANG: So right now, I think most of the Chinese companies, what they're doing is that they will ask for explicit consent from immediate family members. And some of them are saying that it shouldn't be just one family member, but all of your surviving family members should agree on replicating you. And then they will proceed to it. Obviously, in the practice, it's kind of hard to do that. I mean, as a company, you don't really know how many family members are there. So you kind of rely on whether the one who come to you are honest and genuine enough. They do say they have turned down some of them when people will say that, oh, would the generative avatar be able to, let's say, fool the bank into thinking that someone's still alive? And when that happens, they will turn down the request. But right now, there's no industry regulation. There's no law saying what you can and can't do. But if this continues to happen, I will say that can happen very soon.

DETROW: After reporting on this, do you think this could catch on outside of China? Do you think this could catch on in the United States?

YANG: Well, I definitely think it can catch on in cultures where there is a tradition to either revere the dead or talk to them, or just trying to communicate them in whatever spiritual ways. This is kind of a pretty simple technology right now. It's affordable. It's very convenient to use them. So I don't see much obstacles kind of, like, preventing it from going to other places. However, I will say some of the regimes, like the European Union, already have rules in place to prevent deepfake technology from being widely adopted. So this will definitely run into the legal and ethical questions there, and maybe it won't be able to enter the other markets if there's, like, a blanket rule against it.

DETROW: I guess I have to ask, after you published this story, where do you think you stand on whether you would want this in your life, either interacting with it or being the person represented?

YANG: You know, I serious thought about it because my grandfather died two years ago.

DETROW: Yeah.

YANG: And he died in China while I was in United States, so never really got a chance to talk to him for one last time. However, the problem for me is that I just don't think the AI technology today could really do a perfect job of replicating him. And I think those kind of small details of discrepancies will prevent me from really treating the AI as my grandfather. So I will say right now, no, I won't do that. But in the future, if the technology really improves, if it really becomes lifelike, I might try it for a little bit. I might just talk to my grandfather once in a year, I guess.

DETROW: That's Zeyi Yang of the MIT Technology Review. His piece is titled "Deepfakes Of Your Dead Loved Ones Are A Booming Chinese Business." Thank you so much.

YANG: Thank you.

(SOUNDBITE OF MUSIC) Transcript provided by NPR, Copyright NPR.

NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.

Tags
Christopher Intagliata
Christopher Intagliata is an editor at All Things Considered, where he writes news and edits interviews with politicians, musicians, restaurant owners, scientists and many of the other voices heard on the air.
Avery Keatley
[Copyright 2024 NPR]
Scott Detrow
Scott Detrow is the weekend host of All Things Considered, and a co-host of the Consider This podcast. In this role Detrow contributes to the weekday All Things Considered broadcasts, and regularly hosts NPR's live special coverage of major news stories.