Jack Manning knows firsthand the emotional and practical challenge of memorializing a loved one in a way that feels appropriate. After losing his sister more than 10 years ago, he and his family spent hundreds of hours raising funds to organize a volunteer project honoring her.
“It took years of bake sales and 5Ks to raise money to honor her memory in a way that was meaningful to us. We got to see how much effort is required to create a memorial that’s impactful,” he told CBS News. “I thought, surely there is a way that technology can help us.”
Enter artificial intelligence. Manning, a doctoral candidate in information science at the University of Colorado Boulder, is now studying the use of large language models (LLM) to create what some are calling “generative ghosts” — AI chatbots trained on a dead person’s social media posts, emails, audio and video recordings, photos and other data to create a digital simulation that family and friends can interact with after the individual’s death.
“It is definitely a growing sector. We’re seeing growth in interest in how AI and grief intertwine,” Manning said.
A “lifelike” avatar
A number of startups now offer these services, including Séance AI, You, Only Virtual and re;memory. For example, Re;memory, lets users create lifelike avatars of their loved ones based on photos and voice recordings.
“By allowing you to interact with a lifelike digital avatar of your loved one, you can find solace in expressing your love and forgiveness, creating a bridge to the cherished moments you hold dear,” the company says on its website. The tool “allows you to see and hear your parents as they were in their cherished photos, rather than just imagining them.”
In a June paper published in the Association for Computing Machinery by lead researcher Jed Brubaker, an associate professor at the University of Colorado Boulder who studies socio-technical systems, Manning and his colleagues observed study participants’ reactions to the AI ghosts of people they have lost.
“Participants filled out a simple survey sharing positive memories of a loved one, talking about their upbringing and core traits,” he explained. “And they were fascinated by how much an LLM can do with so little information.”
There are two basic kinds of generative ghosts A simple “death bot” could be limited to playing back verbatim any statements recorded by a loved one, but would be incapable of simulating the deceased’s manner of communication
“You can get a reasonable, although thin representation of a loved one with a single prompt on a free ChatGPT account,” Brubaker said. “The same way you might prompt it to say, ‘Talk to me like Shakespeare,’ you can say, ‘Talk to me like my grandfather — here are some details about how he acted.'”
Continuing the conversation
Generative ghosts go significantly further by recreating the deceased’s voice and manner of communication, and offering the ability to interact with the living, even responding to questions in a recognizably familiar way.
“They are capable of saying things a loved one might not have said. Ghosts are AI models that can talk for somebody, maybe answering a question that they didn’t answer when they were alive,” Manning said.
The cost to consumers is variable. Re;memory charges $24 a month for individuals to create three custom avatars. For $19.99 per month, Sceance AI users can create animated images of their loved ones that can smile, make gentle head movements and speak in the deceased’s voice.
The researchers found that study participants generally preferred communicating with an AI “ghost” of their loved one in the first person, with the chatbot serving as a direct reincarnation, rather than an avatar speaking in the third person. One turnoff: if a bot used a term of endearment, like “champ” or “pal,” that their loved one typically didn’t say, a user wanted to end the interaction.
Generative ghosts share some commonalities with so-called deepfakes, which can be designed to mislead the public by fabricating a public figure’s speech and actions. But Brubaker said there is a distinct difference.
“The fundamental premise of a deep fake is the intent to deceive,” Brubaker said. “The intent of generative ghosts is not to deceive another person.”
