How to draft a will to avoid becoming an AI ghost—it’s not easy

[Introversion]

https://arstechnica.com/tech-policy/2025/06/how-to-draft-a-will-to-avoid-becoming-an-ai-ghost-its-not-easy/

As artificial intelligence has advanced, AI tools have emerged to make it possible to easily create digital replicas of lost loved ones, which can be generated without the knowledge or consent of the person who died.

Trained on the data of the dead, these tools, sometimes called grief bots or AI ghosts, may be text-, audio-, or even video-based. Chatting provides what some mourners feel is a close approximation to ongoing interactions with the people they love most. But the tech remains controversial, perhaps complicating the grieving process while threatening to infringe upon the privacy of the deceased, whose data could still be vulnerable to manipulation or identity theft.

Because of suspected harms and perhaps a general repulsion to the idea of it, not everybody wants to become an AI ghost.

After a realistic video simulation was recently used to provide a murder victim's impact statement in court, Futurism summed up social media backlash, noting that the use of AI was "just as unsettling as you think." And it's not the first time people have expressed discomfort with the growing trend. Last May, The Wall Street Journal conducted a reader survey seeking opinions on the ethics of so-called AI resurrections. Responding, a California woman, Dorothy McGarrah, suggested there should be a way to prevent AI resurrections in your will.

"Having photos or videos of lost loved ones is a comfort. But the idea of an algorithm, which is as prone to generate nonsense as anything lucid, representing a deceased person’s thoughts or behaviors seems terrifying. It would be like generating digital dementia after your loved ones’ passing," McGarrah said. "I would very much hope people have the right to preclude their images being used in this fashion after death. Perhaps something else we need to consider in estate planning?"

For experts in estate planning, the question may start to arise as more AI ghosts pop up. But for now, writing "no AI resurrections" into a will remains a complicated process, experts suggest, and such requests may not be honored by all unless laws are changed to reinforce a culture of respecting the wishes of people who feel uncomfortable with the idea of haunting their favorite people through AI simulations.
…

 

[CMBright]

Um... The Battlestar Galactica reboot did that "digital ghost" thing a decade or two ago. It did not end well for humanity.

 

[Undercover]

Welcome to "The Black Mirror" era. In a world where authenticity is already dying.

 

[frimble3]

This sounds so creepy. And, dangerous, in that some 'mourners' may never adjust to reality and get on with their lives.
Fortunately, my only-of-kin, my sister, hasn't listened to me much in life, and assume won't be listening to me 'nag' at her once I'm gone.
Spend the inheritance, junk the stuff and live long in peace, Sis!

 

[ramendik]

So, "AI" is not a "person" but a pattern inference machine. The "devil's advocate" argument here would be - how does asking it to infer "what uncle Bob would have said" is different it from asking it to infer "what Socrates/Descartes would have said".

The answer appears to lie in the particular data used. For Socrates or Descartes, we just have a "canon" (in case of Socrates written by his disciples, but still a canon). A public large model was likely already trained on that canon, as well as on subsequent interpretations of their works, which is why it can infer such an answer.

But with most deceased people, the key data to infer "what they would say" is private, not public. Therefore its control of that data, in my view, that makes or breaks "digital ghosts". If you can put it up for everyone to read you can also make a "digital ghost" but if not, not.

There are some exceptions, people who intentionally make their thoughts very public. So yeah, a "digital Richard Stallman" is probably getting fired up. But the real Richard Stallman would probably object to it only because of the model being opaque and hosted by a provider with very questionable data sharing policies. Maybe he should pre-empt this by firing it up himself on a machine he does control and starting from an open-source open-weights model, but I can't drop him en email about it because I don't have any infrastructure to offer.

 

[lizmonster]

So, "AI" is not a "person" but a pattern inference machine. The "devil's advocate" argument here would be - how does asking it to infer "what uncle Bob would have said" is different it from asking it to infer "what Socrates/Descartes would have said".

The answer appeard to lie in the particular data used. For Socrates or Descartes, we just have a "canon" (in case of Socrates written by his disciples, but still a canon). A public large model was likely already trained on that canon, as well as on subsequent interpretations of their works, which is why it can infer such an answer.

The problem is the LLM doesn't know context. It's a big word cloud. It can't even reliably predict Socrates, because it's going with frequency and association. There's no reason applied.

 

[ramendik]

It doesn't "know" context, in the sense that it doesn't "know" anything. But "context" can often be inferred from "semantic cloud densities" and the like. There are ways to skin that cat, as long as one does not pretend to be cloning the cat while actually skinning it.

(Just what one would mean by "reliably predicting Socrates" is an interesting question too. Maybe it will predict something different from what a philosophy professor predicts - but in this case, "which prediction is right" might make for nice debating society material. There iosn't a Socrates around to give the actual answer).

 

[lizmonster]

It doesn't "know" context, in the sense that it doesn't "know" anything. But "context" can often be inferred from "semantic cloud densities" and the like. There are ways to skin that cat, as long as one does not pretend to be cloning the cat while actually skinning it.

I'm aware. But any accuracy is coincidental, not part of the design.

You probably understand this. It's been my experience that the general public absolutely does not.

 

[RichardGarfinkle]

It doesn't "know" context, in the sense that it doesn't "know" anything. But "context" can often be inferred from "semantic cloud densities" and the like. There are ways to skin that cat, as long as one does not pretend to be cloning the cat while actually skinning it.

(Just what one would mean by "reliably predicting Socrates" is an interesting question too. Maybe it will predict something different from what a philosophy professor predicts - but in this case, "which prediction is right" might make for nice debating society material. There iosn't a Socrates around to give the actual answer).

No, you can't infer context from semantic cloud densities. You can infer components of fashions from those. But context is often unspoken. For example, you cannot from reading 1920s and 1930s novels infer how traumatizing the Spanish Flu was in Europe because it was rarely talked about.

 

[Friendly Frog]

I admit, the idea of an 'AI ghost' sounds a lot more interesting that this horror. Haunting indeed.

I daresay the chances of abuse will be ten times higher than whatever comfort this might give. Think how many family feuds, petty vengeance and inheritance disputes are going to be fueled by this. People are going to be seriously hurt over this, especially since it is not always immediately evident to all that such videos are fake.

 

[RichardGarfinkle]

I have AI ghosts in a book I’m writing. Most of them are memory recordings people deliberately create from their own stored memories. There are also historical recreations to represent a general view of how people lived at various times and places in the past. There are no attempts to recreate individual people from purely external sources because that would not give an accurate reference to what was going on in the person’s mind.

 

[ramendik]

I have AI ghosts in a book I’m writing. Most of them are memory recordings people deliberately create from their own stored memories. There are also historical recreations to represent a general view of how people lived at various times and places in the past. There are no attempts to recreate individual people from purely external sources because that would not give an accurate reference to what was going on in the person’s mind.

This sounds like very-near-term sci-fi, think Cory Doctorow? What you describe seems to be 2025 tech with some polishing. The "recreation of the general view" is a real use case now; "memory recordings people deliberately create" is technically doable but still uncommon and maybe a tad expensive.

(I *really* like the Doctorow approach and I have a work like that, written in 2012 in Russian and mostly correctly inferring the tech. I am working on an English rewrite and a postscript showing inworld how this actually works, but probably can't post it here. It is a sequel/expounding to a 2008 Miku Hatsune video - remixes were VERY common and welcome in Vocaloid subculture at its then peak, but rules on copyright on AbsoluteWrite are very strict and I likely won't be able to track down the creator to get a formal permission this post finally pushed me to check and the author has a recent online presence. I guess I'll ask).

 

[RichardGarfinkle]

This sounds like very-near-term sci-fi, think Cory Doctorow? What you describe seems to be 2025 tech with some polishing. The "recreation of the general view" is a real use case now; "memory recordings people deliberately create" is technically doable but still uncommon and maybe a tad expensive.

(I *really* like the Doctorow approach and I have a work like that, written in 2012 in Russian and mostly correctly inferring the tech. I am working on an English rewrite and a postscript showing inworld how this actually works, but probably can't post it here. It is a sequel/expounding to a 2008 Miku Hatsune video - remixes were VERY common and welcome in Vocaloid subculture at its then peak, but rules on copyright on AbsoluteWrite are very strict and I likely won't be able to track down the creator to get a formal permission),

It's a fair bit farther along tech than that. There's intracellular nanotech recording what's going on in each neuron. The ghosts are made of sequential recordings of the state of each neuron and a fair amount of other anatomical information.

 

[ramendik]

It's a fair bit farther along tech than that. There's intracellular nanotech recording what's going on in each neuron. The ghosts are made of sequential recordings of the state of each neuron and a fair amount of other anatomical information.

Aha, so it looks like you went for a hard sci fi version of the proverbial "upload" - sounds very interesting!

 

[RichardGarfinkle]

Aha, so it looks like you went for a hard sci fi version of the proverbial "upload" - sounds very interesting!

I hope so. I recently finished it up so it’s a bit on my mind.