Who gets to speak for someone who has died?

A familiar voice can offer comfort without settling whose account of the past it represents. AI recreations make that distinction especially important: keeping someone’s words and letting a service produce new words in their name are different decisions.

A family recording has an edge. The person finishes speaking; somebody reaches for the camera; the room goes quiet. A conversational recreation offers to continue. You can ask the question you forgot to ask, or a question that only became possible after the person died. The attraction is easy to understand. The difficulty begins when an answer that sounds familiar acquires the standing of something the person actually said.

A useful starting point is to ask what kind of material you are receiving. Is this an original recording, a quotation selected from an archive, or a newly generated response? Then ask who authorized that particular use. Those questions leave room for comfort while keeping the record of a life distinct from the service speaking about it.

A recording, a quotation, or a new sentence

An original recording preserves an occasion, though it may still be edited or missing context. A retrieved quotation selects words from an existing source. A generated answer creates a new response. The same interface could present all three, and a convincing voice would not tell the listener which operation had occurred.

One technical approach, retrieval-augmented generation, supplies a language model with material retrieved from a collection. That can make relevant documents available during generation; it does not make the resulting response an original statement by the person represented. Lewis and colleagues’ foundational RAG paper concerns language tasks, not the recovery of a deceased person’s lived experience. Lewis et al., 2020

Imagine asking an archive what your father thought of your first job. A letter might contain his actual words. A service might find that letter and quote it. It might instead combine several letters into a sympathetic answer addressed to you now. Each output could be meaningful, but each needs a different description. “He wrote this in 1998” supplies a kind of evidence that “this sounds like him” cannot supply.

The sensible demand is not that a simulation never generate anything. Generation is part of the proposed experience. It is that the service make clear when it has moved beyond the record. A sentence can be welcome and still deserve a label explaining where it came from.

More than one person is present in the archive

The deceased person is the obvious subject. Other people enter through the material: a friend’s correspondence, a sibling’s account of an argument, a child’s face in a video. The person uploading the files may also be the person paying for the service. Neither role automatically makes them the only person whose interests matter.

Hollanek and Nowaczyk-Basińska distinguish the person whose data informs a simulation, the person holding that data, and the person expected to interact with the result. Their ethical analysis recommends attention to mutual consent, transparency, and how a simulation can be retired. It develops speculative scenarios; it is not a clinical trial establishing what griefbots do to bereavement. Research paper, May 2024

Consider a brother who wants to hear his mother’s stories and a sister who does not want messages in her mother’s voice. Their wishes need not cancel each other. A design could allow private listening while preventing unsolicited contact. But a service that assumes every family member is an intended recipient has already made a decision about their relationship.

The archive itself may contain disagreements. One relative remembers a holiday as a rescue from a difficult year; another remembers being left behind. A fluent account does not reconcile those experiences merely by choosing the version that produces the most reassuring conversation. It may simply make one person’s account easier to repeat.

Permission needs an object

“Use my data” is too broad to resolve every later choice. A person might welcome the preservation of an interview and dislike the idea of their voice answering new questions. They might allow their children to hear it but object to public circulation. They might be comfortable with family stories and unwilling to appear to endorse a financial decision, a political cause, or a quarrel they never witnessed.

For discussion, specify the act: preserving, quoting, generating, sharing, or initiating contact. Then identify the intended audience and the person able to change the arrangement. This is an ethical way to make the choices legible, not a claim that one permission form resolves every legal issue. Rights, contracts, and rules differ; a family conversation should not be mistaken for jurisdiction-specific advice.

It is also worth asking whether refusal remains practical after the service begins. Can a recipient stop messages without deleting another person’s archive? Can a participant remove a private letter without losing access to recordings they value? A single button marked “manage legacy” may conceal several decisions that deserve separate answers.

Five decisions hidden inside one upload

The following is an invented discussion exercise. A family has a folder containing interviews, photographs, and correspondence. A service offers to turn it into conversations with a deceased relative. Before selecting “continue,” write an answer to each question. If the answer is unknown, leave it unknown rather than filling the gap with what seems likely.

1. Whose material is being uploaded? Identify the represented person and anyone else whose words, likeness, or private information appears. Separate a record created for public circulation from one shared within a relationship.

2. What may the service do with it? Decide whether the arrangement permits playback, search, quotations, newly generated speech, or some combination. Identify how a recipient would tell them apart.

3. Who has agreed to receive the result? An invitation to try a service is different from an incoming message already speaking in a parent’s voice. Describe a way to decline that does not require explaining the refusal to the whole family.

4. How will disagreement remain visible? Suppose two accounts conflict. Decide whether the service will show both, identify the source it used, or say that the record does not settle the question. Avoid making a comfort setting the final editor of family history.

5. Who can pause or end the arrangement? Include what happens to original files, generated conversations, and access if a subscription stops or a provider closes. Preserving an archive and continuing a simulation need not be the same decision.

Now change one fact. The person who organized the service dies, or a relative discovers a private letter they did not know had been included. Which of your answers still holds? Which requires another person’s agreement? The exercise is useful when it reveals an unfinished conversation; it does not need to produce unanimous permission.

When the service changes

Suppose the original arrangement promised a searchable collection of recordings. Later, the provider offers automatic messages on anniversaries. That may be a welcome feature for one recipient and an unwelcome intrusion for another. The earlier decision to preserve recordings does not, by itself, answer whether the new feature should be enabled.

A family should be able to describe the change in ordinary language. What will the service do now that it could not do before? Who will receive something new? Which records will it use? Who can decline? These questions are more useful than treating every update as either a betrayal or an improvement.

There is a similar distinction between correcting an archive and making a conversation more agreeable. If a date is wrong, correcting it can improve the record. If a disagreement is uncomfortable, removing it changes what the record allows a person to encounter. The family may choose privacy or restraint, but the choice should remain recognizable as a choice.

Comfort and uncertainty can coexist

A person may value a simulation while understanding that it generates responses. Another may prefer an imperfect recording, with the interrupted sentence and the noise from the next room. This guide offers no evidence-based reason to prescribe one experience for every bereaved person. Its concern is whether the arrangement makes the relevant choices visible and gives the affected people a workable way to participate or decline.

The question to carry away is narrower than whether technology should have a place in remembrance. When the familiar voice says something new, what should the listener understand about that sentence, and who should have had a say in its creation?

What a change to the archive should change

Imagine that a relative removes a private letter from the collection. There are several possible effects: the service stops retrieving it, an old generated summary remains, and a saved preference derived from it still influences later replies. This is an invented configuration, not a description of all griefbot products. It shows why the next question is “where was the information used?” rather than simply “is the file gone?”

The AI memory guide separates conversation context, saved records, retrieved material and trained model information. The forgetting question explains why changing one component does not establish what happened to the others.

For this family, a useful response would identify which outputs were based on the letter and what can be corrected or removed. That may matter as much as deleting the original. If a generated account has already been shared, the recipient needs to know that its source has been withdrawn or disputed. No technical edit can make them unhear it.

This adds another dimension to consent: the ability to reconsider should have a practical description. It should also preserve the distinction between withdrawing access to private material and proving that an unwanted account of the past is false. Neither act automatically establishes the other.

Two conversations to explore

The Human Podcast’s interview with Katarzyna Nowaczyk-Basińska and Tomasz Hollanek introduces the researchers behind the ethical analysis discussed here. Waking Up With AI’s episode on griefbots approaches the subject through professional discussion of rights, third parties, and the service relationship. They provide different routes into the question; neither establishes what a particular service will do for a particular person.

See the listening notes “Will We Be ‘Haunted’ By AI Chatbots Of Dead Loved Ones?” and “Griefbots and the Digital Afterlife.” The notes identify the material reviewed and the limits of each selection.

Read the dilemma as fiction

Her mother’s reconstruction has a comforting account of the past. Her brother remembers something else. The Last Person Who Remembers is an eleven-minute standalone story about who gets to tell a family’s history. It is non-canon fiction, with separate notes identifying the real technology and invented service.

For a related question about remembered preferences becoming new instructions, continue with Your Mother Would Have Wanted This. If the question of who may decide for someone else draws you onward, read the open sample of Sovereign Mercy.

Sources and reading context

The family examples and five-decision exercise are original, invented illustrations. Recommendations about making permissions and disagreements visible are editorial judgments. The linked research supplies technical or ethical context; the stories are not evidence of real-world outcomes. This page discusses bereavement and family disagreement without graphic material.

In the glossary: Griefbot and deadbot; Digital legacy; Retrieval-augmented generation; Consent.

Return to the broader AI companionship, loneliness and grief guide or use the reading-group discussion guide.