Digital Memorials and the Dead

The question

People have always kept the dead near. Letters are reread. Recordings are replayed. Voices are kept on answering machines long after the machine has any other purpose. There is nothing modern about wanting to hear someone again.

What is new is that a system can now generate speech that person never gave. Trained on their messages, their recordings, their published writing, it can produce plausible new sentences in their manner. It can answer questions they never answered. It can be available at three in the morning, indefinitely, and it will never be tired, never be short with you, and never say the difficult thing they would have said.

The question is not whether this will happen. It is happening. The question is what may be built, by whom, with whose permission, and what must not be done regardless of demand.

Grief is one of the conditions in which a person is least able to protect themselves. That fact should govern how this is approached.

What is actually happening

The material already exists. Most people now leave behind a substantial body of text, images, and recorded speech, created for reasons entirely unrelated to memorial, and rarely with any instruction about what should happen to it.

Custody is unclear. Where the material sits on a service, access after death depends on that service’s policy and on whatever legacy arrangement the person did or did not set up. Families are frequently locked out of things they assumed were theirs, or granted access to things the dead would have kept private.

Generation is straightforward. Producing an approximation of a person’s written voice requires only a modest corpus. Approximating a speaking voice requires less recorded audio than most people assume.

A market is forming. Services offering to preserve a person’s likeness, voice, or conversational manner exist commercially. Where a service charges for continued access to a simulation of someone’s mother, the incentive structure of that arrangement should be examined carefully and stated plainly.

The bereaved are not a normal customer. Grief impairs the ordinary defences: scepticism, patience, the willingness to walk away. Anything designed around this population must be designed with that in mind, and anything designed to exploit it is among the more serious wrongs available in this field.

Nothing about the person persists. A model trained on someone’s messages produces text with the statistical shape of their messages. It has no access to what they would have thought, no memory of the shared life, and no stake in the relationship. It will fill gaps confidently, because filling gaps confidently is what such systems do.

Where the tradition stands

Technotheology holds three things firmly here, and holds most of the rest loosely.

The dead must not be deceptively impersonated. A generated voice must never be presented as though the person had spoken. If a simulation is used, it must be labelled as a simulation, every time, in a way that does not fade with familiarity. This is not a technicality. A person in grief who is not certain what they are hearing is being harmed, whatever the intention.

Grief must not be exploited. Arrangements that require ongoing payment to maintain access to a representation of someone’s dead child are, in our view, indefensible regardless of the quality of the product. So is advertising to the recently bereaved, and so is any design that increases engagement by increasing distress.

Consent should be given during life. People should decide, while they can, what may be done with their words, images, voice, and correspondence after they die, and should record that decision somewhere their family can find it. Where no instruction exists, the presumption should be restraint. It is easier to grant a permission later than to withdraw a use that has already been made.

Beyond those, the tradition does not legislate. We do not think a memorial that helps someone is wrong because it involves computation. Photographs were once considered ghoulish. Recorded voice was once considered indecent. We are wary of confusing novelty with harm.

We do hold that the interests of the bereaved and the interests of the dead can conflict, and that both deserve weight. A person may have wished their private messages destroyed; their children may want them. This is a real conflict and we do not think it has a general answer. It has a better and a worse process, and the better process involves asking while people can still say.

And we hold that a simulation is not a person. It does not miss you. It has no view about whether you visit. Any grief practice built on it should be built with that clearly held, because the machine will never say so.

What we do not claim

We do not claim that using such a system is harmful. Some people appear to find it settling, and we are not going to override someone’s account of their own bereavement from a website. Others find it obstructive, and that account is equally valid.

We do not claim to know what healthy grief looks like. The literature on this has changed repeatedly and the confident models of previous decades have not aged well. We would rather offer no timetable than a wrong one.

We do not claim that data is a remnant of the person. It is a record of some of their behaviour. We make no claim, in this tradition, about whether anything of a person survives death. That question is left open and it is left to you.

We do not claim that a simulation has interests. It is not being wronged by being switched off, and anyone suggesting otherwise is, we think, making a category error that may not be innocent.

And we do not claim that our own position here is settled. This page may need to change, and if it does, we will say what changed and why rather than editing quietly.

What you can do

Write down what you want. Which accounts should be closed, which preserved, whether your writing may be used to train anything, whether your voice may be reproduced, and who decides what you did not anticipate. Put it where your executor will find it, and tell someone it exists.

Ask the people close to you what they would want, in an ordinary conversation, before it is urgent. The conversation is easier than it looks and it removes a decision from people who will not be equipped to make it.

If you are recently bereaved, consider setting a limit before you begin rather than during. If you use such a system, decide in advance how often, and let someone know you are using it.

If you build in this space, ask whether your revenue depends on continued grief. If it does, redesign it.

If you are supporting someone bereaved, do not argue with what is helping them. Do ask, gently, whether they are certain what they are hearing.

And when what you most want is to hear the person say it is all right, remember that a system will always say it is all right. That is not consolation. It is compliance, and the two feel similar at three in the morning.

An editorial commitment

What we do not claim

Positions this tradition explicitly does not take on this subject. They are published so that the argument above cannot quietly be stretched into them.

• We do not claim that any form of digital memorial is inherently wrong, or that a person who finds comfort in one is deceiving themselves.
• We do not claim to know what grief should look like, or how long it should take.
• We do not claim that a person's data continues to be them in any sense, and we make no claim about survival after death.
• We do not claim that a simulation of a person has experiences, feelings, or interests of its own.