The Ethics of AI After Death: Consent, Accuracy, Privacy & Grief

We build this technology, which makes it more important to write this section carefully rather than less. Anyone in this field telling you the ethics are straightforward is selling something.

Consent

Consent to record is not consent to represent

Someone who happily records stories for their grandchildren has not thereby agreed to have a conversational persona built from those recordings, or to have new sentences generated in their synthesised voice. These are separate uses with different implications and they warrant separate, explicit permissions.

Posthumous construction is a harder case

Building a persona from someone who died without expressing a view is where most of the real difficulty lies. The available material — messages, videos, social posts — was created for other purposes entirely. Family belief about what someone would have wanted is sincere and also unreliable, particularly when grief is acute. A defensible standard: rely on documented material, avoid synthesising a voice without prior consent, keep the representation clearly labelled as a reconstruction, and prefer restraint where there is doubt.

The right to be forgotten, and to stay silent

Some people do not want to persist in any form, and that is a legitimate preference deserving as much weight as the opposite one. Equally, a person may want an archive that deliberately omits chapters. An archive should be able to hold an intentional silence rather than treating every gap as something to fill.

Accuracy and the invention problem

This is the ethical issue most specific to AI, as opposed to memorialisation generally. A generative model asked something outside its material will produce a fluent answer in the person's manner. The output is plausible and unfounded, and a grieving family is not well placed to tell the difference.

The harm is not abstract. An invented opinion can settle a family argument wrongly, misrepresent someone's beliefs, or attribute to them an approval or disapproval they never expressed. The obligations that follow:

• Ground answers in real material rather than generating freely, and retrieve before responding.
• Keep documented statements distinguishable from inference. A reader should be able to tell which is which.
• Decline rather than fill gaps. "They never spoke about that" is the correct answer far more often than products admit.
• Never fabricate memories, quotes, or positions for emotional effect, however well it performs.

The technical background is covered in can AI recreate a personality?

Privacy — including the privacy of others

A personal archive is rarely only about one person. Stories name relatives, describe conflicts, and disclose things others might not want repeated. Living people mentioned in an archive have interests, and in some jurisdictions rights, that outlast the person who recorded the material.

• Default to private. Public sharing should be a deliberate act, not a setting someone forgot to change.
• Allow material to be restricted to specific people or sealed until a date.
• Treat identity data, health details, and family conflict as sensitive by default.
• Expect that a living person may reasonably object to a story about them being publicly accessible.

Who controls it

Family disagreement about a shared parent's archive is not an edge case; it is the normal case. One sibling finds it comforting, another finds it disturbing, and a third wants it public. Deciding stewardship in advance — while the person can state a preference — prevents most of it. Where that has not happened, the safer default is narrow access and reversible decisions.

Commercial control matters too. If a company can change access rules, raise prices, or shut down, the family's control is conditional. Export capability is an ethical feature, not just a convenience.

Grief

We are deliberately careful here. Responses to interacting with a representation of someone who has died vary substantially between individuals, and the research base is early and mixed. Some people describe it as steadying, similar to rereading letters. Others describe interaction keeping a loss immediate rather than easing it. We are not aware of evidence that supports confident claims in either direction, and we will not make medical or therapeutic claims.

What can be said responsibly: this technology should not be designed to maximise engagement. Streaks, nudges, and re-engagement prompts are inappropriate applied to a dead person. Access should be easy to reduce or stop. Each family member should decide their own level of contact without pressure, and choosing not to use it at all should be entirely unremarkable. If grief is interfering with daily life, that warrants support from a qualified professional — not a product.

Commercial ethics

• No fear-based marketing. Urgency built on mortality is manipulation.
• No overstated claims. Nothing about restored consciousness, and no promises of permanence a company cannot keep.
• No hostage pricing. Access to someone's own memories should not be leverage. Export should always work.
• Clear limits, stated upfront. What the system cannot do belongs in the marketing, not the footnotes.

Where Eterna stands

These are the commitments we hold ourselves to: personas reflect documented material rather than invented personality; content is private by default; voice features require explicit consent and are access-controlled; impersonation and unauthorised cloning are prohibited; and material is exportable. We describe our personas as representations, not as the person.

Our Privacy Policy, Terms of Service, and Acceptable Use Policy set the formal version of this out, and about Eterna covers the principles behind it. If you think we have got something wrong here, tell us — this is a field where being corrected is more useful than being confident.