The ethics of AI photos of deceased loved ones: a 2026 framework
A principled framework for thinking about AI photography of deceased loved ones — consent, dignity, minors, simulated speech, content credentials, and the line between honoring memory and exploiting grief.
By Jiuhong Deng · · Updated
AI photography of deceased loved ones is one of the most ethically loaded uses of consumer AI in 2026. This piece lays out a principled framework for thinking about it — both as a user evaluating tools and as a developer building them.
This isn’t an abstract treatise. It’s the framework we use at Lover Snap, published as /memorial/ethics and applied as actual product rules. It’s also a useful evaluation test you can apply to any other AI memorial tool: count how many of these six commitments they publish and enforce.
The six commitments
1. Consent comes first
When someone uploads a photo of another person to an AI tool, they are making an implicit claim: “I am the appropriate steward of this person’s likeness.” For deceased loved ones, this typically means being a family member or close relation with the moral authority to act on their behalf.
No AI tool can verify this perfectly. But responsible tools require an explicit attestation and provide a clear takedown channel for family members who later object to a particular character or generation.
What this rules out: Anonymous upload of celebrity or public figure photos. Anonymous upload of photos of people without any family relationship.
2. No minors
This is the single most important commitment, and the one most likely to distinguish a responsible tool from an irresponsible one. AI photography of minors — living or deceased — generates training data, model artifacts, and downstream content that can be misused in ways the original generator cannot predict or control.
Responsible tools refuse generation of new AI photos of anyone under 18, full stop. Restoration of existing family photographs that include minors is a separate (and more limited) capability that some tools support, but new generation is universally refused by reputable tools in 2026.
What this rules out: Aged-up portraits of deceased children. New scenes featuring children. Anything that could be characterized as “synthetic child sexual abuse material” even at the most ambiguous edge case.
3. No simulated speech
Simulated speech of the deceased — voice cloning, dialogue generation, conversational chatbot impersonation — is a different category of use than photographic remembrance. A photograph captures a moment; a simulated voice claims to speak for someone who can no longer speak.
Responsible tools either don’t offer it at all (Lover Snap’s position) or offer it only with extraordinary safeguards (explicit pre-death consent recordings, time-limited, family-approved). The default should be refusal.
What this rules out: AI-generated audio of a deceased person. AI-generated video with simulated dialogue. Conversational chatbots that impersonate a deceased family member without explicit pre-death consent.
4. Private by default
Reference photos and trained AI characters of deceased loved ones should be:
- Encrypted at rest and in transit.
- Scoped to a single user account.
- Never used to train shared or public models.
- Deletable on request.
- Permanently removed within 30 days of account closure.
This isn’t optional. The dignity of the deceased depends on their likeness not being weaponized in training data for downstream uses no one can predict.
What this rules out: Tools that use reference photos to train improvements to their public models. Tools that share or sell reference photos to AI training data providers. Tools that retain photos indefinitely after deletion requests.
5. Honest representation
Generated images should be identifiable as AI-generated through embedded content credentials (C2PA) and through clear marketing language that doesn’t claim the technology “brings someone back.”
Honesty is part of dignity. A photograph marketed as a remembrance is different from a photograph marketed as resurrection — and the difference matters both for the user and for the broader cultural position of the technology.
What this rules out: Marketing that claims AI photos let you “see your loved one alive again.” Removal of content credentials from generated images. Tools that intentionally make AI-generated photos indistinguishable from photographs of the actual person at the time of their death.
6. Grief support, not replacement
AI memorial photos are a tool, not therapy. Responsible tools link to grief support resources, don’t market themselves as therapeutic, and don’t structure their pricing or features in ways that encourage compulsive use.
What this rules out: Engagement-optimized products that maximize generation frequency over user wellbeing. Tools without any reference to grief support. Tools that market the experience as “healing.”
How to use this framework
If you’re evaluating an AI memorial tool, the easiest test is: does the tool publish all six? Lover Snap does at /memorial/ethics. MyHeritage Deep Nostalgia partially does. Most newer entrants in 2026 do not yet.
The presence of a published framework doesn’t guarantee a tool will live up to it, but its absence is a useful signal. Tools that haven’t thought through these questions are tools that might solve them in ways you wouldn’t like.
The case for AI memorial photography
Despite the ethical loading, there’s a strong case for the technology done well. Grief researchers broadly support continuing-bonds approaches — the integration of the lost person into your ongoing internal life. Photographs have always played a role in that. AI just extends what photographs can do: fill in the milestones that didn’t get photographed, complete albums that feel incomplete, give families memorial artifacts they couldn’t otherwise have.
The technology isn’t inherently exploitative. It depends entirely on what the people building it decide it can and can’t do. That’s the case for publishing a framework and enforcing it.