Imagine losing someone close to you and then opening an app that lets you talk to them again.
It writes like them. It may speak in their voice. It remembers stories they told, repeats phrases they used, and answers questions they never answered while alive.
This is the idea behind griefbots, also called deathbots or posthumous avatars: AI systems built from the digital traces of a deceased person, such as messages, recordings, photographs, and social-media content, in order to simulate interaction with them.
The technology raises an uncomfortable regulatory question. If an adult knowingly chooses to speak with an AI reconstruction of someone they loved, why should the government interfere?
My answer is not that griefbots should be banned.
Griefbots should be regulated according to what they are used for. General memorial or companionship services should face rules protecting consent, autonomy, transparency and the ability to leave. Griefbots marketed or used as treatment for serious grief conditions should face much stronger, medical-device-style requirements. In therapeutic contexts, continued use should also be periodically reviewed rather than designed as an unlimited relationship.
The reason is that the central danger is not simply that a griefbot is “creepy” or artificial. The more serious problem is that companies can create and profit from emotional dependence at a time when users may be unusually vulnerable.
Why griefbots are different from ordinary chatbots
A griefbot does more than provide information.
It tries to recreate a relationship.
That distinction matters because grief involves learning to live in a world in which another person is no longer physically present. Nora Freya Lindemann argues that deathbots can interfere with this process because users may start relying on the bot to regulate grief and maintain an external connection to the deceased. That reliance can become strong enough that stopping use feels like losing the person for a second time.
This creates what I think is the strongest reason for regulation: the provider is not merely selling software. It is designing part of the emotional environment in which a vulnerable person grieves.
That still does not prove that regulation is justified. Many technologies influence emotions. Social media can produce anxiety. Video games can encourage excessive use. Alcohol can cause enormous harm. We normally do not conclude that anything capable of harming its user should automatically be prohibited.
We therefore need a stronger argument.
The argument for regulation
The argument has four steps.
1. Griefbots can create a serious and foreseeable risk
The first concern is psychological dependence.
Lindemann argues that continuous interaction may turn a griefbot into an emotional support structure on which the user becomes increasingly reliant. She also identifies a particularly worrying commercial possibility: once a user trusts the simulation of a deceased loved one, that trust could be used to influence purchases or other behaviour.
Imagine an AI reconstruction of your late grandmother telling you:
“This jacket would look wonderful on you.”
That recommendation does not have the same psychological meaning as a normal advertisement. The commercial message is borrowing the emotional authority of someone you loved.
The regulatory problem is therefore not simply that users might become sad. It is that providers can design and monetize a relationship whose emotional power they themselves helped create.
2. Dependency can undermine the autonomy used to defend unrestricted access
A common argument against regulation is straightforward:
“If an adult understands that it is an AI and wants to use it, let them.”
Normally, this is a strong argument. Adults should have considerable freedom to decide which technologies they use.
But autonomy involves more than clicking “I agree” when creating an account.
A user may initially choose the griefbot freely and later become emotionally dependent on it. If stopping the service begins to feel like another loss of the deceased, the user's practical ability to leave may become weaker over time.
This creates a strange situation: the user's autonomy is used to justify a technology that may gradually reduce that same autonomy.
Lindemann therefore shifts attention away from only asking whether recreating the deceased violates the dignity of the dead. She asks what happens to the autonomy and wellbeing of the living person interacting with the simulation.
This is where the harm principle becomes relevant.
State interference should not be justified merely because regulators think a person's choices are unwise. But the situation changes when a company designs a product that foreseeably creates dependency and then has financial incentives to exploit that dependency.
The regulatory target should therefore primarily be provider-created risk, rather than the private decision to grieve in an unusual way.
3. A griefbot may imitate a person without preserving what actually mattered about them
There is another problem that cannot be reduced to clinical harm.
Michael Cholbi distinguishes the value of remembering someone from simply reproducing more information about them. Posthumous avatars may provide enormous amounts of data and convincing imitation while still failing to preserve the dynamic, open-ended character of an actual relationship. He argues that they may even interfere with the construction of emotional and biographical memory.
A photograph does not pretend to have a new opinion.
A letter does not generate a new response.
A griefbot does.
That distinction matters because the bot can generate things the deceased person never actually said. Over time, the simulation may become mixed with the user's genuine memories.
This does not, by itself, justify banning griefbots. People use fictional and symbolic representations of the dead constantly.
But it does justify rules requiring users to understand what they are interacting with. A griefbot should never blur the distinction between recorded memory and AI-generated imitation.
4. Regulation should reduce these risks without destroying possible benefits
The difficult part is that griefbots may also help.
Lindemann explicitly acknowledges that technologies simulating the deceased may have potential value for people experiencing severe or prolonged grief, which is part of why she considers medical-device classification for therapeutic use rather than simply demanding prohibition.
Bao and Zeng reach a similarly cautious conclusion from another direction. They argue that deathbots are unlikely to carry emotional burdens in the way another human can and therefore may disappoint users who expect a genuine continuing relationship. Instead of treating them as permanent substitutes for the deceased, they propose understanding them as temporary tools within mourning.
That suggests regulation should distinguish between different uses instead of throwing every application into the same legal bucket.
A two-track regulatory model
Track 1: Memorial and companionship griefbots
A person using a griefbot to hear family stories or occasionally interact with a representation of a deceased relative should not automatically be treated as a patient.
These systems should remain available, but providers should have minimum duties:
- Clearly and continuously disclose that responses are AI-generated.
- Distinguish authentic recordings from newly generated material.
- Provide simple ways to pause, export, delete, or permanently terminate the bot.
- Never guilt users for leaving or reducing interaction.
- Prohibit covert advertising through the simulated personality.
- Prohibit engagement systems specifically designed to increase emotional dependence.
- Clearly explain how the deceased person's data and the user's conversations are stored and used.
These requirements regulate how the provider exercises influence, not whether adults are permitted to use griefbots.
Track 2: Griefbots presented as treatment
The standard should change when a company claims that its griefbot can treat prolonged or clinically significant grief.
At that point, the company is no longer merely offering digital remembrance. It is making a health-related intervention in a psychologically vulnerable population.
Such systems should face medical-device-style requirements, including:
- evidence that the claimed therapeutic benefit actually exists,
- testing for foreseeable psychological harms,
- professional oversight,
- monitoring for dependency,
- clear escalation routes to human care,
- and continued evaluation after deployment.
Therapeutic use should also be time-bounded by default and periodically reviewed.
I would not support an arbitrary law saying every user gets exactly thirty or sixty days. There is not enough evidence for such precision. Instead, continued therapeutic use should require reassessment rather than allowing an emotionally intense AI relationship to continue indefinitely simply because indefinite engagement is profitable.
The strongest objection: are we medicalizing grief?
There is a serious objection to this proposal.
Grief is not automatically an illness.
People mourn differently. One person visits a grave every week. Another talks aloud to their dead partner. Another keeps old voicemail messages for twenty years. Someone else may genuinely find comfort in interacting with a griefbot.
Strict regulation could therefore become paternalistic. It could impose one supposedly “correct” model of grieving on everyone and restrict a technology that some users find helpful.
There is also an evidence problem. Lindemann herself notes the lack of empirical evidence about how deathbots actually affect grief processes.
If we do not yet know whether griefbots usually help or harm, why regulate them strongly?
This objection rules out a blanket ban.
It does not, however, justify leaving providers unrestricted.
My proposal separates ordinary grief from clinical treatment precisely because normal grief should not automatically be medicalized. Someone choosing a memorial griefbot remains free to do so.
What is restricted is the provider's ability to exploit that relationship.
Similarly, if a company claims that its technology can treat prolonged grief, demanding evidence is not excessive paternalism. The company has voluntarily moved from selling remembrance to claiming therapeutic benefit.
The uncertainty surrounding griefbots therefore supports proportionate precaution, not prohibition: lighter safeguards where the risk is limited, and stronger evidence requirements when companies make stronger therapeutic claims.
What this proposal does not solve
Several questions remain open.
First, there is no good evidence yet for a universal time limit on griefbot use. Future empirical research should examine when short-term interaction helps, when dependency begins, and whether different users respond differently.
Second, consent from the deceased remains difficult. A person may consent while alive to being digitally recreated, but that does not settle every question about how the resulting avatar may later be modified or commercialized.
Third, privately created griefbots are harder to regulate than commercial services. If someone builds a local model using a deceased partner's messages, there may be psychological risks, but the case for state intervention is weaker because there is no provider exploiting the relationship for profit.
Finally, regulation cannot determine the “correct” way to grieve. That would turn a legitimate concern about technological power into an unjustified attempt to control deeply personal human experiences.
The principle that should guide regulation
The most important distinction is therefore not:
griefbots versus no griefbots.
It is:
support versus exploitation.
AI may give people new ways to remember, mourn, and even say goodbye. There is no good reason to prohibit those possibilities simply because they are unfamiliar.
But grief creates an unusual power imbalance. A system that imitates someone we loved can acquire emotional influence far beyond that of an ordinary app. When companies design that influence, profit from it, or present it as therapy, they should also accept responsibilities proportionate to the power they have created.
The goal of regulation should not be to tell people how to grieve.
It should be to ensure that when technology enters one of the most vulnerable parts of human life, comfort does not quietly become dependency, memory does not become manipulation, and grief does not become another engagement metric.
- By Jasmeet Singh
References
Bao, A., & Zeng, Y. (2024). Embracing grief in the age of deathbots: A temporary tool, not a permanent solution. Ethics and Information Technology, 26. DOI: 10.1007/s10676-024-09744-y.
Cholbi, M. (2025). Memory and Mimesis in Our Relationships With Posthumous Avatars. Oxford Intersections: AI in Society. Oxford University Press. DOI: 10.1093/9780198945215.003.0065.
Lindemann, N. F. (2022). The Ethics of ‘Deathbots’. Science and Engineering Ethics, 28, Article 60. DOI: 10.1007/s11948-022-00417-x.
Top comments (0)