Would you continue paying a monthly fee to hear the voice of someone who has died? The question may seem drawn from a science fiction tale, but it is already circulating in app stores and catalogs of digital services. Our everyday intuition tells us that grief is an intimate bond that no one should package as subscription software.
Nevertheless, the thanabots push right up to that boundary. There is no reincarnated consciousness or real contact with the deceased: there is a data and prediction architecture that generates plausible responses from digital traces. And the genuine conflict is not emotional but about control. Four questions organize the entire debate: who provides the data, who controls the avatar, who benefits financially, and what happens when the company shuts down or changes its terms.
What is a thanabot and why it is not resurrection
A thanabot, also called a griefbot, is an artificial intelligence system designed to imitate the way a deceased person spoke, their memories, or expressive traits. To achieve this, it combines generative models with very specific materials: messages, emails, photographs, audio clips, videos, and social media posts.
Put simply, the system does not “receive” messages from the afterlife nor remember anything. It calculates, with probability, what that person would say in response to a given question. That distinction is not a minor technical nuance: because it predicts rather than remembers, a thanabot can invent memories, attribute opinions the deceased never expressed, or generate emotionally convincing but false responses.
Philosophical framing helps situate it. In an analysis published in the Journal of the American Philosophical Association, Campbell, Liu and Nyholm (2025) examine whether a chatbot can preserve something of our relationships with the dead, and conclude that it can imitate conversational traits, but it does not preserve consciousness nor guarantee the identity of the person depicted. We are talking about symbolic continuity of the bond, not a real presence.
It is also worthwhile to separate two uses that are often conflated. A thanabot can function as a memorial archive or a narrative ritual, something akin to preserving letters or family recordings, but should not be presented as psychological treatment or as validated therapy. The evidence on its effects on grieving is still limited and heterogeneous.
How it works: from memorial archive to generative chatbot
The workflow can be summarized in a simple chain. First, someone provides the materials. Then the company stores and processes them. Those data feed the voice synthesis or the language model. Finally, the result is accessed through an account or a subscription, whether as text, synthetic voice, or audiovisual avatar.
The essential caveat is here: the quality and selection of those data completely condition the outcome. If the material comes from a particular social network or a specific life moment, the avatar will represent an edited personality, not the true complexity of the individual. There is a kind of interference between what the person was and what the model infers they would say.
Not all services do the same thing. There are two models under the same label. On one side, memorial services based on stories and recorded audio, which reproduce existing materials in the person’s authentic voice. On the other, generative chatbots, which produce new phrases with a simulated personality and can answer questions never asked in life. Services like HereAfter AI lean mainly toward the first group; other experimental proposals approach the second. The distance between the two is vast: one retrieves, the other fabricates.
What these services promise and what should worry us
The promises tend to be threefold: preserving memories, offering symbolic continuity, and facilitating access to the voice or the stories of the deceased. On paper, they seem appealing to anyone going through a loss.
However, “to converse” does not mean “to receive a message.” It means interacting with an interface that generates text or voice through statistical calculation. The emotional realism can be so high that it can be misleading: for a vulnerable user, a synthetic voice that mimics timbre and an authentic recording can be indistinguishable, even though their evidentiary value and legal status are radically different.
The qualitative study by Hurtado Hurtado and Glynos (2026), which analyzed ten journalistic accounts of real users and fifteen interviews with potential users, describes experiences of solace and a sense of continuity, but also ethical ambivalence, concern about realism, and fear that the simulation substitutes for grieving work. Thanabots are not inherently beneficial or harmful: their effect depends on whether they invite reflection or merely simulate a continued presence.
How much they cost and what happens if you stop paying
There is no standard price. It is an emerging technology and still minority, without a stable catalog of plans. According to ranges advertised in comparable services gathered in prior research, fees have moved, provisionally, between about $7 and $24 per month. Some apps also offer one-time purchases at higher amounts. It is worth noting that these figures vary by country, app store, and date, so they should be taken as approximate references rather than universal or verified rates for a specific service.
In addition to the subscription, there are the usual limits: maximum number of messages, available voice minutes, or premium features with extra charges. And here comes the nuance that almost never appears in advertising: paying a subscription does not equate to buying the avatar or the data that makes it possible. What you acquire, in many cases, is an access license subject to the terms of use and the applicable jurisdiction.
What if you stop paying? The service may limit messages, block features, keep the account for a grace period, or delete it according to its terms. There is no universal rule obliging the provider to hand over a portable conversational model. A truly useful export should include at least the original audios and videos, the photographs, the contributed texts, the transcripts, and the relevant metadata; downloading only a transcript does not allow reconstructing the avatar on another platform. Without real portability, the emotional legacy can be tied to a private server.
Who provides the data: the family, the deceased, and third parties who never chose
Typically, the data is uploaded by the surviving user, a family member, a digital executor, or the person themselves before death. In practice, the uploader is rarely the sole owner of the data. A family album includes voices from several people; an email reveals third-party information; a photograph may include minors or contacts who never consented to reuse.
There is also a consent distinction that should not be overlooked. The consent of the deceased is not interchangeable with the family’s, and authorization does not automatically cover every use. A person may have agreed to preserve their videos but not to authorize an AI to generate new phrases in their name. Storing a recording and synthesizing unseen speech are very different operations.
That is why the prudent approach is to require specific, informed, revocable consent limited to concrete purposes. The issue is not only technical: it is estate-related and moral. The deceased’s wishes may conflict with those of their survivors, and not always in the same direction: some asked to erase all trace, while others wished to preserve it.
Who controls the avatar: ownership versus technical control
Here is, in my view, the decisive lever. In most services, the company runs the infrastructure, the AI model, and the account. The user may hold a license to use or a right of access, but that does not imply control over the trained model, the metadata, backups, or future versions of the system.
In other words, owning an original photograph and technically controlling the representation generated from it are different things. The voice and the image carry enormous identifying and emotional value, but the control lever typically remains in the provider’s hands. And there is a troubling detail: the user does not necessarily control how the model is updated. The more it is modified, the more it may diverge from the original materials and from the image the family recognized. “Continuity” is not the same as stability.
Who benefits: the grieving turned into a market
The provider can charge for storage, for generating responses, for voice cloning, or for premium features. Surrounding the avatar may also be imagery rights, authorship, performance, trademarks, prior contracts, or the estate’s financial interests. The answer to who earns money changes by country and depending on whether the service uses materials in the public domain, protected works, or the deceased’s identifiable voice.
The underlying twist is subtle but profound: a memory-keeping practice becomes a recurring consumer relationship. Price, usage limits, and automatic renewal introduce a barely visible tension, because access to an affective representation depends on the family’s ability to pay. Soh and colleagues’ systematic review (2026), which analyzed thirty studies on digital technologies applied to mourning, identifies dependence as one of the recurring risks, without proving that every thanabot causes clinical harm. It is wise not to conflate a noted risk with an inevitable consequence.
What happens if the platform closes or changes its terms
A change in terms can broaden the purposes of use, incorporate advertising, allow new ways of training the model, or change the price. The user should check whether a new consent is required, whether they can refuse it, and whether there is a real way to recover or delete the contributed materials.
The most delicate scenario is closure. If a company goes bankrupt, dissolves, or is acquired, the data may be subject to corporate transfer, insolvency proceedings, or a variable deletion policy. The avatar’s continuity should not be taken for granted. That is why contracts should clearly specify who acts as data controller, how long they keep it, where it is stored, whether they sell or share it, and what happens after service termination. In short, a family may bind their memory to a private provider and lose access just when they need it most.
Privacy and AI in a fragmented legal zone
Data protection starts from a striking asymmetry. Recital 27 of the General Data Protection Regulation states that the GDPR does not apply to the personal data of deceased people and leaves regulation to each member state. The result is a non-harmonized protection for digital legacy. That said, this does not mean that any material can be used freely: files often contain data about living family members and third parties, and contracts, intellectual property, and image rights remain relevant.
In Spain, Article 3 of Organic Law 3/2018 allows family members, de facto partners, heirs, or expressly designated persons to request access, correction, or deletion of the deceased’s data, with limits if the deceased prohibited it. Yet, the rule does not by itself resolve who can order the creation of a generative avatar or who controls the AI model. France, for its part, allows leaving guidelines on the preservation and communication of data after death, while in the United States the answer combines state personality rights, copyright, and platform contracts, without a single federal rule. In all cases, the legal fit depends on the country, the type of data, and the provider’s terms.
Beyond this panorama lies a basic transparency requirement: every response should always be identified as AI-generated, so that no vulnerable user confuses a statistical calculation with a message from the deceased.
What we truly know about its effects on grief
The temptation to vendor these tools as guaranteed comfort is obvious, but the evidence calls for caution. Neither the review of thirty Soh et al. studies nor the qualitative study by Hurtado Hurtado and Glynos allows us to conclude that thanabots improve grief, reduce symptoms overall, or function as a validated therapeutic intervention. What they describe are mixed experiences: comfort and symbolic continuity in some cases, dependence, inaccurate representation, and privacy concerns in others.
There are also data on potential acceptance that should be handled carefully. In some surveys, a portion of people who have lost a loved one are willing to talk with a deceased avatar. However, willingness to try something does not measure its clinical efficacy nor sustainable real-world use over time. Put simply: a thanabot can serve as an archive or as a symbolic ritual, but it is not validated as a treatment.
Questions every family should ask before subscribing
Before contracting such a service, it is worth pausing on some uncomfortable questions. Who is the legal data controller and who can manage the account after the user’s death? Does the consent authorize only storage or also the generation of new voice, text, and images? What happens to the data of family members and third parties included in the files? Can the company use those materials to train its own or third-party models?
And practically: can the originals, transcripts, history, and metadata be exported in reusable formats? Is there a complete delete, backups included, and how long does it take to execute? What happens if you stop paying, if the price rises, or if the platform closes? Can the information be transferred to another company in a sale or bankruptcy? Finally, two questions that almost no one asks but are essential: how can it always be identified that the response comes from AI and not from the deceased, and what limits exist to prevent manipulative responses or those that encourage emotional dependence.
Where this is headed
What is striking, when you step back from the astonishment at the synthetic voice, is not the imitation itself but the access, control, and portability architecture that ends up in the hands of a company. There, one decides whether a private memory remains ours or becomes a function tied to an account, a server, and a revisable commercial policy. If clear standards for consent, export, and deletion do not mature, we risk grief becoming dependent on subscriptions and on conditions that can change without notice. Perhaps the most useful lesson is this: before asking whether we want to hear that voice again, we should ask who stores it, who modifies it, and who can switch it off.
References
- Campbell, Stephen M., Pengbo Liu and Sven Nyholm. “Can Chatbots Preserve Our Relationships with the Dead?”. Journal of the American Philosophical Association, vol. 11, no. 2, June 2025, pp. 230-248. Published online February 10, 2025. DOI: 10.1017/apa.2025.1.
- Hurtado Hurtado, Josué, and Jason Glynos. “Thanabots, fantasy, and the ethics of mourning”. AI & Society, published July 6, 2026. DOI: 10.1007/s00146-026-03200-9.
- Soh, Xun Ci, Andree Hartanto and cols. “Digital grief technology to support bereavement: A systematic review of potential benefits and risks”. Computers in Human Behavior Reports, vol. 22, 2026, article 101148. DOI: 10.1016/j.chbr.2026.101148.
- Reglamento (UE) 2016/679 del Parlamento Europeo y del Consejo, de 27 de abril de 2016, considerando 27.
- España. Ley Orgánica 3/2018, de 5 de diciembre, de Protección de Datos Personales y garantía de los derechos digitales, artículo 3. Boletín Oficial del Estado, 2018.
