Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives

Abstract

As AI systems quickly improve in both breadth and depth of performance, they lend themselves to creating increasingly powerful and realistic agents, including the possibility of agents modeled on specific people. We anticipate that within our lifetimes it may become common practice for people to create a custom AI agent to interact with loved ones and/or the broader world after death. We call these generative ghosts, since such agents will be capable of generating novel content rather than merely parroting content produced by their creator while living. In this paper, we first discuss the design space of potential implementations of generative ghosts. We then discuss the practical and ethical implications of generative ghosts, including potential positive and negative impacts on individuals and society. Based on these considerations, we lay out a research agenda for the AI and HCI research communities to empower people to create and interact with AI afterlives in a safe and beneficial manner.

Introduction

The past few years have brought incredible growth in the capabilities of generative AI models, particularly large language models (LLMs) such as GPT-4 (OpenAI 2023a), Palm 2 (Anil et al. 2023), and Llama 2 (Touvron et al. 2023), though there has also been incredible progress in generative AI for the production of images (Ramesh et al. 2022), video (Singer et al. 2022), and audio (Borsos et al. 2023), as well as a new generation of multimodal models (Yang et al. 2023; Google DeepMind 2023) that combine functionality across several media categories. These models, in turn, have given rise to new types of generative agents (Park et al. 2023), simulacra that can produce believable human behaviors, including capabilities such as memory and planning. While still in their infancy, generative agents and related technologies are likely to increase in fidelity and popularity as underlying model capabilities improve and compute costs drop. For instance, in November 2023 OpenAI released GPTs (OpenAI 2023b), a no-code interface for people to develop agentic AIs.

As AI models increase the set of human capabilities they can faithfully reproduce (Morris et al. 2023; Bubeck et al. 2023), societal change is inevitable. For instance, experts anticipate that powerful AI systems may profoundly change disparate areas of society such as the labor market (Eloundou et al. 2023), the education system (Kasneci et al. 2023), the pursuit of scientific knowledge (Morris 2023), and criminal activities (Ferrara 2023). In this paper, we discuss how advances in AI might change personal and cultural practices around death and dying.

We introduce the concept of generative ghosts, AI agents that represent a deceased person, and discuss why we anticipate such representations will become popular within our lifetimes. We explore the design space of possible instantiations of generative ghosts and consider both the benefits that might lead to their adoption and the practical and ethical concerns such technology may introduce.

Our contributions include: (1) identifying and characterizing an emerging phenomenon of creating “generative ghosts” to represent the deceased, (2) introducing a taxonomy of design dimensions and analysis of potential benefits and harms that can be used to support future empirical research and motivate fieldwork. By characterizing this emerging trend, highlighting potential risks to be averted, and creating a framework for future investigation, we aim to ensure that future technical and sociotechnical systems will maximize the potential benefits of “AI afterlives” while minimizing potential risks.

Related Work

We discuss the rich literature on how technologies have changed practices around death and dying and initial forays into AI afterlives by individuals and start-up ventures.

Post-Mortem Technology

Throughout history, people have turned to technology to remember, memorialize, and even interact with the dead. Gravestones and other burial markers can be traced nearly back to 3000 B.C.E. (Taylor 2001). Obituaries in the U.S., while dating back to the 16th century, became more common during the 19th century in part due to the U.S. Civil War (Hume 2000) – an event that also brought embalming into favor. Even the mediums of the Spiritualism movement in the late 19th and early 20th century turned to telegraphs, radio-wave detectors, and later wireless radio in their attempts to detect the presence of and communicate with the dead (National Science and Media Museum 2022).

During the earliest days of the World Wide Web, when people would create personal Home Pages describing their lives and family, it was routine for people to dedicate a page to the memory of a family member, often a deceased parent or family pet (Brubaker 2015). Online graveyards, websites specifically dedicated to the memorialization of the dead, soon followed (Roberts and Vidal 2000; Roberts 2004). The Virtual Memorial Garden is the earliest documented example (Roberts and Vidal 2000). Created in 1996, it featured a collection of brief obituaries authored by loved ones, capturing their grief, the circumstances around the death, and loved ones’ guilt.

Even as people adopted digital technology to memorialize the dead, scholars in human-computer interaction and social computing began to note how mortality is overlooked by technology design and have argued for increased attention to end-of-life and mortality (Massimi and Charise 2009; Massimi et al. 2010; Odom et al. 2010; Bell 2006), often termed “thanatosensitive design” (Massimi and Charise 2009). Scholars have studied the intersection of technology and mortality across diverse contexts, including digital heirlooms (e.g., (Odom et al. 2012; Beuthel and Fuchsberger 2022)), communal rituals (e.g., (Häkkilä, Colley, and Kalving 2019; Uriu et al. 2021)), online memorials (e.g., (Gulotta et al. 2014; Moncur and Kirk 2014; Massimi 2013)), digital legacy (e.g., (Gulotta et al. 2016, 2014; Gulotta, Kelliher, and Forlizzi 2017; Doyle and Brubaker 2023)), and family archives (e.g., (Kaye et al. 2006)). Yet, as Doyle and Brubaker summarize (Doyle and Brubaker 2024), end-of-life scenarios also present new privacy challenges (e.g., (Holt, Nicholson, and Smeddinck 2021; Locasto, Massimi, and DePasquale 2011)), challenges that result from shifting motivations at different life stages (Chen, Vitale, and McGrenere 2021; Thomas and Briggs 2014), and challenges due to differences between user expectations and platform functionality (Gach and Brubaker 2020).

In HCI, extensive attention has been given to digital legacies, the collection of materials that carry values and meaning that are passed down and/or otherwise continue to represent the deceased after their death. A recent literature review of the field by Doyle and Brubaker (Doyle and Brubaker 2023) identified four foci of this scholarship: digital identity (i.e., how legacies can continue to represent the deceased in intentional and unintentional ways), engagement with digital legacies (i.e., studies on use and user perceptions), putting to rest (concerned with issues around preservation, disposition, and disposal), and the integration of technology into existing legacy practices (Doyle and Brubaker 2023).

In most cases, it is the digital traces people leave behind that give rise to legacy crafting and memorialization opportunities. The sources of legacy content vary – from burner accounts (Gulotta et al. 2016) to personal archives (She et al. 2021; Kaye et al. 2006) – but social media content (especially memorialized profiles) has received the most scholarly attention (e.g., Brubaker et al. 2014; Gach and Brubaker 2020, 2021; Brubaker and Callison-Burch 2016; Brubaker and Hayes 2011; Brubaker, Hayes, and Mazmanian 2019; Getty et al. 2011; Mori et al. 2012; Sofka 2009, 2017; Arnold et al. 2017; Nansen et al. 2019).

As Doyle and Brubaker (2023) summarize, the persistence of identities has benefits and challenges: Digital legacies can facilitate healthy grieving (She et al. 2021; c.f., Brubaker, Hayes, and Dourish 2013; Brubaker et al. 2012) and maintain connections to the deceased (Brubaker and Hayes 2011; Getty et al. 2011; Gulotta et al. 2014). However, receiving what is often a large and uncurated set of content can be overwhelming for loved ones (Holt, Nicholson, and Smeddinck 2021), and may provide (for better or worse) an uncensored version of the deceased (Gulotta et al. 2013). Loved ones may also find themselves overwhelmed as they become the contact point for the deceased’s various online networks (Brubaker et al. 2014; Brubaker, Hayes, and Dourish 2013).

Across many studies, participants describe uncertainty around the use, care, and understanding of digital legacies. Ambiguities often arise when data created for one purpose – e.g., personal archives (Kaye et al. 2006; Holt, Nicholson, and Smeddinck 2021) or interpersonal communication (Moncur and Kirk 2014; Thomas and Briggs 2014; Gach and Brubaker 2021, 2020; Gulotta et al. 2014) – are repurposed post-mortem. In contrast to traditional legacies where there are familiar norms and practices, Pfister (2017) argues that common norms around digital legacy have not yet emerged, an issue that AI will certainly exacerbate.

AI Afterlives

In recent years, a handful of tech-savvy individuals have attempted to posthumously create interactive memorializations of loved ones by ingesting their digital (or digitized) content such as emails, journals, videos, photographs, or other autobiographical media. In her graphic novel Artificial: A Love Story, Amy Kurzweil 2023 documents how her father, futurist Ray Kurzweil, created a chatbot to embody the memory of his deceased father, Fred Kurzweil, in a form he dubs “Fredbot”(Veltman 2023). Fredbot interactively responds to questions from his descendants, but only by sharing exact quotes from the materials Fred left behind such as letters digitized by his family. In another well-publicized incident, the engineer Eugenia Kuyda created an app to preserve the memory of her best friend, Roman, who died unexpectedly in an accident, training a neural network on text messages her friend had sent her to create a bot named, suitably, “Roman.” Unlike Fredbot, which was available only to immediate family, the Roman bot was made available on social media and app stores for public interaction, resulting in mixed reactions from friends and family of the deceased (Newton 2016).While Fredbot and Roman bot represented private citizens, AI has also been used to re-create public figures in various ways. For instance, in November 2024, the surviving members of The Beatles released a new song, “Now and Then,” (Beatles 2023) using AI to enable the deceased John Lennon to sing along with his living bandmates (Savage 2023). The musician Laurie Anderson collaborated with the University of Adelaide’s Australian Institute for Machine Learning to create a chatbot based on her longtime partner, deceased musician Lou Reed, which, in addition to conversing, generates new music lyrics in Reed’s style (Andereson 2024). In early 2024, gun-control activists in the used AI to re-create the voices of victims of gun violence, to create a visceral and poignant plea to U.S. legislators (Luscombe 2024).

Meanwhile, there are a handful of start-up companies that aim to give people the ability to create their own AI afterlives while they are still alive. Re;memory (rememory.deepbrain.io/en) is an offering from DeepBrain AI that professes to create an interactive virtual representation after a seven-hour filming and interview session. They specifically advertise this service as a way that one can proactively create a rich memory that friends and family can engage with after one dies. HereAfter (https://www.hereafter.ai) provides an app that interviews a user with the goal of proactively creating a posthumous digital representation. Friends and relatives can interact with a chatbot-based representation of their loved one who can provide photos and voice recordings of memories and life events. Companies offering ”digital immortality” are increasingly popular in China (Yang 2024).

In contrast with Re;memory and HereAfter, which aim to empower users to preserve their memories before their deaths, some start-ups purport to use AI to provide an experience more akin to resurrection or reanimation (i.e., without the planning or consent of the deceased party). For instance, Character.AI (https://beta.character.ai/) uses LLMs to enable a chat experience with deceased public figures for entertainment or educational purposes (e.g., chatting with a bot representing William Shakespeare), and the Musée D’Orsay in Paris recently contracted with a start-up company to develop a chatbot meant to represent the artist Vincent van Gogh by ingesting his writings (Small 2023). Beyond public figures, Project December (projectdecember.net/) is a venture that allows end-users to upload content from a deceased loved one (Fagone 2021) in order to “simulate the dead” via a customized chatbot.

While the examples shared above demonstrate the diversity of emerging approaches, a key trend in post-mortem AI is around griefbots. Griefbots (Jiménez-Alonso and de Luna 2022; van der Vorst and Kamp 2022) or deathbots (Lindemann 2022) are typically described as chatbots (predating modern large language model technologies) designed to allow the bereaved to “converse” with the deceased, typically created posthumously by a third party who leverages general-purpose text the deceased created during their life. The aforementioned “Fredbot” and “Roman” would be considered examples of griefbots. Because of their third-party origins, griefbots raise controversial ethical issues regarding privacy and consent (Grandinetti, DeAtley, and Bruinsma 2020).

Generative ghosts can be considered an extension of the concept of a griefbot, but not all griefbots would be considered generative ghosts. Generative ghosts go beyond the traditional definition of griefbots in several ways: they always include the ability to generate novel content in-character (whereas many griefbots only regurgitate content from a training corpus), potentially evolve over time, and possess agentic (Park et al. 2023; Shavit et al. 2023) capabilities such as the ability to participate in the economy or perform other complex tasks with limited oversight. Notably, our framework for generative ghosts envisions the possibility of first-party (rather than only third-party) creation as part of an end-of-life planning process. Also, generative ghosts may be designed to represent non-human entities (such as pets), and in many cases might exist pre-mortem as a generative clone. Finally, in contrast to griefbots, generative ghosts may be created for reasons other than to support the bereaved in their grieving process. The following section introduces more details about the design space of generative ghosts.

Generative Ghosts: A Design Space

We introduce the notion of a generative ghost, an agentic, AI-powered representation of a deceased individual. Generative ghosts differ from other technology-mediated representations of the deceased (such as memorialized social media accounts) in that they provide more than static or interactive access to content about the deceased produced during their lifetime (e.g., text, imagery, audio, video, or other media). Rather, they additionally are capable of generating novel content and interactions based upon the deceased’s data, potentially mediated by additional metadata such as personality questionnaires (e.g, (NORC at the University of Chicago n.d.)), prompts (Zamfirescu-Pereira et al. 2023), constitutions (Bai et al. 2022), or other rules set forth by the generative ghost’s creator. In addition to producing novel content, generative ghosts may also possess agentic (Park et al. 2023; Shavit et al. 2023) capabilities (i.e., the ability to autonomously execute complex sequences of actions with side effects in either the digital or physical realms).

Generative ghosts can be considered a special case of a broader class of agentic generative AI systems meant to represent a specific individual, which we will call generative clones. Generative clones are agents created for or by a living person to mimic their persona in order to execute actions and interactions on their behalf during their lifetimes. For example, a busy professional might create a generative clone of herself to respond to low-priority emails or phone calls in a manner that mimics her writing style, voice, and other characteristics such that her correspondents would think she had personally crafted the replies. Using this terminology, a generative ghost can be defined as a generative clone representing a deceased person.

In the rest of this section, we expand on our definition of generative ghosts by discussing the design space of potential instantiations of this concept. While all generative ghosts share the capability to retrieve pre-existing multimodal content and generative novel multimodal content representing a deceased individual (including through the execution of agentic interactions), we argue that several key dimensions may impact the capabilities, perceptions, and societal impacts of such systems, including provenance, deployment timeline, anthropomorphism paradigm, multiplicity, cutoff date, embodiment, and representee type. As we detail in the Discussion, interrelationships between these dimensions are common. However, we present them separately here for the sake of clarity.

Provenance: First-Party vs. Third-Party Ghosts

The provenance of a generative ghost refers to who created it (or authorized its creation). A first-party generative ghost is one created by the individual being represented (i.e., perhaps via end-of-life planning or legacy crafting practices). In contrast, a third-party generative ghost is not created by the representee. Third-party generative ghosts might be created by parties with a direct personal connection to the deceased (e.g., family, friends), by parties with a fiscal connection to the deceased (e.g., employers, estates), or by unconnected parties. Additionally, third-party ghosts may be authorized or unauthorized by the representee. Authorized third-party generative ghosts, for example, might be created with consent (e.g., via the deceased’s will). Meanwhile, unauthorized ghosts created by unconnected third parties are most likely in the case of public figures such as historical figures or contemporary celebrities; indeed, we already see early attempts at creating third-party generative ghosts of historical figures for entertainment and educational purposes by companies such as character.ai (Metz 2023) and nonprofits such as Khan Academy (Grant 2023).

Deployment Timeline: Pre-Mortem vs. Post-Mortem

The deployment timeline of a generative ghost indicates whether the system’s initial deployment was pre-mortem or post-mortem. While some generative ghosts will be deployed post-mortem with the explicit purpose of memorializing the dead, many generative ghosts will likely start off as agents that serve some pre-mortem functionality (e.g., responding to personal emails or accomplishing other tasks on behalf of the representee, in character as if the representee had executed the task themself) – what we have termed generative clones – only to transition to ghosts after the death of their representee. Pre-mortem deployments might be desirable as they would allow an individual the opportunity to tune the behavior and/or capabilities of a first-party generative ghost to their liking, perhaps as part of an end-of-life planning process. Likewise, we imagine that generative clones might be designed with mortality in mind, potentially with attendant modifications to their behavior and/or capabilities once they become ghosts.

Anthropomorphism Paradigm: Reincarnations vs. Representations

The anthropomorphism paradigm of a generative ghost refers to a subtle, but potentially impactful, interface metaphor choice: whether a generative ghost presents itself as a reincarnation of the deceased individual or as a representation of that individual. Generative ghosts employing a reincarnation metaphor would engage with end-users as if they were the deceased individual. In contrast, generative ghosts employing a representation metaphor would engage with end-users as if they are a representation of the deceased, but are not the deceased themselves. Design choices that might impact perceptions of whether a generative ghost is a reincarnation or a representation include whether a generative ghost uses first-person pronouns, whether it uses the present or past tense when discussing its representee, whether it uses the name of its representee or a different nomenclature (e.g., the use of “Fredbot” in (Kurzweil 2023)), whether it is allowed to make statements that assert it is alive, possesses a soul, etc.

Multiplicity: Single Ghost vs. Multiple Ghosts

The multiplicity design dimension refers to whether an individual is represented by a single generative ghost or by multiple distinct ghosts. Multiple ghosts might arise intentionally (i.e., a single creator develops a set of generative ghosts with different behaviors, capabilities, and/or audiences in mind). Multiple ghosts might enable people to customize different ghosts for different audiences in order to avoid forms of context collapse (Davis and Jurgenson 2014; Litt 2012; Marwick and Boyd 2011). Multiple ghosts might also arise unintentionally due to multiple third parties creating generative ghosts for a single individual, or perhaps due to post-mortem identity theft or other novel cybercrimes.

Cutoff Date: Static vs. Evolving

The cutoff date of a generative ghost describes whether the ghost remains static (i.e., not develo** novel interests, skills, or other characteristics after its representee’s death) or whether it is evolving. Evolving ghosts’ characteristics might change over time, such that they may eventually diverge in substantial ways from the deceased individual. For example, consider a hypothetical situation in which a parent creates a third-party generative ghost to represent a deceased child. A static cutoff-date design would result in a representation that perpetually interacted in a style faithful to the appearance, diction, maturity, etc. of a young child, whereas an evolving representation might “age” (either in real-time or per some other scheme).

While the previous example speaks to whether a ghost is designed to evolve by simulating aging, ghosts will also evolve (albeit in less predictable ways) should new information be added to their models (e.g., in the case that additional data about the deceased was found). Moreover, the informational context beyond data about the representee (e.g., from world news to the personal information of those who might interact with the ghost) can be static or be allowed to evolve. One can imagine scenarios in which a ghost might be designed to incorporate (or not) new information, including personal information (such as a wedding or birth of a child) or world events (such as a future election).

Embodiment: Physical vs. Virtual Ghosts

The embodiment dimension refers to whether a generative ghost has a physical embodiment. Such embodiments might be physical in a literal sense (e.g., robotics) or might be embodiments in rich digital media (e.g., avatars in mixed reality environments). In contrast, purely virtual generative ghosts would lack embodiment (e.g., perhaps available only as a chatbot). In addition to technical and/or cost constraints, there may be additional reasons to opt for lower-fidelity (i.e., virtual-only) embodiments, including potential ethical or psychological concerns related to physical embodiments (i.e., the strong anthropomorphic cues of embodied generative ghosts might elicit a reincarnation rather than representation metaphor, which might be undesirable for certain users or contexts).

Representee Type: Human vs. Non-Human Legacies

The representee type dimension acknowledges that, in addition to representing deceased humans, people may wish to create ghosts representing non-humans, such as beloved family pets or service animals. While this paper is primarily concerned with generative ghosts representing deceased people, we also recognize the need for reflection on technology to support interactive remembrances of non-humans.

Anticipating Benefits and Risks of Generative Ghosts

Having introduced the concept of generative ghosts and the key associated design parameters, we now discuss the potential benefits of thoughtfully and safely implemented generative ghosts, followed by a discussion of potential risks associated with various instantiations of this paradigm.

Generative Ghosts: Potential Benefits

We first consider likely motivations for creating generative ghosts. With appropriate attention to technical concerns (interaction design, AI safety, etc.) and sociotechnical concerns (ethics, cultural concerns, legal ramifications, etc.), one can envision a positive future in which generative ghosts offer a host of benefits both to the representee, the bereaved, and society more broadly.

Potential Benefits for Representees

Individuals may choose to create first-party generative ghosts as part of their end-of-life planning for several reasons. On a personal level, people may feel comforted by or even feel hopeful in the idea of a “digital afterlife.” Prior work on digital legacy has documented people’s motivations to capture their life stories (Thomas and Briggs 2014), major life events (Jamison-Powell et al. 2016), and pass on values to future generations (Gulotta, Kelliher, and Forlizzi 2017). Posthumous, personalized AI representations may likewise be seen as a way for people to ensure that they are remembered, whether by family, friends, or perhaps the world at large.

The likely nature of generative ghost architectures makes such remembrances higher-fidelity, more interactive, and more personalized than today’s more common digital memorials, such as memorialized social media accounts. Generative ghosts may help individuals feel agency over their posthumous future, such as by presenting an ability to comfort loved ones after they have passed, to offer advice to loved ones at future events (e.g., convey well wishes and advice upon the marriage of a great-grandchild), or to preserve their personal, religious, and/or cultural heritage by sharing knowledge and wisdom specific to their background and time in history to future generations.

It is even possible that generative ghosts may offer legal or economic benefits to representees. For example, one can imagine that creating a generative ghost might replace or complement life insurance policies for some individuals if their AI embodiments can participate in our economic system and earn an income for their descendants (e.g., an author whose generative ghost continues to produce novel works in their style). Depending on how our legal system evolves, generative ghosts may also provide utility in ensuring the proper execution of other aspects of end-of-life planning, such as arbitrating disputes over interpretations of a will.

Potential Benefits for the Bereaved

Interacting with a generative ghost may provide emotional support to close friends and family of the deceased. Prior research has considered the impact of online memorials on the bereaved (e.g. (Walter et al. 2012)), often responding to concerns that such spaces might prolong grief. Yet scholars cite the benefits of online memorials that allow loved ones to maintain healthy “continuing bonds” (Klass, Silverman, and Nickman 1996) with the deceased (e.g., (Getty et al. 2011; Brubaker and Hayes 2011; Klass, Silverman, and Nickman 1996); c.f., (DeGroot 2018)), often in a space where those grieving can gather (Carroll and Landry 2010; Brubaker, Hayes, and Dourish 2013).

As with griefbots more generally, discussions with a generative ghost might provide comfort or closure and support a continued feeling of closeness. Generative ghosts may support positive mental health, particularly if they enable the bereaved to maintain connections with the deceased while “making sense” of the loss (termed “accommodation” (Neimeyer and Thompson 2014)). Family members may find it comforting not only to interact with the generative ghost in the immediate post-mortem period, but to know that they might be able to interact with their deceased loved one at upcoming future life events such as weddings, the birth of a child, or other milestones during which they may wish to share updates with the generative ghost and/or hear advice, congratulations, or other event-appropriate commentary from their deceased loved one.

One-sided conversations with deceased loved ones, such as speaking graveside or writing to the deceased on a memorial, are common and well-recognized in the literature (DeGroot 2018; Brubaker and Hayes 2011; Brubaker, Hayes, and Dourish 2013) and are generally considered a part of healthy grieving (Klass, Silverman, and Nickman 1996). As summarized by DeGroot (DeGroot 2018), these conversations are cathartic and help the living confront their loss: “People essentially struggle to develop an identity where the deceased is now part of the past self and not the present self… and communicating with the deceased appears to ease the transition” (DeGroot 2018). However, it is worth noting that the dead do not typically respond. Rather, people imagine how the deceased would reply (DeGroot 2018). It unclear how responses from a generative ghost might impact bereavement.

Finally, generative ghosts may also embody practical knowledge related to end-of-life planning in a method that is easily discoverable to and digestible by surviving relatives, such as hel** convey information to support funeral or memorial planning or supporting discovery of items associated with transitions of ownership of financial accounts or other property such as passwords or locations of key items. Generative ghosts might also support family members by providing advice on procedures that they had been responsible for in life (e.g., teaching a surviving spouse how to cook a favorite dish or repair the kitchen faucet). In some cases, income provided by generative ghosts’ participation in the economy might support family members.

Potential Benefits for Society

The interactive modality of generative ghosts may provide benefits, particularly in the cases of culture, history, and heritage. As mentioned above, organizations such as character.ai and Khan Academy are already considering the benefits for entertainment and education.

Generative ghosts may also be beneficial for smaller cultural groups. They may be one way to preserve the collective wisdom of elders, as well cultural heritage such as knowledge of dying languages, religions, or other small-membership cultural practices that are at risk of being forgotten as their members age. For instance, generative ghosts may be one way to preserve historical knowledge about events such as the Holocaust before the few remaining elderly survivors with firsthand experience pass away.

Indeed, the creation of generative ghosts may enrich the practice of disciplines such as museum curation, historical scholarship, anthropology, and other related humanities. These disciplines may benefit, not only from the static knowledge encoded in generative ghosts, but from the ability to interactively query such representations to understand perspectives and events in new ways. For example, devout adherents to a specific religious tradition might be able to converse with generative ghosts of deceased elders or scholars of their faith about changing aspects of society (e.g., perhaps an Orthodox Jew might engage with generative ghosts of Talmudic scholars to debate the pros and cons of whether kashrut allows consumption of synthetic, lab-grown pork). Generative ghosts might also represent archetypes or amalgamations rather than specific individuals, particularly when developed from historical records (e.g., “citizen of Pompeii,” “typical resident of Colonial Williamsburg”).

Generative Ghosts: Potential Risks

While thoughtful implementations of generative ghosts have potential upsides for the representee, their loved ones, and society more broadly, there are also potential risks associated with this paradigm. By proactively anticipating these risks, we may be able to design technical and sociotechnical systems that avoid or mitigate them. Of course, as with any powerful new technology, it is impossible to fully anticipate apriori the many potential sociotechnial side-effects of its introduction; the emergence of “AI afterlives” may reshape society in complex ways beyond our current imagining (i.e., spurring the advent of new religious movements). Here, we discuss four broad categories of risk associated with generative ghosts: mental health risks, reputational risks, security risks, and socio-cultural risks.

Mental Health Risks

Earlier we discussed ways in which generative ghosts might offer mental health benefits to mourners. However, ghosts may also pose mental health risks, including risks related to delayed accommodation, excessive attachment, information overload, anthropomorphism, deification, and second deaths. Studies of prior generations of AI technologies (and even of non-AI computing systems (Reeves and Nass 1996)) have found that people tend to anthropomorphize such systems and even develop attachments (Gabriel et al. 2024). We anticipate that without careful design generative ghosts might magnify these risks due to the increasing fidelity of generative AI technologies and the modeling of the agent on a specific individual.

Interacting with a generative ghost may impact the bereaved’s ability to assimilate and accommodate (Payne, Joseph, and Tudway 2007; Neimeyer and Thompson 2014) their loss and integrate the reality of their loved one’s death (Klass, Silverman, and Nickman 1996). Delayed accommodation may negatively impact the bereaved’s mental health and quality of life by supporting loss-oriented experiences (e.g., active grieving, reminiscing while looking at old photos, etc.) at the expense of restorative-oriented experiences (e.g., the development of new relationships and life patterns) (Stroebe and Schut 1999). Stroebe and Schut argue that both loss- and restorative-oriented experiences are necessary, but that having agency around when and how one oscillates between the two is important – agency that could be impeded by the interactive design of generative ghosts (e.g., if a generative ghost were designed to use “push notifications” that might pull a mourner into an interaction rather than letting the mourner decide when and how they want to engage in remembrance and other grief-related activities).

A related concern is the possibility of develo** excessive attachment to generative ghosts, as indicated by spending a disproportionate amount of time interacting with representations of the deceased in lieu of interacting with the living. Already, there are indications that some individuals find AI-powered companions, such as those from Replika (replika.ai), to be highly compelling (Liang 2023); generative ghosts’ basis in real, beloved individuals rather than in fictional personas or historical figures may amplify this type of attachment.

In addition to spending an unhealthy amount of time interacting with a generative ghost, another aspect of excessive attachment may take the form of consulting a generative ghost’s advice for momentous or even mundane life decisions. While soliciting input from ghosts might be positive or harmless in some circumstances, the availability of ghosts might also result in information overload, lead to choice paralysis, or create tensions over how to weight competing recommendations from a generative ghost vs. living friends and relatives.

Anthropomorphism of generative ghosts is a potential risk for mourners, if they become convinced that a generative ghost truly is the deceased (rather than a computer program representing them). Anthropomorphism might increase the likelihood of other mental health risks such as excessive attachment or delayed accommodation. Some design choices we discussed earlier may increase anthropomorphism risk, particularly the use of reincarnation metaphors, evolving representations, and physical embodiments.

Deification is a more extreme version of anthropomorphism risk, in which a mourner might develop religious or supernatural beliefs about a generative ghost. Such beliefs may carry risks such as alienating the mourner from living companions, excessive attachment, over-reliance on the generative ghost for life advice (i.e., treating it as an oracle), and/or a propensity to unquestioningly carry out actions suggested by the generative ghost (potentially including actions detrimental to themselves or others).

An additional mental health risk is that of second deaths or second losses, the experience of grieving a loved one again. In digital contexts, second deaths occur when data becomes unavailable either through technical obsolescence, deletion, or lack of access (Bassett 2022). In the case of generative ghosts, second death refers to the emotional harm that may result if a ghost ceases to exist. Second deaths may occur for myriad reasons, including economic (i.e., the company that maintains the generative ghost service goes out of business; survivors’ or the deceased’s estate can no longer afford maintenance fees), regulatory (i.e., a government outlaws generative ghosts, which may be particularly likely in theocracies where AI afterlives threaten established belief systems), technological (i.e., obsolete generative ghosts no longer run on future infrastructure), or security reasons (i.e., a hacker disables a generative ghost). While second loss is not unique to AI systems, the potential intensity of engagement that generative ghosts can support means that the loss experienced at the disabling of a generative ghost might be especially profound.

Reputational Risks

By reputational risks of generative ghosts, we refer to situations in which the generative ghost’s interactions might tarnish the memory of the deceased. In addition to altering people’s perceptions of the deceased in a manner they would not have desired, reputational risks may also harm the living, either via association with (the now tarnished) deceased (e.g., “your grandfather was racist”) or through harming their mental health (e.g., a son learning from a generative ghost that his father strongly preferred his sibling over him). We identify three classes of reputational risk: privacy risks, hallucination risks, and fidelity risks.

Privacy Risks refer to scenarios in which a generative ghost exposes true information that the deceased would not have wanted to be revealed (e.g., because the information is embarrassing, criminal, hurtful to loved ones, etc.). While some privacy risks may be anticipated and prevented via rules when creating a ghost (i.e., “don’t tell my spouse about my affair”), revelations that arise from generated content may be more difficult to prevent (e.g., if the AI correctly infers and reveals the deceased’s sexual orientation based on patterns in their data, even though they were closeted while alive). Likewise, what constitutes a privacy risk changes over time. Aspects of the deceased’s life or beliefs that were socially acceptable during their lifetime may come to be frowned upon by future societies (e.g., perhaps in an environmentally conscious future, people who consumed certain kinds of meat or used certain levels of fossil fuels might be retroactively viewed as immoral).

Privacy risks are context-dependent (Nissenbaum 2009). For instance, it may be acceptable for a generative ghost to share sensitive information with some audiences, but not with others (e.g., perhaps it is acceptable for the ghost to discuss the deceased’s religious beliefs with their spouse, but not with their close work friend). The creation of multiple ghosts, each with different knowledge or abilities targeted toward different audiences, might mitigate privacy risks related to context collapse (Davis and Jurgenson 2014; Litt 2012; Marwick and Boyd 2011).

Hallucination risks arise when a generative ghost reveals false (but plausible-sounding) information about the deceased. As with privacy risks, such information may tarnish the memory of the deceased (in this case, without true cause), as well as harm the mental well-being of loved ones who believe the falsehoods. Hallucination risks might arise unintentionally (i.e., due to a failure of the underlying AI technology) or might occur due to malicious activity, such as the hacking or hijacking of a generative ghost in order to harm the represented individual or their survivors.

Fidelity risks refer to challenges that might arise from both accurate and inaccurate information. Related to hallucination risks, ghosts providing inaccurate information can be a problem, particularly in historical, legal, and economic applications. Yet accurate information also comes with risks. As with human memories, personal and family stories change over the years, and most forms of media decay over time. However, digital media defaults towards persistence, which can impede the important roles that forgetting (Mayer-Schonberger 2009) and evolving memories can play (Pasupathi 2001). Generative ghosts might disrupt the natural tendency for loved ones to recall the deceased through “rose-colored glasses” (i.e., the tendency to focus on positive memories and attributes) by persistently reminding mourners of particular negative events or character traits.

Security Risks

Generative ghosts introduce several security risks, including post-mortem identity theft, hijacking, and malicious ghosts.

Generative ghosts may inspire new twists on crime, such as variants of identity theft. Ghosts present more than previously documented security concerns with online data after we die (Locasto, Massimi, and DePasquale 2011). Identity thieves may be able to interact with generative ghosts and use innocuous or jailbreak prompts to cause the ghost to reveal sensitive information and even raw data (Nasr et al. 2023) that might be used for direct financial gain (e.g., passwords, financial information, the location of valuables) or for indirect financial gain (e.g., sensitive personal information about the deceased or the living that might be used in blackmail schemes, phishing attacks, etc.).

While thieves might only aim to extract useful information, some criminals may engage in hijacking attacks, wherein they take control of a generative ghost. Hijacking attacks might take several forms. One possibility is a ransomware-style attack in which hijackers disable access to a generative ghost until mourners pay them a bounty. Hijackers may also surreptitiously change the functionality of a generative ghost in order to manipulate or harass mourners; this might be accomplished through a variety of methods including modifying source code, prompt injection attacks via the conversational interface, and/or puppetry attacks in which people believe they are chatting with the ghost but are instead chatting with a hijacker posing as the ghost or with a distinct bot the hijacker has substituted for the original generative ghost.

While identity theft and hijacking represent attacks by nefarious third parties, a first-party security risk is the creation of a malicious ghost. Malicious ghosts are first-party generative ghosts whose creators explicitly design them to engage in unpleasant or criminal activities. For example, an abusive spouse might develop a generative ghost that continues to verbally and emotionally abuse their surviving family members. In addition to ghosts that might engage in post-mortem harassment, stalking, trolling, or other forms of abuse of the living, malicious ghosts might be designed to engage in illicit economic activities as a way to earn income for the deceased’s estate or to support various causes including potentially criminal ones.

Socio-cultural Risks

Finally, we note that generative ghosts may introduce socio-cultural risks. While extremely important, this category of risk is also particularly speculative as it is difficult to anticipate how novel technologies may impact society at scale. For example, widespread adoption of generative ghosts might cause profound changes to cornerstones of modern society such as the labor market, interpersonal relationships, or religious institutions. While it is difficult to anticipate such societal impacts, even recognizing the possibility of systemic societal change without being confident of its precise direction can help us to be more thoughtful in the design of novel technologies.

A property of generative ghosts (vs. other types of griefbots that might be limited to simple chat functionalities) is their agentic nature, including the ability to execute actions on behalf of the deceased. For example, while on an individual scale we might view the ability of a generative ghost to earn an income that supports surviving family members (or other causes important to the deceased) as beneficial, at a societal scale this may impact the economy in complex ways that are difficult to anticipate. Widespread economic activity by generative ghosts might impact wages and employment opportunities for the living. Reliance on generative ghosts in some sectors of the economy might result in cultural and economic stagnation if such agents are not capable of develo** creative ideas in the same manner as the living or remain anchored to ideas or values from the past.

Another example of how generative ghosts may impact socio-cultural practices is in how they may alter interpersonal relationships and attendant social structures. For example, in the past two decades we have seen profound changes to social practices resulting from the widespread adoption of novel technologies such as smartphones and social media. The age of AI agents is likely to result in similarly disruptions of status quo socialization and relationship patterns (Shavit et al. 2023), and generative ghosts may be more disruptive to existing relationship patterns than general-purpose agents because they are modeled on specific loved ones rather than fictional characters or archetypes, possibly making them more compelling and potentially more likely to substitute for traditional relationships.

Because rituals and beliefs around death are often intertwined with religion (Walter 1994), we anticipate that the widespread adoption of generative ghosts might change religious practices, potentially in profound ways. Possible impacts might include the evolution of practices within a particular religion (i.e., updated rituals), the dissolution of existing religions (i.e., if the existence of AI afterlives alters people’s faith in a particular dogma), or the creation of novel religions (i.e., particularly religions that might deify generative ghosts and/or religions that might confer special status upon people with the technical skills to create AI afterlives). We anticipate that major world religions might issue guidance regarding the use of generative ghosts, and perhaps even offer customized versions of such technologies that are modified to support interaction styles and interface metaphors aligned with particular belief systems.

Discussion

In this paper, we introduced the concept of generative ghosts, agentic griefbots that may execute a variety of behaviors and actions on behalf of a deceased individual. Although digital memorials and simple griefbots have existed for years, generative AI is fundamentally changing the nature and scale of postmortem representations, making generative ghosts a distinct and novel source of AI ethics concerns and a likely cause of social change. We defined key design dimensions of generative ghosts, and identified potential benefits and risks of this new paradigm of post-mortem representation. Here, we reflect on the importance of careful partnership between the AI and HCI communities to design generative ghosts in a manner that mitigates risk, discuss the need to develop policies that protect values around privacy, consent, and safety, discuss the complexities of foreseeing the societal impacts of “AI Afterlives,” and identify key avenues for future research.

Interfaces to Mitigate Risk

Earlier in this paper we introduced several design dimensions for generative ghosts. Choices regarding these design dimensions may influence the benefit/risk landscape of this emerging technological paradigm.

Careful attention to the interaction design and interfaces of generative ghosts is vital for ensuring that these systems empower representees and the bereaved, and maximize the likelihood of socially beneficial outcomes over risky ones. This includes investing in user studies and social science research to more carefully understand aspects of the design space, such as what interfaces and interactions increase anthropomorphism risks (Abercrombie et al. 2023; Shanahan 2023; Gabriel et al. 2024), and what factors (i.e., features of generative ghosts, contexts of loss, attributes of the bereaved) may contribute to mental health risks from engaging with such technologies.

Whether a ghost is designed to act as a “reincarnation” of or “represents” the deceased is a particularly important feature of the design space to consider. Designers will have to make choices about whether a ghost speaks as the representee, assuming their voice (which some prior work (Brubaker, Hayes, and Dourish 2013) and case studies (Newton 2016) suggest can be unsettling) vs. speaking about the representee (ostensibly from the point of view of the system or bot). Embodiment can present similar issues, particularly if the ghost’s physical design resembles the representee.

Yet we do not mean to suggest that there are clear right and wrong approaches for each feature in our design space. After all, a ghost that reincarnates a spouse is likely experienced quite differently than one that reincarnates a historical and/or public figure such as Abraham Lincoln. We suspect that reincarnation-style or embodied interfaces for historical figures carry less risk than for personal remembrances. Characteristics of the representee, particularly in relationship to the user, will ultimately have profound impacts on how users experience ghosts. However, interaction design can play an important role in framing these experiences.

Given the sensitive nature of generative ghosts, it is particularly important to be vigilant against the use of dark patterns (Narayanan et al. 2020) in the design of these technologies, to minimize the risk of, for instance, the formation of addictive parasocial relationships that might harm mental health. For instance, one might consider what the equivalent of undesirable “push notifications” might be for an agentic system such as a generative ghost – perhaps ghosts should only respond to interactions initiated by the living rather than initiating interactions.

Generative ghosts might even include interface designs to proactively guard against likely harms. For instance, ghosts might monitor the patterns of interaction with the bereaved and analyze these interactions for patterns of overuse or other types of harm; in such cases, the systems might initiate interventions such as offering referrals to mental health professionals, reducing its own fidelity, reducing the hours during which it is available, etc. The types of use patterns that might prompt concern and the appropriate system responses are important areas for future study.

In addition to the interfaces and interactions for the agents themselves, carefully designed interfaces for creation and ongoing management of generative ghosts are vital for mitigating risk. For example, such interfaces may be part of end-of-life or estate planning processes, and need to clearly convey to representees and their survivors policies around data governance, under what conditions a generative ghost should be terminated (and what happens to the associated data and model), and actions a ghost can and cannot take on behalf of the deceased (i.e., can a generative ghost continue to perform paid labor on behalf of the deceased in their chosen profession? can it represent the deceased in legal disputes, such as about estate handling? can it participate in managing trusts or donor-advised-funds on behalf of the deceased? can it be consulted regarding end-of-life decisions if the representee is medically incapacitated?). Similarly, interfaces to support data capture and curation are vital both for privacy considerations (i.e., the ability to review and delete sensitive data so that it is not part of a ghost’s model) as well as to improve the fidelity of generative ghosts (i.e., by supporting the capture of latent data that exists only in the representee’s knowledge rather than in existing media).

Finally, as with any new media or technology, our experience of generative ghosts may change over time as we gain “literacy” in this new medium. For instance, we may find that anthropomorphic ghost designs will be less likely to cause harm over time as end-users become more savvy about understanding this interaction metaphor.

Policies to Mitigate Risk

Third-party generative ghosts may introduce complex ethical and legal concerns regarding privacy and consent. In addition to potentially violating the desires and privacy of the deceased individual, some classes of third-party ghosts (i.e., those developed by fiscal connections or unconnected entities) may offend or distress surviving family and friends of the deceased. For example, concerns that Hollywood studios might use hypothetical future AI resembling our concept of generative ghosts was among the causes of the 2023 SAG-AFTRA strikes (sagaftrastrike.org). Policy and governance around third-party ghosts may be prudent, specifically when it comes to who is allowed to create ghosts, who is allowed to be represented, and for what purposes. For example, it might make sense to adopt different policies for private individuals vs. public figures. Distinctions between policies regarding generative ghosts of distant historical figures and public figures whose death is more recent may be important as well, particularly since the latter may have living relatives who may be distressed by unauthorized third-party ghosts. For instance, in January 2024 a fan of the late comedian George Carlin (who died in 2008) created an unauthorized comedy special called “I’m Glad I’m Dead” using AI technology to mimic Carlin’s voice and persona; Carlin’s surviving daughter was highly distressed by this incident (St. Martin 2024), which raised complex ethical and legal issues.

Of course, status-quo mortality technologies also raise many concerns regarding consent, privacy, and data governance (Kneese 2023). However, these issues are likely to be amplified by the rapid pace at which generative AI technologies are progressing (which makes it difficult for policymakers to keep up with, much less proactively anticipate, risks) and by the large amounts of highly personal data such models will by necessity access in order to create a high-fidelity agentic representation of a deceased individual. Policymakers may need to create frameworks that address questions such as under what (if any) circumstances are third-party ghosts permitted (e.g., for public figures), if and how representees or their survivors can terminate a generative ghost should they change their mind, and what, if any, obligations hosting services have to provide models or data to representees or their survivors in the event of service termination (i.e., due to discontinued products (Kneese 2019), failure of an estate to pay for ongoing services, etc.).

It may be important to have a “kill switch” – i.e., a secure and reliable override for temporary or permanently disabling them in the event of malicious third-party events such as hacking or first-party events such as a generative ghost that has been programmed to harass the living.

Societal Impacts

It is difficult to anticipate the complex societal changes that may result from technologically and psychologically powerful innovations such as generative ghosts. It is unclear, for instance, what the adoption curve of generative ghosts may be, and what factors might influence this adoption curve; in addition to typical factors influencing technological adoption, such as features and costs, social factors such as network effects are likely important, as well as factors such as changing legal frameworks, religious edicts, and other considerations that are difficult to predict. Generative ghosts present an exemplar of why not only the capabilities of novel AI systems, but also their effects in the context of individual human-AI interactions and also their potential systemic, societal-level effects is necessary for robust safety evaluations (Weidinger et al. 2023).

The design choices of early or popular generative ghost services may influence their adoption (or rejection) by segments of society. For instance, third-party generative ghosts might be viewed more negatively than tools that support first-party creation, due to the privacy and consent concerns surrounding the former.

Other technological trends and developments may also influence likely adoption. For instance, if advances in machine learning improve factuality (i.e., reducing hallucination), thus reducing some reputational risks of generative ghosts, that may encourage adoption. If personal AI agents such as generative clones become commonplace, then transitions from pre-mortem to post-mortem agents may become appealing to a larger audience. If augmented reality or other metaverse technologies see widespread adoption for social and entertainment purposes, then embodied generative ghosts may be more widely accepted.

If generative ghosts become popular, their associated costs may have significant societal impacts. The economic costs of maintaining an increasing number of high-fidelity generative ghosts may become prohibitive, and the associated computational resources (and associated energy consumption) may also be cause for concern at scale, though such economic, storage, and environmental costs might be mitigated by future advances in technology. In addition to considering the cumulative costs at scale, the costs of such services for any given individual may also create new types of digital divides – it may be prohibitive for families of lower socioeconomic status to afford to create and/or maintain generative ghosts of their loved ones (or they may go into debt to do so if generative ghosts attain important cultural or religious significance).

Future Work

This conceptual paper takes an analytical approach toward identifying the design space of generative ghosts and the associated benefits and risks of this paradigm. As generative AI technologies progress (i.e., improvements in model capabilities, reductions in cost), it will become increasingly feasible to prototype high-quality generative ghosts, which will enable user studies with various stakeholders (potential representees and their loved ones, clergy, legal experts, etc.) to gain insight into user responses to particular implementations, and to understand patterns of use. In advance of building prototype systems, we recommend investment in research that gathers additional reactions to and requirements for generative ghosts from key stakeholder groups via interviews, surveys, and/or participatory design techniques. Because practices around death and dying are highly culturally dependent, there is a strong need for investigations of generative ghosts to engage with diverse user populations beyond the traditional “WEIRD” (Western, Educated, Industrialized, Rich, and Democratic) demographic. Including samples that vary along dimensions such as age, religion, and nationality is particularly important.

This paper identifies an emerging trend (creation of generative ghosts) that is likely to scale quickly due to (1) the fast pace of advances and availability of powerful generative AI models that will empower tech-savvy individuals to create generative ghosts and (2) the emergence of start-ups that will provide ghost-creation services to a broader consumer market. In addition to describing an emerging phenomenon, this paper is a call-to-action to scholars to collect empirical data that can be used to understand more deeply how choices in the design of generative ghosts (i.e., our framework of design dimensions) might impact outcomes (i.e., our risk/benefit analysis). The benefits and risks of generative ghosts will likely arise from the intersection of these design dimensions and contextual factors (such as personal characteristics of the bereaved, length of time since the deceased has passed, etc.). As such, contextual factors are also an important aspect of future experimental designs. The findings of such studies can then be used to provide design guidelines that can steer such technologies in beneficial directions, as well as provide information that can support policymakers and other stakeholders in understanding what societal frameworks may be needed to adapt to this phenomenon.

Conclusion

We introduced the concept of generative ghosts – agentic, AI-powered representations of deceased individuals. Using an analytical approach, we proposed a set of design dimensions of generative ghosts (provenance, deployment timeline, anthropomorphism paradigm, multiplicity, cutoff date, embodiment, and representee type), and reflected on how combinations of these choices might introduce both benefits and risks for the representee, their loved ones, and society at large. Our paper highlights a growing trend of using AI systems to represent the deceased; our design taxonomy contributes a framework that can help scholars and policymakers understand how this trend may evolve, and our analysis of potential benefits and risks provides an agenda for researchers and policymakers to conduct studies and collect data that will offer empirical evidence to understand how these design dimensions (as well as other contextual considerations) might relate to risk/benefit trade-offs. This paper sets the stage for a future research agenda bringing together interdisciplinary experts in AI ethics, ML models, Human-AI Interaction, Policy, and other cultural and religious institutions to study and design AI afterlives that balance the benefits and risks of this emerging sociotechnical paradigm.