PaperlayerView exact revision on arXiv

Source-linked AI summary

Beloved Afterlives: Governing AI Resurrection Beyond Consent

Hanjing Shi, Dominic DiFranzo

arXiv:2609.29544v1cs.CYcs.HC

TL;DR

AI resurrection raises a governance problem beyond whether consent existed at creation: authority can shift as representations change, circulate, and depend on providers. A 93-system public-record audit and four public cases reveal distinct gaps across system types and motivate relational authority as an accountability chain.

  • Problem

    Consent framed only at creation leaves unclear how authorization remains legible as representations change models, hands, providers, and audiences.

    Source: §1 · p. 1, §7 · p. 16

  • Method

    The paper audits 93 public systems across three analytical positions and compares four public cases tracing participation, creation, circulation, and contestation.

    Source: §1 · p. 2

  • Results

    Human systems show consent becoming incomplete over time, companion-animal systems require authority allocation without subject consent, and adjacent systems reveal portable representational capacity.

    Source: §9 · p. 20

  • Takeaways & Limitations

    Relational authority reframes governance as an accountability chain linking creation authority, source boundaries, circulation, contestation, preservation, and exit.

    Source: §9 · p. 20

    Public documentation cannot verify private onboarding, live model behavior, moderation, deletion, security, or support outcomes.

    Source: §8 · p. 19

Abstract

from arXiv · show

AI resurrection is often framed as a question of consent: did the represented person authorize being made to speak? That question matters, but it freezes authority at the moment of creation. A representation can later change models, pass to relatives, depend on a provider, incorporate records shared with others, or circulate far beyond its intended audience. We argue that the central governance problem is therefore not whether authorization exists once, but whether it remains legible as the representation moves. Across a public-record audit of 93 systems, creation was far easier to inspect than the conditions for speaking, contesting, preserving, or leaving: consent or authority information was thin in 82 systems, objection or redress in 82, and deletion or export in 77. The differences among systems reveal why these gaps cannot be reduced to one transparency score. Human afterlives show consent becoming incomplete over time. Companion-animal afterlives begin where subject consent is unavailable and shared care must allocate authority. Adjacent persona and mimetic systems show how voices, likenesses, and personalities can travel into later afterlife uses. Four public cases follow the same movement from premortem participation, through intimate postmortem creation, to third-party circulation and family contestation. From this evidence we develop relational authority: authorization is distributed across people, records, providers, and audiences, and must remain traceable as those relations change. This reframes AI resurrection from a product authorized once into an accountability chain linking creation authority, source boundaries, circulation, contestation, and exit. The study measures what users and affected parties can inspect publicly; private implementation and lived outcomes remain open empirical questions.

1 Introduction

AI afterlives can become consequential within existing relationships even when users know the representations are artificial. The paper asks how authority remains legible as representations move beyond creation, and develops relational authority and an accountability chain in response.

1 Introduction

  • Representations can occupy a relational place for mourners even when they know the replies are generated, shifting the question beyond deception to permission, attributed speech, persistence, and exit.Affected parties may include siblings, former partners, co-caregivers, and estate representatives; the public record may be their only common evidence.Source: §1 · pp. 1–2
  • Prior research examines digital remains, griefbots, generative ghosts, and mimetic systems, but does not establish what users and affected people can inspect across a product corpus.The cited work includes Brubaker et al. (2014), Harbinja et al. (2023), Hollanek and Nowaczyk-Basińska (2024), and Morris and Brubaker (2025).Source: §1 · p. 2
  • The study asks what authority, source, contestation, deletion, export, and exit information is publicly inspectable, how gaps differ across system types, and what controls follow.Its audit distinguishes human afterlives, companion-animal systems, and adjacent enabling systems, then uses four public cases to examine changing relationships.Source: §1 · pp. 2–3
  • Relational authority distributes authorization across people, records, providers, and audiences, while affected-party standing identifies who may seek notice, review, correction, or selective removal.The accountability chain links creation authority to source boundaries, circulation checkpoints, contestation, preservation, and exit.Source: §1 · p. 3

2 From Digital Remains to Beloved Afterlives

Digital afterlife research moves from stewardship of records toward interactive, generative presence, raising questions about who may authorize new attributed acts. The paper situates those questions in existing relationships, provider power, and the limits of public documentation.

2.1 Postmortem data became interactive presence

  • Generative systems can turn preserved records into first-person replies, new spoken words, or avatars in new scenes, shifting the issue from storage to authority over new attributed acts.Prior work includes Harbinja et al. (2023) and Hollanek and Nowaczyk-Basińska (2024).Source: §2.1 · p. 3
  • Pre-recorded, retrieved, scripted, and open-ended outputs differ in their relation to sources, while premortem participation may leave inheritance, provider changes, and shared records unresolved.The paper uses “beloved afterlives” for representations tied to deceased humans or companion animals and encountered through an existing relationship.Source: §2.1 · pp. 3–4

2.2 Continuing bonds do not settle platform authority

  • Continuing bonds can support remembrance, but a commercial AI product adds a provider that determines upload options, output, retained evidence, and who can end the interaction.The cited literature cautions against assuming continued interaction is inherently unhealthy.Source: §2.2 · p. 4
  • A public-document audit can identify visible support and exit features, but cannot determine whether a representation comforts, unsettles, isolates, or supports a user.Companion systems may carry emotional weight, but that does not authorize transforming a caregiver’s account into an animal’s voice without making the transformation inspectable.Source: §2.2 · p. 4

2.3 Public documents are an accountability surface

  • Public inspectability measures what people can verify without an account, private upload, or support exchange; it is distinct from a provider’s internal controls.Public materials may inform prospective users, relatives seeking takedown routes, regulators, and journalists, without verifying implementation.Source: §2.3 · p. 5
  • The legal scan identifies possible control points rather than determining compliance, and the paper frames existing instruments as not beginning from the complete social arrangement of a beloved afterlife.The scan includes data-protection, synthetic-content, digital-replica, and memorial-chatbot sources.Source: §2.3 · p. 5

3 Method

The study audits 93 publicly documented systems across three analytical positions, using a separate evidence base for scholarship and legal context. It combines structured coding with four public cases while limiting its claims to what public records show.

3.1 Evidence layers and study object

  • The audit treats human afterlives, companion-animal systems, and adjacent persona or mimetic tools as distinct positions, and inspects public product materials rather than pooling scholarship or legal sources into the product denominator.The product corpus is the only denominator; 91 scholarly sources and 16 legal or policy documents provide context.Source: §3.1 · pp. 5–6

3.2 Retrieval, inclusion, and corpus closure

  • Fixed query families and inclusion rules yielded 29 human afterlife services, 24 static memorial systems, 17 companion-animal systems, 17 persona or companion platforms, and 6 mimetic services.Corpus closure bounds the retrieved corpus rather than making it representative of an international market; retrieval relied mainly on English search terms and pages.Source: §3.2 · pp. 6–7

3.3 Coding and decision rules

  • One primary coder assessed authority, verification, source grounding, labeling, redress, deletion, export, shared-care conflict, and support-related fields using a codebook and decision rules.“Thin public information” means public readers cannot reconstruct the process with confidence; it does not mean an internal control is absent.Source: §3.3 · p. 7
  • After removing 20 low-confidence systems, the direction of the main findings remained, but the audit did not establish intercoder reliability.Among the remaining 73 systems, authority information was thin in 64 and objection or redress in 62.Source: §3.3 · p. 7

3.4 Four-case public-record comparison

  • Four cases were selected to vary premortem participation, creator relationship, provider involvement, generative scope, circulation, and visible objection, tracing authority through distinct movements.Dadbot and Eternos begin with premortem participation; Project December with intimate postmortem creation; Qiao Renliang videos with third-party creation and public circulation.Source: §3.4 · p. 7, Figure 1 · p. 8
  • Case records tracked creators, sources, circulation, affected parties, contestation, persistence, and exit; public reporting may omit private discussions, later deletion, or technical change.The analysis distinguishes a reported removal request from a verified platform-wide outcome.Source: §3.4 · p. 8

4 Findings from the 93-System Audit

The audit finds distinct public-information gaps across human, companion-animal, and adjacent systems rather than one uniform transparency deficit. Across the findings, creation is more visible than continuing authority, source grounding, contestation, or exit.

4.1 The three analytical positions reveal different pressure points

  • In 53 human systems, consent or authority information was thin in 51; all 17 companion-animal systems were thin, while adjacent systems were thin on source grounding in 19 of 23.Human deletion or export was thin in 45 of 53; adjacent deletion or export was thin in 20 of 23.Source: §4.1 · p. 8, Table 2 · p. 9

4.2 Creation authority is visible before refusal authority

  • Represented-person verification was thin in 65 systems, and family or co-caregiver conflict visibility contributed to thin consent or authority information in 82.Public pages commonly gave creators a specific workflow while leaving affected parties with a generic contact channel or no clear route.Source: §4.2 · p. 9

4.3 Presence claims outpace source grounding

  • Public grounding information was thin in 44 systems; this matters especially for generative speech, which may produce plausible first-person claims without showing their source.A synthetic label can disclose that content is generated without distinguishing recorded, retrieved, or invented speech; the audit does not show that weakly documented systems hallucinate more or users believe them.Source: §4.3 · pp. 9–10

4.4 Contestation and exit lag behind creation

  • Objection or redress information was thin in 82 systems, while deletion or export was thin in 77, leaving affected parties with less specific routes than creators.Export, deletion, billing termination, transfer, and objection serve different purposes; hosted conversational systems may lose interaction functionality even when recordings are retained.Source: §4.4 · p. 10

4.5 Companion-animal systems reveal distributed authority

  • None of the 17 companion-animal systems described handling caregiver disagreement, although shared records and caregiving histories may underlie generated speech.The animal cannot authorize speech or correct a false memory; uploader permission therefore does not settle shared-media separation or notice to another caregiver.Source: §4.5 · p. 10

4.6 Support language is easier to find than escalation guidance

  • Among support-framed systems, 40 did not pair those claims with an explicit crisis resource, and age or life-stage information was thin in 17 systems.The audit cannot infer crisis access, worsened distress, or minors’ use without protection; it measures public features, not lived outcomes.Source: §4.6 · pp. 10–11

5 Four Public Cases

Across four cases, authority shifts with participation, source grounding, platform dependence, and circulation; the cases do not form a simple good-to-bad scale. Premortem participation clarifies creation authority but does not settle continuity, while public circulation brings new affected parties and contestation.

5.1 Dadbot: participation can bound a representation

  • Dadbot’s recorded interviews and audio made its source grounding comparatively inspectable, but public materials did not establish formal revocation, inheritance, deletion, or continuity procedures.Its bounded structure could make some absences visible, whereas smoother open-ended generation can obscure statements that exceed the record.Source: §5.1 · p. 11

5.2 Michael Bommer and Eternos: self-authorization enters a platform

  • Bommer’s direct participation established strong creation authority, while commercial voice synthesis and external models made new responses’ relation to his archive less visible.Claims that control could pass to relatives or that his wife could delete the representation depended on provider description and reporting, not an inspected workflow.Source: §5.2 · pp. 11–12

5.3 Project December: intimacy supplies authority that the deceased did not document

  • Barbeau’s intimate access to Pereira’s messages supported generative simulation, but the reviewed public record did not establish her premortem authorization for this use.Generated statements could extend beyond her records, with no output-level source record to distinguish remembered phrases, prompt instructions, and model invention.Source: §5.3 · p. 12

5.4 Qiao Renliang videos: public circulation produces contestation

  • The Qiao Renliang videos involved third-party creation and public circulation; his father said the family had not been consulted and wanted the video removed.Reports of removal did not verify platform-wide removal, re-upload prevention, evidence preservation, or a formal appeals process.Source: §5.4 · p. 13

5.5 Cross-case finding: authority changes with circulation

  • Across the cases, premortem participation clarifies creation authority but leaves inheritance, model change, recipient choice, and provider shutdown unresolved.The same audit field can conceal very different authority bases, so the cases restore relationships without serving as prevalence evidence.Source: §5.5 · p. 13

6 Keeping Authority Legible from Creation to Exit

The proposed controls make relational authority traceable as representations move from creation through use, contestation, transfer, and exit. They distinguish authority records, source boundaries, review routes, and forms of leaving, while accommodating shared care.

6 Keeping Authority Legible from Creation to Exit

  • Relational authority becomes governable when a record follows each transition, linking creation authority, source boundaries, circulation, contestation, and exit.The proposed controls rebuild the chain at the same public boundary measured by the study; their effects remain questions for evaluation.Source: §6 · p. 14

6.1 Record authority before creation

  • Authority records should identify the creator’s role, the basis for the upload, represented-person participation, and whether another person may seek review.Affected-party standing requires providers to accept objections and explain review outcomes, but does not give every affected person an automatic veto.Source: §6.1 · p. 14

6.2 Separate recorded, retrieved, and inferred speech

  • Public materials and output cues should distinguish replayed, retrieved, scripted, and newly generated speech, including when responses may extend beyond the record.These controls support inspection but do not guarantee truth or emotional safety.Source: §6.2 · pp. 14–15

6.3 Make objection a workflow

  • An objection becomes contestation when a documented route connects a claim to review and possible remedy, while preserving evidence before removal.Selective remedies can address disputed media or voices without automatically deleting an entire memorial or allowing unrestricted persistence.Source: §6.3 · p. 15

6.4 Treat export, deletion, transfer, and exit as distinct

  • Export, deletion, transfer, and interactional exit require separate documentation, including what is retained, who can inherit access, and whether recipients can decline.Pause, silence, billing cancellation, and archiving are not equivalent; providers can also disclose what happens if a service closes or changes models.Source: §6.4 · p. 15

6.5 Design for shared care

  • Companion-animal systems should record uploader relationships, allow identification of other caregivers, and separate source media from memorial text and generated speech.The same shared-care architecture can help human families review shared records without stretching the represented subject’s consent to cover every contributor’s data.Source: §6.5 · pp. 15–16

7 Discussion

The discussion reframes AI resurrection as a moving distribution of authority and argues that public accountability must follow representations across sources, custodians, audiences, contestation, and exit. Its audit and cases show why initial consent and broad transparency labels cannot by themselves make that authority inspectable.

7.1 Public inspectability is where authority becomes governable

  • Public inspectability should let affected people trace a claim from authority evidence to review and remedy, not merely find a disclosure or contact channel.The needed information changes across creation, interaction, circulation, and exit; a deletion policy, for example, may not say whether shared recordings can be preserved.Source: §7.1 · pp. 16–17
  • Consent or authority was thin in 51 of 53 human systems, objection or redress in 49, and all 17 companion-animal systems lacked caregiver-disagreement procedures.Adjacent systems were thin on source grounding in 19 of 23 and deletion or export in 20; one transparency score would obscure these differences.Source: §7.1 · p. 16

7.2 Consent begins authority; it does not finish it

  • Cases show that even strong premortem participation did not govern an entire afterlife, while pet systems defaulted to account control despite shared care.None of the 17 animal systems made caregiver allocation contestable.Source: §7.2 · p. 17
  • Affected-party standing requires providers to receive claims, preserve disputed evidence, identify the authorization basis, and explain a remedy without granting every claimant an equal veto.Potential grounds include a person’s image, a shared photograph, family responsibility, or direct exposure to a circulating artifact.Source: §7.2 · p. 17

7.3 Fluency can conceal the edge of the person

  • Dadbot’s authored branches and recordings left a visible path to participation, while Project December expanded a comparatively small record into a broad conversational presence.A familiar voice can make a thin source record feel complete; provenance, uncertainty cues, and non-response can keep its limits visible.Source: §7.3 · p. 18

7.4 Circulation changes who can make a claim

  • Circulation brings new claimants and evidence needs; a synthetic label does not establish authorization or whether generated words were grounded in the represented person’s speech.The Qiao case involved relatives, fans, platforms, commercial promotion, and possible re-uploads, each changing what evidence mattered.Source: §7.4 · p. 18

7.5 Relational authority turns disclosure into an evidence chain

  • The proposed evidence chain links authority records and source boundaries to traveling provenance, affected-party review, and distinct options to export, pause, transfer, or delete.A break at any stage weakens the authority carried into the next.Source: §7.5 · pp. 18–19
  • Public documentation cannot verify private safeguards; future work should test provider responses and examine research consent, withdrawal, and quotation practices.The authors propose field audits of objections, corrections, export, and deletion, and recommend consent before combining living participants’ relational records.Source: §7.5 · p. 19

8 Limitations and Future Work

The authors bound their audit as a map of public evidence rather than a measure of global coverage, operational practice, or lived effects. They propose multilingual replication, independent coding, deeper case studies, and longitudinal research to address these limits.

8 Limitations and Future Work

  • The corpus overrepresents English-language retrieval, stable pages, and publicly documented services, and may miss small, discontinued, app-only, and region-specific systems.A multilingual replication could test whether the analytical positions travel across regions and legal and cultural settings.Source: §8 · p. 19
  • Public documentation captures pre-use promises and routes but cannot verify private onboarding, live model behavior, moderation, deletion, security, or support outcomes.Standardized field requests could test evidence demands, pauses, response times, and remedies after model or provider changes.Source: §8 · pp. 19–20
  • One primary coder applied the codebook, and family assignments, thin-information thresholds, and generic contact-channel interpretations still involve judgment.Removing 20 low-confidence systems preserved the headline direction but did not estimate intercoder reliability.Source: §8 · p. 20
  • The four public cases are analytic contrasts, not prevalence evidence or ethnographies, and their records may omit private authorization, family disagreement, deletion, and later system changes.The audit also cannot establish dependence, distress, healing, closure, or effects on families and caregivers.Source: §8 · p. 20

9 Conclusion

Across 93 systems and four public cases, the paper finds that authority becomes harder to inspect as representations move beyond creation. It proposes relational authority as an accountability chain that follows representations across changing sources, custodians, providers, and audiences.

9 Conclusion

  • Across 93 systems, creation was easier to inspect than subject verification, source grounding, objection, deletion, export, or exit.Human systems showed consent becoming incomplete over time; companion-animal systems began without subject consent, while adjacent systems showed representational capacities traveling into afterlife uses.Source: §9 · p. 20
  • Four cases trace movement from premortem participation to intimate postmortem creation and third-party circulation, where family objection exposed a missing review route.Dadbot and Eternos had premortem participation; Project December lacked documented authorization from the deceased; Qiao Renliang’s videos circulated publicly.Source: §9 · p. 20
  • Relational authority shifts governance from a creator–product transaction to a chain linking creation authority, source boundaries, circulation, contestation, and distinct exit paths.The chain requires authorization and provenance to travel with an artifact and affected-party review to preserve evidence and explain a remedy.Source: §9 · p. 20
Loading 2609.29544v1…