People want to keep the voice and stories of a parent or partner, and museums want future generations to "ask" the last Holocaust witnesses questions. Griefbot products record a person's interviews, messages and voice and turn them into a conversational avatar, either made before death as a life-story archive or built from the estate's material afterwards. The technical part is now easy; consent, post-mortem rights, honest labelling and protection of grieving users are the hard part.
Problem & value
- Today: families keep photo albums, voicemails and videos that nobody can query; stories die with the person. Museums run recorded testimony as passive video. Funeral providers look for memorial services beyond a gravestone page.
- AI value: an interviewer bot elicits life stories while the person is alive; afterwards, retrieval over the archive lets relatives ask questions and hear answers in the person's own voice.
- Evidence: the USC Shoah Foundation "Dimensions in Testimony" project lets visitors ask questions and get real-time answers from pre-recorded interviews; each witness answers around a thousand or more questions, and answers are the witness's own recorded clips [1]. Cambridge LCFI researchers (Hollanek and Nowaczyk-Basinska, 2024) analysed griefbot risks and argued for design protocols, consent and dignified retirement of such bots [2]. No reliable figures exist on grief outcomes; do not claim therapeutic benefit.
Variants
- Pre-mortem life-story archive (private): the person is interviewed over weeks, approves what the avatar may say, and names who may access it. This is the safest variant because consent is explicit.
- Post-mortem griefbot: built by relatives from messages and recordings of someone who left no consent. Highest ethical and legal risk; many providers and ethicists advise against it or want strict estate authorisation.
- Funeral and memorial SMB add-on: funeral homes offer a memorial page with an interview-based archive, a QR code on a gravestone or an annual remembrance session.
- Interactive testimony for museums and memorials: curated, pre-recorded answers matched to visitor questions (retrieval of real clips, no free generation). Fits cultural-heritage institutions.
- Generative persona: a fully generated voice or video avatar that can say new sentences. Most flexible, most risky; keep it clearly separate from the retrieval-only variant.
How it works
- consent + persona rules (person or estate, written)
- structured interviews (audio/video, interviewer bot or human) + document import
- transcription + diarization + metadata (date, topic, people, sensitivity tag)
- index (chunks + embeddings; every chunk linked to its source recording)
- dialogue: question
- retrieval
- answer (A) retrieval-only: play the best matching real clip (B) generative: LLM answers only from retrieved material, in the person's style
- optional voice (cloned TTS) / talking-head video
- guardrails (topic exclusions, no new facts, no advice, crisis handling) + AI label
- access control, usage log, retirement/deletion workflow
- Grounding: the key design choice. An answer must be traceable to something the person said or wrote; "I don't know / I never talked about that" is the correct fallback, not invention.
- Human-in-the-loop: the person (pre-mortem) or an authorised executor reviews sample answers and blocked topics; museum curators approve the clip set and the question-matching quality.
- Voice and face: a few minutes of clean audio suffice for cloning; limited archives, noisy old recordings and age of the voice reduce quality.
Data preparation
The archive is the product. Material comes from structured life-story interviews (audio/video), and from what families already hold: voicemails, letters and paper photo albums, e-mails and chat exports, social-media posts, old tape or phone recordings, often scattered across devices and cloud accounts. Typical gaps: noisy or low-quality audio, undated and unlabelled recordings, handwritten letters, messages that mention living third parties, and no record of what the person agreed to.
Preparation steps: record a written consent and persona rules first (PREP-025), then run interviews with a topic guide, scan paper material with OCR (PREP-006), convert old formats (PREP-007), transcribe with Whisper-class tools and tag each clip with date, topic, people and sensitivity (PREP-010, PREP-019). Pseudonymise or exclude passages about third parties and health or religion (PREP-022). Create a test set of typical family questions with the person or executor (PREP-018) and keep every chunk linked to its source (PREP-013). Tools: Whisper, Paperless-ngx, pgvector/Qdrant.
Rough share of effort: about 30-40 percent; the interviews and tagging take longer than the technical setup. GUIDE-003 gives a checklist for the inventory.
Implementation paths
Which path for whom
- private: A, because a ready-made interview service needs no technical skill; verify export and deletion rights first, since several providers are unstable.
- smb: A for funeral homes and similar providers who resell a memorial service, with contractual export rights and insolvency protection; B only for an own branded product.
- public: C, because museums and memorials must keep long-term custody of very sensitive testimony, avoid vendor lock-in and decide themselves how answers are generated.
Path A — Buy (recommended for private users and funeral SMBs)
- Legacy-interview services: HereAfter AI (recorded interviews, voice avatar answers from the person's own recordings; a closure notice on its website was reported in July 2026, with users asked to request their recordings by e-mail; treat the service as discontinued), Eternos (voice and personal AI built from user-supplied data), StoryFile (interactive video conversations, used by museums; StoryFile filed for Chapter 11 in May 2024 [8], so check its current status and data-custody terms before relying on it). Project December and similar services generate text personas, which is the high-risk generative style.
- Museum projects: typically bespoke production (green-screen capture, question list, matching engine) via heritage specialists, as in the USC project [1].
- Check: EU hosting, deletion and export on termination or provider insolvency, explicit consent flow, labelling, who may access, price model (subscriptions for memorials create long-term lock-in).
Path B — Build on cloud APIs
- Models: an EU-hosted frontier or Mistral LLM for grounded answers; ASR such as Whisper-class or vendor ASR; TTS and voice cloning from ElevenLabs (check data-residency options and its consent-based voice rules) or Azure custom neural voice (gated access with a recorded consent statement) [4].
- Stack: vector store (pgvector, Qdrant) with provenance metadata; avatar video through a talking-head API.
- Choose B when: a funeral provider or museum wants its own branded product with custom archive handling.
Path C — Self-host (open source)
- Models: open-weight LLM (Mistral, Qwen, Llama families) via vLLM or Ollama; Whisper for transcription; open TTS such as F5-TTS, Kokoro or XTTS (check licences, some are non-commercial); embeddings such as BGE-M3.
- Why: keeps the most intimate data of a family in their own control; no voice data leaves the premises.
- Hardware: HW2 for a 7-30B quantised LLM plus TTS; ops effort moderate. Video avatars need more GPU time.
Path D — Custom training
Fine-tune a voice model only for fidelity; retrieval beats LLM fine-tuning on a person's writing.
State of the art (2026)
- Retrieval-grounded personas with a cloned voice are standard; real-time voice-to-voice latency is under a second in good setups. Few-shot voice cloning works from short samples.
- Interactive testimony still relies on pre-recorded clips plus question matching because it protects the authenticity of what the witness actually said; generative extensions are debated in the heritage sector [1].
- Limits and failure modes: invented anecdotes that never happened, drift into generic chatbot tone, wrong emotional register, voice artefacts, the person's views on events after death being fabricated, and users who return daily and substitute the bot for mourning. Cambridge researchers highlight risks of unwanted advertising by the bot, distress for children and the "haunting" experience when a bot cannot be switched off [2].
Shared resources & combinations
- Same stack: voice cloning with SPEECH-007, avatar rendering with GEN-003, retrieval pipeline with KNOW-002, persona and companion safety patterns with CONV-023, CONV-009 and CONV-016.
Effort, cost & KPIs
- Path A: hours of interviews; subscription or one-off fee. Path B: 2-4 months plus legal and ethics review.
- Path C: 3-6 months; one GPU server.
- KPIs: share of answers traceable to a source, rate of "I don't know" versus fabricated answers (red-team review), consent coverage, family satisfaction surveys, healthy usage patterns (session frequency, no escalation), deletion SLA. Avoid engagement KPIs.
People & change
Involve the person themselves (pre-mortem) or an authorised executor first; their written consent and persona rules decide everything else. For funeral providers and museums, management and legal/compliance must settle post-mortem personality rights, estate law, contract terms and insolvency protection (GUIDE-010). The data protection officer reviews third-party data in messages, special-category data, processor contracts and retention; a DPIA is advisable for large archives (GUIDE-005). The museum curators or family own the content and approve the clip set.
Works council: no in the typical private and memorial setting, because no employee behaviour is monitored. It becomes relevant (depends) only if a funeral home or museum with a Betriebsrat or Personalrat logs staff use of the system; then §87(1) no. 6 and §90(1) no. 3 BetrVG apply (GUIDE-004).
AI literacy: staff who offer the service need to understand fabrication risk, labelling duties and crisis hand-over, and the Act asks for measures supporting literacy (GUIDE-006). Communicate openly to families what the avatar is and is not, always label it as AI, offer session limits and a retirement ritual, and involve grief counsellors (GUIDE-008). Start with a small pilot family or visitor group and red-team for invented anecdotes. Humans, the person or executor, review sample answers and can switch the avatar off at any time.
Risks, regulation & compliance
- EU AI Act: transparency under Art. 50: users must be told they interact with an AI system (Art. 50(1)), and synthetic voice, image or video that resembles a real person is a deepfake that must be disclosed (Art. 50(2) and (4)) [3]. Classification is normally limited risk. Art. 5 prohibitions apply as red lines: a bot using subliminal or manipulative techniques, or exploiting the vulnerability of grieving or elderly users, to materially distort behaviour and cause harm (for example, pushing purchases or subscriptions) is prohibited [3]. Do not infer emotions of users in workplace or school settings.
- GDPR: the regulation does not apply to deceased persons (Recital 27), but the archive contains personal data of living third parties (relatives, friends in messages), and special-category data such as health or religion [5]. Pre-mortem consent must be explicit, granular, revocable while alive. Many member states allow national rules on data of the deceased.
- Post-mortem personality rights: national, not harmonised. In Germany the general personality right continues after death in part and heirs can enforce it; voice and likeness rights, estate law and contractual terms must be checked per country. Get written authorisation from the person or executor.
- Psychological safety: crisis detection with hand-over to human help, session limits, no pretence of being alive, no "I love you" claims, no ads, and a ritual for retiring the avatar [2]. Involve grief counsellors.
- Data custody: insolvency, acquisition or price changes of the provider can switch off a loved one's voice; demand export rights.
Getting started
- Decide the variant: pre-mortem archive, museum testimony or post-mortem bot; start with the first or second.
- Draft the consent form and persona rules (allowed topics, blocked topics, who has access, retirement date).
- Record 5-10 hours of structured interviews with the person and tag each clip with topic and source.
- Prototype retrieval-only answers first; add cloned voice only after sample review by the person or executor.
- Run a red-team session for fabrications and emotional manipulation, then pilot with a small family or visitor group.
Sources
- USC Shoah Foundation — What is Dimensions in Testimony? — https://sfi.usc.edu/dit
- Cambridge LCFI — Deathbots, griefbots and the digital afterlife — https://lcfi.ac.uk/news-events/news/deathbots-griefbots-and-the-digital-afterlife
- Regulation (EU) 2024/1689 (AI Act), Art. 5 and Art. 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Microsoft — Custom neural voice limited access and disclosure — https://learn.microsoft.com/azure/ai-services/speech-service/custom-neural-voice
- Regulation (EU) 2016/679 (GDPR), Recital 27 — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Eternos — https://www.eternos.life
- CORDIS — Move over chatbot, here comes the deadbot — https://cordis.europa.eu/article/id/451347-move-over-chatbot-here-comes-the-deadbot
- AI Business — Startup behind AI William Shatner (StoryFile) files for bankruptcy (May 2024) — https://aibusiness.com/verticals/startup-behind-ai-william-shatner-files-for-bankruptcy