An interface to your story.

Start with the archive, then add conversation

A useful question is simple: What did my father say about his first business? A responsible response should locate his interview, show the date, and let the listener hear the actual passage. It may also offer a clearly marked summary. The proposed system would distinguish a quotation, a synthesis of several records, and information it cannot establish.

That distinction is central to the product. An answer that sounds emotionally convincing can still be wrong. NIST's Generative AI Profile identifies confidently false output, privacy risks, and problematic human reliance among the risks that need to be managed. Death X has not demonstrated that its planned interface adequately controls these risks. [1]

What a model can and cannot represent

Research on generative agents shows how software can simulate believable behavior using language models, stored observations, retrieval, and reflection. Park and colleagues demonstrated such agents in a simulated social environment. That work supports the feasibility of constructed behavioral representations; it does not show that a model contains the consciousness or personal memories of a real person. [2]

For Death X, an avatar would be an optional presentation layer over selected records. Its generated response would remain new software output even if it resembles your vocabulary or voice. It should never be added back into the source collection as something you actually said.

Three proposed ways to explore a life

Archive search would return original evidence with dates and context. Guided narrative would connect approved records into a clearly attributed account. Conversational representation would generate responses within a defined scope, displaying its synthetic nature throughout the experience. Each mode would have its own permission and evaluation requirements.

Families would not need to choose the most realistic interface. Some may prefer transcripts and recordings. Others may want a conversational guide without a cloned voice. The product should support those choices without treating increased realism as increased truth.

Permission must follow every representation

The planned consent model separates permission to store a recording from permission to create a synthetic voice, animate a likeness, allow particular audiences, or connect the interface to a physical device. A public interview should not automatically become permission to simulate its speaker. Rights involving other people in a recording must also be considered.

Users would need a clear way to restrict topics, choose who may activate an interface, and state whether synthetic responses are allowed after death. These are proposed controls; final availability and legal effect depend on implementation and the applicable agreement.

How we intend to measure quality

Evaluation would test whether answers cite the right records, preserve quotations, acknowledge missing information, and resist requests to invent unsupported life events. Reviewers would also test whether viewers understand they are interacting with software. A convincing voice would not compensate for poor source accuracy.

The near-term research objective is a private, reviewable archive assistant. A later objective is an optional, carefully governed representation that a participant can examine while alive. Death X is not offering an autonomous digital continuation or claiming that an avatar makes biological revival possible. Progress would be reported through measured results and documented limitations, rather than resemblance alone.

Sources & further reading

1. Autio et al. — NIST AI 600-1, Generative Artificial Intelligence Profile (2024) ↗
2. Park et al. — Generative Agents: Interactive Simulacra of Human Behavior, ACM UIST (2023), DOI 10.1145/3586183.3606763 ↗

Linked research supports the specific statements discussed. It does not validate the proposed Death X service. Sources checked 14 October 2026.