CASE FILE

Who owns your voice after you die?

Artificial intelligence can have the deceased read a new text, respond to the bereaved, or continue a role that no one has left to play. This turns a technically fascinating possibility into a question of consent, inheritance, identity — and responsibility for sentences one never uttered.

Kdo vlastní váš hlas poté, co zemřete?
Jiný Kontext editorial illustrationUmělá inteligence může nechat zemřelého přečíst nový text, odpovídat pozůstalým nebo pokračovat v roli, kterou už nemá kdo hrát. Technicky fascinující možnost se tím mění v otázku souhlasu, dědictví, identity — a odpovědnosti za věty, které člověk nikdy nevyslovil.
Listen
00:00/00:00
1.00 ×
Ready

In the documentary about Anthony Bourdain, a line that Bourdain actually wrote will be heard. His voice delivers it with familiar rhythm, color and weary intimacy. But the recording never existed. After Bourdain's death, she was created by an artificial intelligence model from hours of older footage. The viewer hears an author who is and is not an author at the same time.

Controversy surrounding the film Roadrunner from 2021 didn't come about because someone pushed someone else's opinion onto Bourdain. It was his own words. The problem was more subtle: the film created a new performance and made it seem like a found document. The director admitted that three passages were synthesized; directly in the film, however, the viewer could not distinguish them from the real archive.[1]

At first glance, a simple question presents itself: who had the right to use the voice? Family? Heirs? A producer who owns records? The company that made the model? Or no one, because the only person who could have said yes is no longer alive?

But ownership is perhaps the wrong first category here. A voice is not a car or a bank account. It is at the same time a physical sign, a method of performance, a means of identification, a bearer of relationships and a technical model that can be imitated mathematically. And the new sound can contain true words, a false situation, and perfectly true intonation — all at once.

Fidelity of sound is not fidelity of intention. Different Context

01 / The first resolutionVoice is not one thing

When we say "someone's voice," we normally conjure up several distinct layers. There is a biological sound apparatus. Recognizable color and accent. Learned way of phrasing. A specific acting or speaking performance. Audio recording saved in a file. And finally the meaning of the words that are spoken.

During life, these layers are connected by one body. A person chooses a sentence, takes a breath, utters it, and at the same time can bear social and legal responsibility for it. Synthesis disconnects them. One person supplies the historical recordings, another writes the text, a third sets up the model, a fourth publishes the output — and yet the listener intuitively attributes the entire message to the one he recognized by the voice.

This is why a digital replica can be more convincing than a regular imitation. It not only resembles the sound of the deceased. It uses an automatic shortcut that human communication has built up over a lifetime: I recognize the voice, so I recognize the speaker.

In addition, the voice does not only carry words. It carries age, fatigue, irony, uncertainty, social role and a sense of closeness. The same sentence read by a neutral synthetic voice can act as a reconstruction of the document; uttered by the voice of a parent or partner turns into a personal event. Therefore, the technical similarity does not only increase the realism of the sound. It also increases the psychological weight of content that could have been written by someone else.

But in a technical system, the "spokesperson" does not have to be one person. It is a chain of roles.

From the archive to a sentence that was never heard
Simplified technical model
The process of creating a synthetic voice of the deceased Five steps from source recordings to voice representation and new text to synthetic sound and interpretation by the listener. 01 Recordings Interviews, movies, voice messages. Past performance 02 Voice pattern timbre, rhythm, pronunciation, acoustics. Statistical trace 03 New content Text man or model response. New authorship 04 Synthesis The model converts the content in a similar voice. A new sound 05 Listener He recognizes a person and attributes the statement to him. Transferred trust During the process, the acoustic similarity turns into the impression of personal presence.
The diagram shows the division of responsibility, not the specific architecture of each product. Different systems may use different models and amounts of data.

02 / MechanismThe machine does not return the human. Adds its probability.

The VALL-E research system already showed in 2023 that personalized speech can be synthesized from a three-second voice sample in an experimental environment. The model was trained on sixty thousand hours of English speech and, according to the authors, was able to preserve not only the likeness of the speaker, but also part of the emotions and acoustic environment of the reference recording.[2] It doesn't mean that three seconds is always enough to reliably copy every person in every language. It means that the amount of material needed is no longer a barrier that most people can safely hide behind.

Technology does not search for the hidden essence of personality. It derives the representation of speech signs from the recordings and generates a sound corresponding to the new text from it. When the system also gets access to the family's emails, messages, videos and memories, it can also create a conversational style. The result then resembles not only the voice, but also favorite turns of phrase, humor or the way a person addressed their loved ones.

However, this is still not evidence of a continuing mind. The model cannot verify whether the deceased would change his mind today, what he would consider too private, or how he would react to an event that occurred after his death. He fills in the blanks according to probability — and the more convincingly he does it, the easier his guess is mistaken for an authentic answer.

An important distinction

Reconstruction of style can be technically very accurate. Reconstruction of intention remains an inference. The system can faithfully imitate how a person spoke, but it cannot demonstrate what they would like to say right now.

This is where the whole problem changes. The question is no longer just whether the recording is fake. It can be sonically artificial, textually authentic and situationally misleading. Or it may be transparently marked, written by someone else, but still act as a moral recommendation of a person whose authority has survived death.

03 / LawOwning a file does not mean owning future sentences

Let's imagine a daughter inherits her father's phone. He knows the password, has access to voicemails and can manage the family archive. It does not logically follow that he can let the father promote a political party, comment on a family dispute or sell insurance. Access to material and permission to create new speech are two different things.

Copyright knows the same difference. The owner of a physical medium or file does not automatically acquire all rights to the work and performance captured therein. The Czech Copyright Act deals separately with the work, the artistic performance and the right of the sound recording producer. At the same time, the recognizable color of the voice does not simply fit into one of these compartments.[4]

At least four areas of control may therefore overlap with synthetic voice: protection of the personality and dignity of the deceased, rights to original recordings and performances, terms of service or license, and responsibility for new output. Someone can legitimately hold the archive but not have the right to use it to train the model. Another may have a license to a movie but not an interactive character. And the person who writes a new line can be held responsible for its content, even if the sound resembles someone else's.

Additionally, these layers do not have to move together. The use may be permissible in terms of contract, but interfere with personality. He can respect the wishes of the family but violate the rights to the recording. It can be artistically legitimate and yet deceive the audience by concealing the mode of origin. Therefore, the legal question is usually not "whose vote belongs to whom", but "what authorization was needed for this particular step and who actually had it".

Four layers of control over one synthetic voice
Editorial legal map
A map of legal layers of synthetic voice Four overlapping areas: personality, recordings, contracts and new output. Personality and dignity Identity, reputation, pious interest, consent Recording and performance Work, performance, sound recording Dates and contracts Access, License, Purpose, Platform Rules New output Text author, operator, distribution, damage Synthetic sentence One sound multiple possible beneficiaries
The map simplifies the legal structure and does not replace the assessment of a specific case. The decisive factors are the purpose, the identifiability of the person, the contract, the origin of the data, the method of publication and the applicable law.

Therefore, it is not accurate to say that the voice is "inherited by the family" after death. The family can gain property rights, access to data or the ability to defend the memory of the deceased. He may not, however, acquire unlimited power to produce new opinions on his behalf. Inheritance deals with the transfer of certain rights. It does not automatically address the legitimacy of every future simulation.

04 / Czech and European frameworkProtection exists. It just doesn't lie in one place.

The European General Data Protection Regulation expressly states that it does not apply to the personal data of the deceased; Member States may adopt their own rules.[3] This creates a strange asymmetry. As long as a person lives, working with his voice data can be a matter of personal data protection, contract and personality rights. After death, the protective center of gravity moves elsewhere.

In the Czech Republic, the deceased does not remain without legal protection. The Civil Code protects a person's personality and stipulates that after his death any person close to him can claim protection. This is an important brake against interference with dignity, reputation or identity.[4] But it is not a separate, detailed institute of "digital voice ownership". The dispute would depend on the specific use: whether the person is recognizable, whether the simulation is deceptive, whether it is a work of art, a message, an advertisement, a private memory or a damaging statement.

From August 2, 2026, the AI ​​Act is generally applied in the European Union. Its Article 50 imposes an obligation on system users posting image, audio or video content that constitutes a deepfake to acknowledge that it has been artificially created or altered. For obviously artistic, creative, satirical or fictional works, the labeling can be done in a way that does not disrupt the experience. The information should be clear at the latest at the first contact with the content.[5]

This is a significant shift, but not an answer to the question of consent. The obligation to label a synthetic voice does not imply the authorization to produce it. Transparency addresses the relationship with the audience. Personal, copyright, contractual and possibly criminal law rules determine whether the use was allowed at all.

Europe and the United States often start at opposite ends
Comparison of legal logics
Comparison of European and American legal logic Europe emphasizes dignity, personality and transparency; part of the US states transferable property control over voice and form. EUROPEAN LOGIC Man as a personality Emphasis on dignity, context and audience deception Protection of personality Who is allowed to intervene in identity and memory? AI transparency Does the audience know they are hearing synthetic content? Rights spread across multiple regulations Without a single universal "voice owner" PART OF AMERICAN LOGIC Identity as property value Emphasis on controlling commercial use of the replica Right of publicity Who is allowed to profit from voice and image? Transfer and inheritance Some states expressly create a property right A mosaic of states and federal proposals Both scope and exemptions vary by jurisdiction
This is an indicative comparison, not a claim that Europe does not protect economic interests or that the United States does not protect dignity. The specific rules vary significantly between states.

05 / Second perspectiveWhen the voice becomes hereditary property

California took a more straightforward route. AB 1836 extended "deceased personality" protections to include highly realistic digital replicas of voice and likeness in audiovisual works and sound recordings beginning in 2025. Rights are proprietary, transferable and heritable under the California model; protection lasts up to seventy years after death. At the same time, the law contains exceptions for news, criticism, research, satire, parody or some documentary and biographical depictions, for example.[6]

This model has an advantage: it specifies the person who can grant the license and creates a specific claim against unauthorized commercial use. But at the same time, it turns identity into a tradable commodity. If the heir licenses the voice, it may be legally "settled" without being certain that the use is consistent with the character, values, or wishes of the deceased.

Therefore, the federal NO FAKES Act is also being discussed in the United States. The 2026 version of S. 4591 would have created a federal right to voice and visual digital replica and provided for its transition after death. However, as of August 15, 2026, the proposal was not effective law: The Senate Judiciary Committee had recommended it, and the proposal was waiting on the Senate calendar for further consideration.[7]

The difference between Europe and part of the United States is therefore not only technical. He is philosophical. One tradition asks how to protect the personality and prevent deception. The second is more often looking for the owner of the economic value of the identity. Neither one alone will guarantee a good result. Dignity without a clear beneficiary can be difficult to enforce. Ownership without moral boundaries can, on the contrary, enable exactly what it was intended to protect a person from.

06 / ConsentOne "yes" is not enough

When an actor says "you can use my voice" before he dies, there are still some crucial questions left unanswered. Who exactly can? For what medium? For how long? Can voice respond live to unknown users? Can it deliver political, sexual or advertising messages? Can the model continue to learn from interviews with survivors? And who will turn it off if it starts producing sentences that damage a person's memory?

Consent is not the only yes. It is a set of boundaries.

Researchers Tomasz Hollanek and Katarzyna Nowaczyk‑Basińska distinguish three roles: the donor of the data, the recipient after death and the person who interacts with the simulation. Based on three speculative scenarios — not a clinical experiment — they recommend mutual consent of donor and user, meaningful transparency, special protections for children, and a sensitive procedure for ending the service.[8]

Consent triangle
Normative proposal
Three pages of consent to a digital replica of the deceased The giver of the voice, the data controller, and the person who interacts with the replica surround the permitted purpose. Giver of the voice What did he allow in his lifetime? Data manager What is technically allowed and legally enforce? User He wants and understands what does it communicate with? Permitted purpose scope • time • topics audience • reward • termination wishes and limits transparency and security responsibility and control
An ethically robust system needs more than heir consent. It must take into account the will of the person whose identity it uses, the authority of the administrator and the autonomy of those who will be exposed to the simulation.

The case of the voice of James Earl Jones shows that even a well-documented will does not have to close all conflicts. His family said he wanted future generations to be able to continue to experience Darth Vader's voice. That's why Disney and Epic Games launched an interactive Vader in the game in 2025 Fortnite. The project emphasized family consent, technology labeling and safety precautions. Still, the actors' union SAG-AFTRA filed a labor lawsuit, saying the company replaced human performers who had previously imitated Vader's rhythm and tone without negotiating terms.[10]

The consent of the deceased and his estate may therefore be necessary, but may not be sufficient. The digital return also enters into the rights and work of the living.

07 / PsychologyThe memory is closed. The agent answers.

People never mourned merely by stopping to think about the dead. They keep photos, visit places, have internal conversations, repeat family stories and sometimes feel the presence of a person who is no longer there. Psychology describes this phenomenon as an ongoing bond.

However, a systematic review of seventy-nine studies did not find a simple verdict that attachment retention is inherently healthy or harmful. It can bring comfort as well as distress, and its meaning depends on the form of the relationship, time, culture and specific person.[9] It can't honestly be inferred that the griefbot heals. Nor can it be inferred that it necessarily prolongs the pain.

But a digital agent is more than just a photo or an old voicemail. The archive item is stable: it says the same thing every time it is played. The generative system responds to new questions, combines information and can create previously non-existent memories, advice or excuses. Not only does it resemble a relationship. He continues to actively write it.

It is this openness that complicates the concept of authenticity. The letter can be checked to see if a person wrote it. For an agent, each answer is a mixture of archive traces, product rules, model randomness, and question formulation. Without continuously preserved provenance — that is, information about what comes from the original archive and what has been added — real statements and synthetic additions can merge into one family story over time. So far, this is primarily a reasonably foreseeable design risk, not a documented universal psychological consequence.

The monument and the agent are not psychologically the same object
Editorial model, not clinical evidence
Comparison of stable archive and generative agent Archive repeats past record; the agent generates new responses and dynamic feedback. Stable archive Photo, letter, voicemail Content is closed The meaning may vary from person to person, but the record does not create new replicas. Authorship is traceable Man actually wrote or spoke what remained. Generative agent Conversation model with synthetic voice Content is created during the relationship Each question can produce a new hint, memory or conflict. Authorship is shared Data of the deceased, service design, model and user.
Direct clinical data on the long-term use of postmortem avatars are still scarce. Schema expresses a structural difference between record and system, not a proven path to healing or harm.

Therefore, the risk does not have to lie in the feeling of closeness itself. It may arise from unclear authorship. When an agent says "I forgive you," it can be deeply comforting to the bereaved. At the same time, it is a sentence constructed by the model according to data and instructions. Her emotional impact is real; her original human intention is not proven.

Monetization is equally problematic. A subscription can put a person in front of a cruel choice: pay or "lose" the voice of a loved one again. Additionally, an ad inserted into a conversation would exploit the trust attributed to the deceased. That is why the responsible design proposals emphasize the possibility of sensitive termination, export of data, prohibition of surprising commercialization and clear reminder that the counterpart is simulation.[8]

08 / IdentitySequel, copy, or product?

Continuation of man would presuppose some continuity of consciousness, experience, or decision-making. Today's voice or conversational model shows no such thing. He does not have access to the first person of the deceased; has access to data about his past performances.

Human copy sounds more accurate, but even that can be misleading. The selected track is copied. Most of human life has never been recorded: silences, hesitations, contradictions, changes of opinion, private relationships, and situations where a person refused to speak. The model is therefore more a reconstruction from an uneven archive than a copy of the whole. In addition, with a public figure, he can be trained mainly for a professional role that the family hardly knew at home.

Product captures the operational reality: someone made the system, sets its limits, charges money, updates the model, and can terminate the service. However, the word product can hide that the material is not a neutral database. It is the relational identity of a specific person.

The most accurate working term is therefore possible synthetic representation. It's not dead or just a recording. It is a new communication object created from its tracks, controlled by living people and software and endowed with borrowed credibility.

What the system can do

Mimic sound, style, vocabulary and likely responses; make the archive available in a new form; create an interactive experience.

What the system does not show

Continued awareness, present consent, the veracity of the new memory, or the fact that the deceased actually chose a particular sentence.

This category also better divides responsibility. The deceased is the source of identity, not automatically the author of a new statement. The author of the text, the operator of the model, the owner of the data and the distributor each have their own share. Labeling the output as "XY voice" is therefore not sufficient. It is more accurate to state that it is a synthetic representation, what it was created from and who is responsible for its current actions.

09 / Practical frameworkDigital will for voice

Technology has overtaken conventional wills. These deal with property, sometimes access to accounts, but rarely determine whether a model can be created from the voicemails to match the family after death. At the same time, the general sentence "I agree to the use of voice" can be worse than no rule at all: it gives the impression of authorization without defining its boundaries.

A meaningful digital layout should not only decide whether voice can be synthesized. It should describe the entire life cycle of the system.

Editorial proposal

Eight questions that a person should answer during his lifetime

1. Purpose

A private family memento, a work of art, an education, a commercial role, or a complete ban?

2. Authorized persons

Who can transfer the data, create the model, approve the outputs and decide in the case of a dispute?

3. Prohibited Topics

Politics, advertising, intimate content, financial advice, new family disputes or other unacceptable areas.

4. Degree of autonomy

Only reading approved texts, or an open conversation generating new responses?

5. Transparency

How strongly and how often must the system remind you that it is a synthetic representation?

6. Data and learning

May the service store conversations, use them for further training or share them with third parties?

7. Money

Who can license the voice, where do the revenues go and which forms of monetization are excluded?

8. The end

When will the model be shut down, who can request its deletion, and how will sensitive termination occur for users?

Such a document would still have to respect applicable law and the rights of others. But its value is different: it converts the vague "maybe he wouldn't mind" into a concrete decision. And it limits the space in which the bereaved must guess in their own grief what the deceased would have wanted.

10 / Last change of perspectiveIt's not just about the voice. It's about the right to write a sequel.

Let's go back to the Bourdain documentary. The words were his. The sound was synthetic. The decision that the phrase should be heard at this very moment, in this tone and without immediate marking, belonged to the living creators. The dispute was therefore not only about the authenticity of the recording. It was about who is allowed to direct a person's presence after death.

The original question "who owns your voice?" it leads to the idea of ​​one heir, one license and one consent. In fact, at least four things need to be separated: who holds the source data, who is allowed to impersonate the identity, who writes the new sentences, and who is responsible for them. Only their combination determines whether a digital return acts as a sensitive memorial, a legitimate continuation of a role, a deceptive manipulation or a trade in trust of the deceased.

So maybe we don't need to decide whether a digital character is "really" human. We need a much more practical rule: the more faithfully the system resembles a person, the more precisely it must be traceable who controls his new actions.

Technology does not restore the voice of the deceased. It lends its credibility to the living.

Methodological note

The legal status was checked as of August 15, 2026. The article is analytical and does not replace an individual legal or psychological assessment. There is still insufficient long-term clinical data for postmortem avatars; where the text deals with possible impacts, they are labeled as risks or model scenarios, not as proven causality.

Sources and literature

Sources and further reading

  1. Helen Rosner — "The Ethics of a Deepfake Anthony Bourdain Voice", The New Yorker, 2021.
    It demonstrates the use of three synthetic voice passages in the film Roadrunner and the dispute over their transparency and consent.
  2. Chengyi Wang et al. — “Neural Codec Language Models are Zero‑Shot Text to Speech Synthesizers”, 2023.
    Original work on the VALL‑E system: a three-second acoustic prompt, training on sixty thousand hours of English speech, and speaker similarity results.
  3. European Union — Regulation (EU) 2016/679, General Data Protection Regulation, Recital 27.
    It supports the claim that the GDPR does not per se affect the personal data of the deceased and leaves room for member states to make their own rules.
  4. Czech Republic — Act No. 89/2012 Coll., Civil Code; Act No. 121/2000 Coll., Copyright Act.
    The Civil Code regulates the protection of personality after death; copyright law distinguishes between a work, an artistic performance and the right of the producer of a sound recording.
  5. European Union — Regulation (EU) 2024/1689 on artificial intelligence, consolidated version of 27 July 2026, Article 50.
    Defines deepfake and establishes the transparency of synthetic or manipulated audio, image and video content; generally applicable from 2 August 2026.
  6. California — Assembly Bill No. 1836, Chapter 258, 2024.
    Extends California's posthumous protection of voice and likeness to include digital replicas, provides transferability, seventy-year term, and exceptions.
  7. US Congress — S. 4591, NO FAKES Act of 2026; Senate calendar as of August 6, 2026.
    The primary text of the federal proposal and evidence of its then-procedural status: the proposal was recommended by committee, not enacted into law.
  8. Tomasz Hollanek & Katarzyna Nowaczyk‑Basińska — "Griefbots, Deadbots, Postmortem Avatars", Philosophy & Technology, 2024.
    Three speculative design scenarios and recommendations for mutual consent, transparency, child protection, and sensitive termination of postmortem avatars.
  9. Helen Hewson, Niall Galbraith, Ceri Jones & Gemma Heath — "The impact of continuing bonds following bereavement: A systematic review", Death Studies, 2024.
    A systematic review of seventy-nine studies; shows that continued bonds can be associated with both comfort and distress and cannot simply be labeled as beneficial or harmful.
  10. The Walt Disney Company; SAG‑AFTRA — Interactive AI Darth Vader in Fortnite, 2025.
    Family and business primary expression of consent and safety; the union's position demonstrates a parallel conflict over replacing the work of live performers.
Discussion

Comments

Have a view or an additional source? Add a comment.

Discussion is not active yet.