Our digital footprints now outlive us by default. Messages, photographs, voice notes, emails and social media posts can remain stored long after the person who created them has died.
Generative AI adds a new possibility. These digital traces can be used to produce an interactive approximation of someone’s voice, appearance, language or personality. The result may be called a griefbot, deadbot, thanabot, post-mortem avatar or generative ghost.
The technology is already commercially available. The difficult questions are no longer purely technical. They concern consent, dignity, memory, ownership and responsibility.
What digital ghosts actually are
A digital ghost is an AI system designed to simulate the communicative presence of a deceased person. It may be built using:
- Text messages and emails
- Social media posts
- Voice notes and recorded conversations
- Photographs and video
- Personal writing and online activity
- Interviews or information supplied by relatives
Some systems are text-based. Others use voice cloning or audiovisual avatars to create something closer to a conversation. Unlike a static memorial page or a folder of photographs, a generative ghost can produce new responses that the person never actually gave.
That difference matters.
A recording preserves something that happened. A generative system predicts something that might plausibly be said. It can sound familiar while still inventing facts, opinions and emotional reactions.
Research published in Ethics and Information Technology describes digital afterlife systems across nine dimensions, including consent, data sources, interaction method, fidelity, disclosure, purpose, audience, governance and behavioural agency. The researchers propose that consent, disclosure and purpose should act as threshold requirements for ethical design.
In plain English, a system should not quietly impersonate someone, use their information for an unclear purpose or encourage people to mistake generated speech for the person themselves.
The technology feels real for reasons that are not entirely about accuracy
The most revealing recent study came from Jack Manning and colleagues at the University of Colorado Boulder. Their controlled study, presented at the 2026 Designing Interactive Systems Conference, involved 16 volunteers speaking with AI ghosts built from a short survey about someone they had lost.
The volunteers generally found the conversations comforting. One participant said, “It felt like I was actually talking with my grandpa.” Another said, “It just feels like I’m getting the closure I needed so bad.”
The systems still made factual mistakes. One invented a job the deceased person had never held. Participants often overlooked those errors. What troubled them more were small departures in tone and manner.
One bot did not use emojis, despite the participant having regularly communicated with their loved one through text. Another called someone “champ”, a word that person would never have used.
This suggests that emotional and social expectations can matter more than factual accuracy. A familiar phrase, rhythm or habit may create a stronger feeling of presence than a large database of correct information.
There was an important limitation. A human researcher reviewed every response before it reached the participant. Commercial systems operating in the real world may not have that safeguard.
One participant worried about becoming too reliant on the bot. Another said, “I don’t know if I would like the person I would become if I kept using this.”
Manning, whose 13-year-old sister died when he was 11, has acknowledged both the risks and the possible value. His conclusion is measured rather than absolute. “There is a version of generative ghosts that could be a really positive thing in the world.”
Acceptance depends on the person and the situation
A 2026 qualitative study in Frontiers in Digital Health found that public attitudes towards griefbots were conditional rather than simply positive or negative.
Acceptance depended on:
- The user’s emotional state
- The purpose of the interaction
- The stage of grief
- The nature of the relationship
- The person’s readiness to engage
- The level of disclosure and control provided
Some people may value a structured way to express something left unsaid. Others may find the experience intrusive or distressing. A system that feels supportive several years after a death may feel deeply unsettling shortly afterwards.
Reported usage also appears to be growing. A September 2026 survey reported by Fast Company found that nearly one in four people who use AI had used it during the grieving process. Among Gen Z users, the figure approached 40%. Around 97% of those users said they found value in the experience, while 28% expressed concern about emotional dependency.
These figures should be treated as evidence of emerging behaviour, not proof that griefbots are beneficial for everyone. Value and risk can come from the same feature, namely the system’s ability to simulate continued presence.

The central problem is that you cannot ask the dead
Consent is the most obvious difficulty.
A living person can agree to a voice clone, limit its use, withdraw permission or correct an inaccurate representation. A deceased person cannot do any of those things.
Relatives may provide permission, but their wishes may not match those of the person who died. A platform may hold the technical data, but that does not automatically give it moral authority to recreate someone. A will may deal with financial assets, yet say nothing about messages, recordings or digital likenesses.
Researchers have therefore proposed the idea of mutual consent. This means considering both:
- Whether the deceased person authorised the creation and use of a representation, where possible
- Whether the living person interacting with it has given informed consent and understands its limitations
A further proposal is to move beyond one-off consent. The 2026 preprint Beloved Afterlives argues for relational authority, where authorisation remains traceable as a representation moves between providers, relatives and audiences.
That is a useful distinction. Consent should not disappear once a file is uploaded. It should remain connected to:
- The source of the data
- The intended purpose
- The authorised audience
- The organisation controlling the system
- The ability to challenge, suspend or delete it
When a simulation begins to overwrite memory
There is also a difference between remembering someone and allowing an AI system to speak in their place.
A photograph may prompt a memory. A voice note may preserve a real moment. A generative ghost can create a new interaction that never happened.
Over time, those invented conversations could become difficult to distinguish from genuine memories. A person may remember what the system said, even though the deceased never said it. The AI version might gradually become more available, more responsive and more emotionally convenient than the complex human being it represents.
This creates a risk of memory distortion. It may also alter the reputation of the deceased by making them appear to hold views, use language or express emotions that were never theirs.
The ethical literature also raises concerns about:
- Emotional dependency
- Prolonged or maladaptive grieving
- Unwanted contact or automated messages
- Commercial exploitation of grief
- Public or political use without meaningful consent
- The use of private messages for purposes far removed from their original context
This does not mean that every digital memorial is harmful. It means that the distinction between a bounded aid to remembrance and an unbounded substitute for a person is important.
The UK legal position remains incomplete
The UK’s Property (Digital Assets etc) Act 2025 came into force on 2 December 2025. It recognises some digital assets as personal property in England and Wales, including assets that are neither physical objects nor traditional legal claims.
That is significant, but it mainly addresses succession. It does not solve every practical problem involving digital remains.
There are still questions about:
- Platform terms of service
- Password recovery
- Access to cloud accounts
- The transfer of licensed services
- The treatment of messages involving other people
- Post-mortem privacy
- The use of data to train or operate AI systems
As the Law Society explains, digital assets may be property while many online accounts remain governed by contractual licences that cannot simply be transferred. The BCS also notes that data protection rights largely end at death, leaving many decisions to platform rules.
This is legal context, not legal advice.
The UK GDPR protects living identifiable people. Its principles still matter when an organisation holds records connected to a deceased person, because those records may also identify living relatives, colleagues, customers or correspondents. However, the deceased person is not generally treated as a data subject with enforceable UK GDPR rights after death.
That creates a regulatory grey zone. The 2026 IEEE paper on post-mortem data protection describes similar gaps across privacy and AI regulation.
What organisations should be thinking about now
The business relevance is straightforward. Organisations may hold the raw material from which a future digital likeness could be built.
That material may include:
- Customer records
- Call recordings
- CCTV footage
- Email archives
- Support tickets
- Chat logs
- Voice messages
- Website enquiries
- Photographs and identification documents
The immediate question is not whether an organisation intends to create griefbots. It is whether the organisation is retaining more personal information than it needs, for longer than it needs, without a clear lifecycle.
Under UK GDPR, two principles are especially relevant.
Purpose limitation
Personal data should be collected for specified, explicit and legitimate purposes. Information collected to provide customer support is not automatically available for unrelated AI training or personality reconstruction.
Storage limitation
Personal data should not be kept longer than necessary for the purpose for which it was collected. “We might need it one day” is not a complete retention policy.
A practical review should ask:
- What information do we hold?
- Why was it collected?
- Is the original purpose still active?
- How long should it be retained?
- Who can access it?
- Is it copied into backups?
- Is it shared with suppliers or AI tools?
- What happens when a customer or employee dies?
- Can we suspend or delete it in a controlled way?
For organisations reviewing these questions, our managed IT services context is relevant because retention is not only a legal or policy issue. It is also a systems issue. Deletion must be considered across live databases, email platforms, archived files, endpoint devices and backup cycles.

A practical priority list for organisations
High priority
- Create a data inventory covering records, recordings, messages, images and backups.
- Assign a retention period to each category based on its purpose and regulatory requirements.
- Restrict access to sensitive personal information using role-based permissions.
- Review third-party AI tools before staff upload customer messages, recordings or photographs.
- Document deletion procedures for live systems and backup environments.
Medium priority
- Add post-mortem considerations to data governance policies, even where the law is not yet comprehensive.
- Record the source and purpose of sensitive datasets so future use can be assessed.
- Review supplier contracts for training, reuse, model retention and deletion provisions.
- Train staff to recognise that voice recordings and conversational data may be highly sensitive.
Low priority
- Consider an organisational position on synthetic representations, including whether a deceased customer, employee or public figure could ever be recreated.
- Monitor regulatory developments such as proposed post-mortem rights to be forgotten and data-donor directives.
- Include digital legacy questions in wider continuity planning, particularly where senior people’s accounts or business knowledge are involved.
The priority is not to predict every future AI product. It is to make today’s data holdings proportionate, understandable and governable.
The takeaway
Digital afterlives are no longer purely fictional. AI can already reconstruct aspects of a person from the traces they leave behind, and some people find that comforting.
The evidence does not support a simple verdict. Griefbots may help some people express feelings or explore unresolved loss. They may also distort memory, encourage dependency or allow commercial systems to speak with authority they do not possess.
The responsible question is not simply whether AI can recreate someone. It is whether the person consented, whether the living user understands the simulation, whether the purpose is legitimate and whether the representation can be challenged or retired.
For organisations, responsible stewardship begins with more ordinary questions. What do we retain? Why do we retain it? Who can access it? When should it be deleted? Those questions protect people in life and may also reduce the possibility of unwanted digital afterlives after death.
A final question for all of us
Who would you trust to decide what happens to your messages, photographs and voice notes after you are gone, and would you want an AI version of you to speak at all?
