Every few months another AI recreation of a dead person arrives, and the argument restarts. Someone calls it a tribute. Someone else calls it digital resurrection. Both sides talk past each other, because the question is never whether the technology works. It is whether anyone had the right to use it.
I run an engineering company that builds these systems, so I have a direct interest in that argument going well. Here are the five questions we ask before we agree to build one. They are not legal advice. They are the questions that decide whether a project is worth doing.
Table of Contents
Toggle1. Who holds the rights, and did they ask for this?
The first question is not technical. If the person is living, their informed consent is the whole foundation, and it should cover what the persona may say and for how long. If they have died, the rights usually sit with an estate, a family, a foundation or an institution that holds the archive.
The warning sign is a client who wants to build a persona of someone they admire rather than someone they represent. Enthusiasm is not permission. If the organisation commissioning the work does not hold the rights to the material, there is no project. We have written elsewhere about how permission, boundaries and disclosure fit together in practice.
2. What may it draw on, and who approved that?
A persona built on everything ever recorded is a persona nobody has approved. We ask for a defined archive: these recordings, these writings, these talks, signed off by a named person who can authorise changes.
This is also the quality control. A model given an unbounded corpus will produce plausible sentences the person never said. A model restricted to approved material produces a narrower range of answers, which is exactly the point.
3. What happens at the edge of what it knows?
This is the question most clients have not thought about, and it is the most important one. Someone will ask the persona about an event after their death, a controversy, or something personal. The system has to have an answer for that, and the answer should be honest: this is outside what we have.
A persona that confidently improvises at the edge of its archive is not representing a person. It is impersonating them.
4. Are you telling people it is AI?
In the European Union this is now law. Article 50 of the AI Act has applied since August 2026 and covers avatars alongside chatbots: people must be told they are interacting with AI, clearly and at the first interaction. The European Commission has published guidance on how that applies. India has no equivalent statute, but the practical case is the same everywhere.
Disclosure is often treated as a risk to the magic. In our experience it is the opposite. People relax when they know what they are dealing with, and they engage more freely. The installations that make people uneasy are the ones that are coy about what they are.
5. What would make you stop?
Ask the client what they would not want the persona to say, and listen carefully to the answer. An organisation that has thought about this will have a list. One that has not is treating the persona as a marketing asset rather than a representation of a person, and the project will go wrong later.
Why this is worth getting right
There is a version of this technology that is genuinely valuable. An archive that answers questions is more accessible than an archive that sits in a library. A teacher whose recorded work spans decades can reach a fourteen-year-old who would never read the books. Institutions holding significant material have a real reason to want this.
And there is a version that is exploitative, built without permission, invented at the edges, presented as though the person were really there. The second version is easier to build and gets more attention. The first requires an organisation to accept constraints that make the result less dazzling and more honest.
Which one gets built is not decided by the model. It is decided in the first conversation, by five questions.
The objection worth taking seriously
The strongest criticism of this work is not that it is technically dishonest. It is that it interferes with grief, or with how a community remembers someone. Researchers studying AI recreations of the dead have raised exactly this, and it deserves a better answer than a disclaimer.
Our answer is partial and I would rather state it plainly than pretend it is complete. A persona built for a public institution, from an approved archive, presented as an interface to recorded teaching, is a different object from a private service that offers a grieving family a simulation of a parent. The first has an institutional owner, bounded material and a public purpose. The second involves one person in distress and a company with an incentive to deepen the attachment. We work on the first and decline the second, and I think the distinction holds, though I accept that not everyone will agree.
What a good brief looks like
When these projects go well, the client arrives with material they have already catalogued, a clear view of who the audience is, and at least one worry about what could go wrong. That worry is usually the most valuable thing in the room, because it tells you where the boundaries need to be.
When they go badly, the brief is a format rather than a purpose: somebody saw an installation elsewhere and wants one. Those projects produce a screen that gets photographed once and then ignored, and they are the ones most likely to cut corners on permission, because nobody involved cares enough to insist.
