The Ghost in the Machine

I was watching a programme the other evening about people creating AI versions of dead relatives. Husbands reconstructed from text messages, mothers rebuilt from old emails, voices pieced together from recordings. The idea is that you can keep talking to them. The programme treated this as something startlingly new. I am not entirely convinced.

I sometimes catch myself having imaginary conversations with my father. He has been gone a long time, but the mental map of him is still there. I know the tone he would use when he thought I was talking nonsense, and the small pause before he explained why. It is obviously not him. It is my memory of him, built from years of experience, but it can still be a useful reference point when I am thinking something through.

Which is when the penny dropped while watching the programme. Humans have always done this. The real question is simply where the conversation happens. Sometimes it happens inside the skull and sometimes it is pushed outside it.

Take prayer. Many people pray to saints, ancestors or deceased relatives. They speak to them, ask for guidance, sometimes imagine the reply. Imaginary conversations with someone you once knew work in much the same way. In both cases you are consulting a mental model built from memory. The voice is internal and you know it ultimately comes from your own reasoning. That internal version has a natural constraint because it is a mental map built from real experience. It cannot easily wander far beyond what you actually knew about the person.

Now consider the Victorian seance. People gathered around a table while a medium tapped out messages from the dead. The voice, of course, was the medium's. The dead relative was only being interpreted through someone else's words. Which brings us to the modern twist, AI relatives. Here the conversation is external again, like the seance, but with one crucial difference. The voice of the machine can be trained to sound exactly like the dead person, their phrases, their cadence, even their manner of argument.

So the structure ends up looking rather simple. Seances and AI relatives externalise the conversation. Prayer to the dead and imaginary conversations keep it internal. In all four cases the human brain is doing the same thing, modelling another person's mind. The difference is where the reply appears to come from.

The AI version, however, introduces something the Victorians never had to contend with. The medium never sounded like your father. The AI can. That similarity creates trust. When the reply arrives in the recognisable voice of someone you loved, the brain is far more inclined to treat it as authentic. Yet the machine is not recalling anything. It is extrapolating, generating sentences the real person never spoke.

The internal voice you carry from memory is limited by the map you built from real life. The external machine is not limited in that way. It can extend the pattern indefinitely. Which is where the danger lies. The technology does not resurrect the dead. It produces a persuasive imitation of them, a voice that feels familiar but which can now say things the real person never said.

The Victorians dimmed the lights and waited for the table to move. We sit on the sofa and wait for the typing bubble to appear. The instinct is the same, but the modern version sounds much more like the person we lost, and that makes it far easier to trust a voice that ultimately belongs to a machine.