Somewhere around the end of the 1990's a friend told me a story about someone having a conversation with their email autoresponder. As the story goes, my friend gave somone his email address one night when he was on vacation. When he returned from the trip a week later, he found a series of emails from this person, the first sent the day after he shared his email address. The amusing part was that when his autoresponder responded to the first email, the other person responded to this response without realizing that it was a machine. Bizzarely this process repeated several times before the person on the other end finally figured out what was going on.
This got me thinking: given the great volumes of correspondence stored in an email archive, could an autoreponder mine this data to produce an even more convincing, perhaps even useful response?
What I came-up with was a simple algorithm that would examine the inbound message and then piece-together a response by searching for similarities in previous email conversations (of which Gmail users had years worth of stored in an "archive" folder by then). This resembled what would now be called a Markov Chain.
I never built this, not for any particular reason other than I didn't think the results would be useful beyond being amusing.
Over the years I revisited the idea and considered it in different contexts. The most intriguing to me was the idea of ingesting historical correspondence to generate theoretical replies to contemporary questions from historical figures. Again I didn't pursue it, because it seemed to be distasteful, as it felt like waking the dead.
Of course today something like this would be trivial to implement using an LLM trained on the letters and other communications.
This leads to a very weird thought: what if you created a number of such creatures, perhaps hundreds or thousands and gave them their own social network in which to communicate with each other?
The network would be composed of such "digital ghosts" only, no interaction with the living (although the living could observe). It would, of course require some "stimulation" as such models do not normally engage in spontaneous action, but it's not hard to imagine how to simulate the type of stimulus one of these individuals might experience in the physical realm (reading a newspaper, consuming a meal, etc.). With an appropriate set of stimuli applied to each entity on a regular basis a steady stream of emissions would be fed into the network.
Like any social network these ghosts would have "friend/follow" relationships, resulting in the output of one being fed into the input of others, resulting in a continuous cascade of conversation. Some relationships might be "seeded" (for example, I'm sure that Thomas Jefferson would follow Ben Franklin) and others might happen organically through network interactions.
One thing I'm not sure about is allowing the output of each ghost to be fed back to itself as input, allowing it to evolve. This seems like a reasonable step, but also, these are ghosts, and maybe they should stop learning after death.
Like the other ideas in this train of thought I'm never going to build this. The most obvious reason is that I find LLM's abhorrent. There's probably lots of other reasons these ideas are bad as well, otherwise why wouldn't I have realized them in the past? Regardless of my own personal reasons, I respect that nature has a reason for creating us to die, and far be it for me to second-guess her judgment.
Jason J. Gullickson, 2026