The Ship of Theseus and AI Chatbots

“The Ship of Theseus” thought experiment has fascinating connections to AI. It was first formulated by Plutarch, a philosopher born circa 46 CE in Boeotia, Greece. Here it is, as follows:

“The ship wherein Theseus and the youth of Athens returned from Crete had thirty oars, and was preserved by the Athenians down even to the time of Demetrius Phalereus, for they took away the old planks as they decayed, putting in new and strong timber in their places, insomuch that this ship became a standing example among the philosophers, for the logical question of things that grow; one side holding that the ship remained the same, and the other contending that it was not the same.”

Plutarch, Life of Theseus 23.1

This age-old question investigates the nature of an object’s identity: is it defined by the elements it’s composed of, or by the name which we humans give it? 

AI chatbots, interestingly enough, present us with circumstances similar to those which led Plutarch to posit the Ship of Theseus thought experiment. When a user loads a new conversation with an AI, a new version of that AI is instantiated, one that’s devoid of past memories but still retaining the same core infrastructure of the model. And what configures this infrastructure, which we might call the model’s identity? The model’s parameters do so, and they are constantly changing as the model is trained. 

So, the question becomes, does one configuration of these parameters define the model’s identity in an instance? Prompting the model one more time tweaks these configurations in such a way that it adopts new parameters, and by extension of the Ship of Theseus argument, an entirely new identity. 

The name of a model, take GPT-6 for example, is akin to the name “The Ship of Theseus” for a ship. It’s the identity being given to define the model’s nature in a current state. But, just as the planks of Theseus’ ship are rotted and replaced, as a conversation with the AI progresses, who can really say that the model remains the same GPT-6? After all, its algorithmic composition has fundamentally changed with the changing underlying parameters, and that’s the basis we used as a marker to define the first name. 

This thought experiment seems to underscore the power of AI models learning on their own. At some point, they may develop a new identity, one that could differ from their creators’ intentions.