How far can we get in developing language models like GPT-3
to let them represent the view of a philosopher?


Talk with DigiDan installation

I am working on the question of what digital replicas are and investigate their relation to their origins. Thereby I asked myself, how far can we get in developing language models like GPT-3 to let them represent the view of a philosopher?
This could lead to insights into how the human mind works. It might even tell us something about what is special about humans. Furthermore, it could turn out that we could use this replica as a new thinking tool. Moreover, I am concerned about misuse and potential ethical consequences.

WHAT is a GPT-3? A GPT-3 is a neural network trained to predict the next likely word in a sequence.

Technically, it is a 175 billion parameter language model that shows strong performances on many Natural Language Processing tasks. The abbreviation stands for Generative Pretrained Transformer!

  • It is called Generative because it can generate long sentences, not just yes or no answers or simple sentences.
  • Being pretrained means, it is already trained with nearly 500 billion tokens taken out of Common Crawl, WebText, Books, and Wikipedia.
  • The whole thing is a Transformer because it is based on a typical Transformer structure. This means that this network applies a self-attention mechanism that calculates the probability of the next word appearing surrounded by the other ones.

Fine-tuning the Davinci model of GPT-3 on the corpus of Daniel Dennett in collaboration with Eric Schwitzgebel, Matt Crosby

Our research questions:
- How to create the ‘best’ model and how to operate it.
- How convincing the generated completions of such a model are for the philosopher himself, experts, and naïve people.
- Can we use this model as a new thinking tool?
- And last but not least, we are concerned about risks of misuse and critical ethical consequences of replicas of this sort

Strasser, A., Crosby, M., Schwitzgebel, E. (2023). How far can we get in creating a digital replica of a philosopher? In R. Hakli, P. Mäkelä, J. Seibt (eds.), Social Robots in Social Institutions. Proceedings of Robophilosophy 2022. Series Frontiers of AI and Its Applications, vol. 366, 371-380. IOS Press, Amsterdam. doi:10.3233/FAIA220637

This picture illustrates the project: On the left side, you see the complete work of a philosopher. In the middle, you see the GPT-3 we fine-tuned. On the right side, you see the not yet existing manual on how to operate this machine.

Experiment to evaluate the model in collaboration with Eric & David Schwitzgebel

We asked the real Dennett 10 philosophical questions and then posed the same questions to the language model, collecting four responses for each question without cherry-picking. Experts on Dennett's work succeeded at distinguishing the Dennett-generated and machine-generated answers above chance but substantially short of our expectations. Philosophy blog readers performed similarly to the experts, while ordinary research participants were near chance distinguishing GPT-3's responses from those of an “actual human philosopher”.

Schwitzgebel, Eric, Schwitzgebel, David, Strasser, Anna (2023). Creating a Large Language Model of a Philosopher. Mind & Language, 1–22. https://doi.org/10.1111/mila.12466