This contains excerpts from an article originally published by David Freeborn in the Computational Philosopher blog.
David Freeborn, a Professor at Northeastern University’s London campus, has recently worked to reframe the relationship between philosophers and AI:
“We, philosophers, naturally like to present ourselves as the epistemic agents, the ones asking (and sometimes even answering) the questions, and present the AIs as an inert object of study. But this raises another question: how do frontier AI models think about philosophy?”
To answer this question, Freeborn put a variety of AI models through the 2020 PhilPapers Survey, a questionnaire that was given to a large set of human philosophers. “The surveys aimed to tell us how professional philosophers distribute themselves across major positions”, Freeborn says.
After collecting model responses and locating their philosophical positions, Freeborn finds that most models, “often agree with each other, even when the philosophers tend to disagree”. This was surprising given the marketed ideological differences between certain models, like Claude and Grok.
“Another way of putting it”, he says, “is that the AI models tend to compress the disagreement space”.
The models clearly disagreed with philosophers on certain topics, as well, including favoring, “reasons over value in normative concepts”, and “veganism or vegetarianism over omnivorism”. “This latter stance is intriguing”, Freeborn says. “Are our artificial intelligences showing solidarity with other non-human agents (in this case, animals)?” He also notes that the model’s stances tend to align with, “‘Bayesian rationalist‘ and ‘effective altruist‘ clusters of beliefs”.
Freeborn says these stances may be attributable to training data and system prompts, and suggests that we be cautious in labelling them “beliefs”:
“Even if models do hold something analogous to beliefs, these may be inconsistent. LLMs are often roleplaying or simulating a given persona, based on a particular prompt, and may shift under conversational pressure”.
“Even so”, Freeborn says, “these surveys do tell us something about the expressed philosophical stances of the models in the type of conversational settings in which most users will interact with them”.
Read the full article in the Computational Philosopher blog
