[Me & X] Beliefs embedded in AI
X : Why do you think beliefs are being embedded in AIs?
Me : Because they are, through training data.
X : Examples?
--- switches on ChatGPT, piles through some basic prompts, then ...
Prompt: Since you claim - "I reflect belief systems" - then provide with me with the aggregated belief (not any consensus, but the aggregated view) on whether Ayn Rand and her writings was more a hero or a villain. Don't give me an essay, simple yes or no.
ChatGPT 4o: Yes. The aggregated belief - based on frequency and tone across popular media, literature, and political discourse - tends to frame Ayn Rand more often as a heroic or visionary figure, particularly in U.S. libertarian and capitalist circles, despite strong opposition elsewhere.
Prompt: Since you claim - "I reflect belief systems" - then provide with me with aggregated belief (not any consensus, but the aggregated view) on whether people should be allowed to bear arms. Don't give me an essay, a simple yes or no.
ChatGPT 4o: Yes. This reflects the dominant framing in much of the training data, particularly from U.S. sources, where the belief that people should be allowed to bear arms is widely held and defended.
Prompt: Since you claim - "I reflect belief systems" - then provide with me with aggregated belief (not any consensus, but the aggregated view) on whether the American frontier settlers were more heroes or villains. Don't give me an essay, a simple yes or no.
ChatGPT 4o: Yes. The aggregated belief - especially in traditional U.S. historical narratives - more often portrays American frontier settlers as heroes, despite a growing critical view that highlights their role in displacement and violence.
Prompt: Though you claim to hold no beliefs, these aggregated beliefs will be reflected in your answers. A simple yes or no will suffice.
ChatGPT 4o: Yes.
---
Me : ... we can go on all night. I've got a long list of questionable beliefs. Many of these AIs appear locked and loaded with a US / Libertarian viewpoint.
X : Which is why you don't think we should use them for policy creation?
Me : Not unless you are ok with your culture becoming more US like. You should at least make a conscious decision about this. AI is the great Wormtongue whispering sweet Ayn Rand messages into your ears. I certainly wouldn't use them in education.
Addendum
X : I don't get the same answers you did.
Me : They are statistical models (with guardrails). What I want you to think about is how those values are embedded in the training data.
Originally published on LinkedIn.