In high school we had maatschappijleer, where we learned about our Dutch constitution and democracy and the role of media in it. I clearly remember lessons on the value of pluralism in your media consumption. The importance of informing yourself through variable sources, through different lenses, was key in these lessons.
That lesson applies to AI too. I have discussions with my wife around bias in AI: gender bias, language bias, normative bias; many of these layers reside, more or less visible, in LLMs. And just like a newspaper, an LLM is not a neutral window on the world.
Part of that bias is by design. I use Claude heavily, but Claude is a product of Anthropic, a company with a vision and a goal to make money. How does this affect their AI models? Luckily Anthropic is open about this, as one can see on the Claude Constitution page. One can draw a worldview from this constitution: it’s embedded in Western philosophy, drawing from Aristotelian ethics, virtue ethics, rule following and how we think about what is good.
Part of that bias is inherited from the data. That constitution fits our world views pretty well, but it is also the lens through which the model is trained and behaves. The training data probably consists of a largely Western paternalistic corpus of texts; bias comes through English being the main language for all training data (even when it concerns non-Western ideas), and the training data is mostly from Western academia.
Here lies my tension. I want to use AI in a pluralistic way, just like I read multiple news sources. But relying on one model is like reading one newspaper. And so far I prefer Claude over other models so I know I focus too much on one source.
Hence I decided to add a line to my account’s base prompt:
“When answering normative/contested questions, briefly name the lens you’re using and what it doesn’t capture.”
This requires Claude to reflect on the lens it uses, or misses, when giving me normative or non-factual answers. With this line, Claude tells me:
- what lens is used,
- what is missing from the answer, and
- thereby it allows me to consider consulting another lens.
It’s not pluralism in the full sense since I’m still using one source. But it’s the next best thing my maatschappijleer teacher would recognize: at least know which paper you’re reading, and what it leaves out.