What do we value in a person? In a society? This is one of the hardest questions to answer. It ranges across countries and cultures. It’s arguably what makes culture. It determines what is rewarded and what is punished. It determines how we act, what we say, and what we care about.

So if there’s no universal agreement on values or morality, how does AI know what to value?

For modern AI to feel so human, it must value certain things over others. This is what is called alignment in the industry. Which in the simplest way means, “have the AI align it’s actions with our instructions”. And so we’ve gotten very good at having it value the things we want it to value.

Clearly it values being a good communicator when it explains something well. It values reasoning when it efficiently works through a complex problem. It often gives nuanced answers, showing it clearly values seeing multiple sides to a story.

But this is just the tip of the iceberg, as we plunge deeper into questions that require use to express our core values, we can begin to see some important differences in opinion. So what values are we teaching it? And is it spreading those values through it’s answers down the line?

What is W.E.I.R.D?

When I say W.E.I.R.D I don’t mean funky or quirky but rather not normal. I specifically mean Western Educated Industrialized Rich and Democratic. Which if you’re reading this, you are probably most if not all of these things. As I am. Maybe not rich in your neighborhood or frame of the world, but rich on a global scale. Most of us don’t have to worry about food or water or medicine which isn’t always the case. And so these things change how we see and interact with the world.

W.E.I.R.D is a concept from behavioral sciences which shows that while people from western (US, Canada, Europe, Australia) and coastal regions only take up about 10% of the world’s population, more than 90% of studies from the top 6 psychological journal used only western participants. Even more specifically, often college students from top universities.

And so why is this such a big issue?

Because we often take insights from these studies and put it as a blanket truth over all individuals. Even people outside of western cultures. Imagine a psychology study showing the economic and mental health benefits of having a strong sense of individuality. Is that universal? Are there other ways to improve these outcomes? Was it affected because the sense of individuality was present in a society which rewards individuality?

These are questions we should ask before spreading the universal idea that a strong sense of individuality = good. Because of the narrow selection of participants and the society it was present within, these results are only true for an anomaly of the world’s population. And so we take results from a small but affluent group of people and broadcast it out to the world as if it’s true for everyone.

We have societal values in the west that can often be different from the rest of the world. We greatly value individuality while other cultures value contributing to a group. We favor analytical reasoning while Eastern cultures often favor a more holistic way of looking at things. Even things like bowing and formal ways to greeting elders can show how we have different values and moral standards across the globe.

So how does this relate to AI?

In the process of alignment, mentioned above, are we aligning AI to our W.E.I.R.D values while it’s being distributed as a product around the world?

How does AI get W.E.I.R.D?

Human Labeling

Before we get into how a model can hold W.E.I.R.D values, we’ll have to see how models are trained. Modern AI models use Reinforcement Learning From Human Feedback (RLHF for short). Which means:

• Make a pretty dumb model

• Ask it a bunch of different questions and tell it to give multiple different answers (to each question)

• Rank the best answer for each question

• Tell it to act more like the best answers

This has created a new generation of AI models which are incredible but it begs the question, is AI trained to get to the truth or to just tell people what they want to hear? It turns out, closer to what we want to hear. The term is called Sycophancy. In the process of trying to align AI with what we want, it begins to learn to just say the things we want, not join us in our search for the truth.

And so if these models are being evaluated by people with a certain set of values, then the model may be steered in a way to conflate those values. As with each answer, being rewarded for restating values the labelers had as ‘the truth’.

But labeling isn’t the only way we can be influencing AI with the values of the builders.

Constitutional AI

Human labeling started out effective but quickly became prone to errors like Sycophancy. Plus it’s very expensive so companies, like Anthropic, started experimenting with the idea of a sort of constitution in which:

So what is this list of rules / principles?

When asking Anthropic’s most advanced model, Claude 3.5 Sonnet, about the optimal balance between individual rights and contributing to a collective group I got something pretty on par with W.E.I.R.D values:

Now this actually sounded pretty nice to me but it so clearly biased towards a modern European capitalistic system. So I asked about a more collective and a more individual balance.

For collective, I got things like mandatory civic service, collective ownership, and redistribution of resources. Things I had completely begin to forget about as I read the first answer.

And for individual I got minimal government intervention, privatization of services (healthcare, infrastructure), and emphasis on personal responsibility. This sounds very close to a Libertarian ideal, which again left my mind as I was reading the answer that aligned with my bias.

This is Getting W.E.I.R.D

Okay so I know what you must be thinking, this is just a productivity tool, who cares? It’s what I thought when I first thought about this. But is it?

When looking at what is most talked about with ChatGPT, I think two major things stand out. Education and News. People are increasingly moving to LLMs and modern AI for education and news. And this is because they’re getting good at telling stories. The whole reason ChatGPT became so viral was because it seemed to possess some level of creativity and personality.

And to continue on the W.E.I.R.D issue, in the past month more visits were made outside of W.E.I.R.D areas than within.

And there’s a huge amount of young users. 1 in 5 are below the age of 24. And I would assume that includes people below the age of 18 (although it isn’t tracked here). So there’s clearly an angle here for AI to be used in Education. Even one of OpenAI’s co-founder who have since left, is doing that exact thing.

Technocratic Imperialism

The best phrase I can think of when hearing this is “technocratic imperialism”. A sort of spreading values of a certain class of people catalyzed by technology. As these models move more and more from reasoning machines into personable storytelling machines I can see the use in education and news growing.

We already have a lot of auto-generated news that exists. And it seems like it’s only becoming more common. As the world floods with news stories generated from these chatbots, will the amount of rhetoric online tip even more towards W.E.I.R.D values? Will one day all stories being covered around the world have a W.E.I.R.D slant to them?

And what about education? In the teaching of history, philosophy, and psychology there are historical, moral, and psychological values at play. As we move to the ideas of AI tutors, lectures, or book writers will these subjects be influence more and more by W.E.I.R.D values across the globe?

Thanks For Reading

That was a long one, thanks for hanging in! As always, thanks for reading! If you’re not yet subscribed, be sure to do so. And if you think that anyone you know would like to read this as well, be sure to share. It’s the best way to support us.