AI & REWARDMAXXING: Part 1 AI isn't neutral—it's deeply Eurocentric. Most large language models treat Africa and the Global South as an afterthought. Here is why passive internet data is corrupting AI alignment, and how we fix it. ︎

Eurocentric; more like 'western' centric .  Agree with the rest. However there is the 'start-up' problem (very similar to the plant 'shut-down' problem in continuous flow systems), in that it's not clear if Africa / Global South would actually want to 'copy' the exploitative euro capitalism (who or what to exploit, and how to create that 'difference'?). 

We also have the hindsight narrative bias that fails to acknowledge the many alternative destinies and ambitions of the past (e.g. Hippies & flower power) All we are getting out of AI is the summarisation of the endless regurgitation in print of the status quos of the time. As humans we tended to accept the adequate and move on, unless there was a pressing need to further refine the (in)adequate (with p=0.5, ...) 

Th algorithms that are pushing the social media doom scrolling (of AI etc), the new 'masses', will inevitably create a niche sector of 'elites' that have avoided being captured by the doom scroll. Exactly what that new elite will do with their 'economy' is anyone's guess.

Philip Oakley said:

Eurocentric; more like 'western' centric

Possibly even US-centric rather than European. And likely with a huge bias toward English-language data sources too (whichever side of the pond the originated from).

   - Andy.

I really enjoy your point of view, even though I disagree with some of your points. For example, I believe the hippie era came much later, largely because the West was experiencing one of its most prosperous periods; some might describe it as a period of privilege.

Interestingly, some argue that the new economy AI could create by 2035 might produce a level of prosperity and abundance comparable to the conditions that helped give rise to the hippie era.

In Part 2, I explore the different biases AI has in depth. Stay tuned until the end of the 30-part series—you might find some of the concepts I explore quite interesting.

Absolutely, and I think the English-language bias is an important point. However, I think the deeper issue goes beyond language. AI can now translate, transcribe, and process written historical material regardless of the original language, so written history is still much more accessible to the training corpus.

The real gap, in my view, is with cultures whose histories were preserved predominantly through oral traditions rather than written records. That potentially puts many African cultures, Aboriginal peoples, and Pacific Islander communities at a particular disadvantage. So even if AI becomes less Eurocentric and more multilingual, the underlying data imbalance may still persist because the history simply isn’t available in the same machine-readable form.

So the problem may eventually become less about which language AI speaks and more about which histories were ever written down for AI to learn from in the first place.

Sultan-Othman Adekoya said:

the hippie era came much later

My 'point' there was that we only have the one history. It's like asking what if Donal Trump hadn't won the elections; it's a hypothetical question. There was a hippie flourishing but then it didn't achieve predominance, so we don't have the data inputs for the AI of what a proper 'hippie' economy would actually look like, what they'd get right and what they'd get wrong.There were other alternates. The hippies were just one of many. 

Every generation appears to find scapegoats among the mainstream actions of the past, that weren’t seen at the time. (e.g. all the mother & baby units, mum's out of wedlock, forced adoptions etc.) We don't yet know what our future faults will be, and AI doesn't know either.

Sultan-Othman Adekoya said:

AI can now translate

There is the 'breakfast' problem with translations though. An English breakfast and a Continental breakfast are completely different [content] morning meals. Same happens with school/ecole as to what is mandatory/prohibited teaching (e.g. religious education).

We also have the "Quran can only be read in Arabic" aspects (and / or the bible was written by white English speakers..) that can lead to misunderstandings.

The point that current AI only ingests written capta (becomes data in context?) does highlight a great omission, even if the 'rise of the west' is often attributed to our [the elite's]  ability to read & write - Catch 22)

Sultan-Othman Adekoya said:

cultures whose histories were preserved predominantly through oral traditions rather than written records. That potentially puts many African cultures, Aboriginal peoples, and Pacific Islander communities at a particular disadvantage.

Many older European cultures were the same - traditional Celtic (the underpinning of Welsh and Scots & Irish Gaelic) for instance almost never wrote anything down, but history was passed down orally with great precision from one generation to the next for thousands of years. It wasn't until the arrival of the Romans did that start to change (and then not all at once). Of course much has been written down since, but still a lot of the original was lost, or only persisted as 'legend'.

  - Andy.

You’re absolutely right. Things can easily get lost in translation—we have the whole of history to prove that. But as AI moves towards superintelligence, do you think it’ll eventually become better at understanding context and preserving the meaning behind different perspectives?

I touch on the topic of fixing AI models in Part 6 of this AI & Rewardmaxxing series. Stay tuned.

The question I was really getting at is: what happens if different cultural viewpoints on the past, morality, lifestyle, and our relationship with the natural world are fed into AI? Would it produce different outcomes depending on which school of thought it was given?

Different cultures can have fundamentally different ideas about morality and how we interact with nature. For example, some people might view taking a crab, using it as bait to catch a fish, dragging that fish for miles until it is exhausted, bringing it out of the water and essentially suffocating it just to take a picture before putting it back as extreme cruelty towards animals. Others might simply view it as fishing, while arguing that you should only hunt or kill what you actually intend to eat.

History is often what the people who wrote it make it out to be. But, coincidentally, there are many other histories intertwined within it, predating some Western historical narratives, that may unfortunately never make it into an AI’s dataset—and the AI may therefore never know about them.

As for the hippie era, it could be said that it was arguably a by-product of the prosperity and social conditions of its time, just as many other norms that emerged from that era may have been more of a cultural façade than a permanent shift in human behaviour.

AI, as a prediction model, may very well be able to predict certain patterns of the future if it is given the right information because it doesn’t merely learn facts; it learns statistical patterns about what humans consider normal, acceptable, moral, authoritative, historical, etc. If some cultural frameworks are massively overrepresented in the training data while others are primarily preserved orally, can the model really give those frameworks equal epistemic weight?

I’ve always believed that nothing is truly new in the world. Everything that could happen may very well have happened before, and perhaps history is simply going to continue repeating itself in different forms.

History is often what the people who wrote it make it out to be. But, coincidentally, there are many other histories intertwined within it, predating some Western historical narratives, that may unfortunately never make it into an AI’s dataset—and the AI may therefore never know about them.

As for the hippie era, it could be said that it was arguably a by-product of the prosperity and social conditions of its time, just as many other norms that emerged from that era may have been more of a cultural façade than a permanent shift in human behaviour.

AI, as a prediction model, may very well be able to predict certain patterns of the future if it is given the right information because it doesn’t merely learn facts; it learns statistical patterns about what humans consider normal, acceptable, moral, authoritative, historical, etc. If some cultural frameworks are massively overrepresented in the training data while others are primarily preserved orally, can the model really give those frameworks equal epistemic weight?

I’ve always believed that nothing is truly new in the world. Everything that could happen may very well have happened before, and perhaps history is simply going to continue repeating itself in different forms.