Cultural Heterogeneity and AI - what do we risk losing?

Will culture heterogeneity be lost as AI models optimise based on the greater volume of learning material from specific cultures?

This is the question I asked Reid Hoffman, co-founder of LinkedIn, Manas AI, Inflection AI and part of the original team at Paypal. He has also been the building force behind companies that have shaped the internet and AI revolutions of the twenty-first century.

To my question, this was his response:

“Cultural heterogeneity. This is actually one of the issues around how do we try to not create monocultures. Now, some of it is like the dominance of training materials in English. For example, one of the early tests I ran that was interesting when I had GPT-4 is I had it create some poems in Hindi, and then I had it create poems in English and then translate them to Hindi. I had some friends of mine who are poets who are fluent in Hindi, and I did a blind test with them and they all thought the poems that were initially in English and then translated to Hindi were better. It is part of the subtlety of having these massive training sets in English that creates one set of heterogeneity.“

I had the chance to see Hoffman‘s return to the Intelligence Squared stage in London last month, as he discussed how AI is rapidly changing our world and how we can use it inclusively to improve our lives and create positive change.

I wanted to ask Reid a question about how something seemingly harmless could actually lead to potential problems in the future. Previously in the discussion, Reid had reminded the audience that the most used programming language is English, and several coding languages are based of the English language‘s syntax and linguistic structure.

Therefore, I was wondering what could be the result of this use of English not only to code AI programs, but also with English-language texts predominantly being used by AI’s to learn. Not only in terms of language, but also considering that the majority of these resources may also be ingrained with a Western-centric viewpoint. What would this mean for AI’s ability to understand and provide insight with the vast variety of cultures present in our modern society?

In terms of Hoffman’s poems example, models such as Chat-GPT will have ten, twenty, a hundred times the number of English-language poems as part of their data sets, than they will Hindi-language poems. As a result of having this greater set of data, they will have a greater understanding of the features present and the type of poem requested, and therefore will do a better job coming up with its own.

In answering my question, Hoffman continued with the following:

“Then there’s the question of companies competing with each other which will create some differentiation but will also create heterogeneity. I think this question is ongoing, and almost bends to the epistemology question which is that we want to make sure that we’re continuing to think about the odd things, the blind spots, the differences along with learning collective truths. And I think that is a subtle and ongoing work that is really important.”

There is a risk that currently AI is primarily designed to act in the most pragmatic, efficient, objective and optimised way. However, we do not want to lose the important details or intricacies that are an integral part to individuals and to each unique culture. There is a beauty in the diversity and variety present in each community, city and nation on our planet. It is essential that even with our increased use of technology, we do not fall into this ‘monoculture trap’.

And this problem of AIs such as Chat-GPT being monocultural is not a new one. It is a concern that has existed ever since generative AI and LLMs started taking the centre stage. Jill Walker Rettberg from the University of Bergen, in article written in December 2022 which was around the time public Chat-GPT usage really took off, discusses how OpenAI trained their models off information available on the internet.

Rettberg’s findings demonstrate that Chat-GPT‘s ‘Wikipedia learnings’ (sites which may be dubious information sources themselves however that is a different conversation!) include only English-language Wikipedia pages, not all of them. Furthermore, the AI’s use of ‘Common Crawl’ researching - “an open repository of web crawl data in more than 40 languages” - retrieves 51.3% of its information from pages originating in the United States.

While this data was collated in 2021/2022, there is the assumption that with increased usage, AI models have been updated and due to OpenAI stating that their models are no longer limited to data before September 2021. However, the numbers are concerning nonetheless. These are technologies that are being used by individuals all around the world, in a great number of different languages and who are a part of myriad cultures. Risking losing our cultural heterogenity is a very real possibility for future generations if our AI usage increases as it is projected to.

Hoffman’s question of how we can use AI and technology for social good and beneficial change for the largest number of people, is a central one present in the topics I am focusing on. Being a part of the current technological landscape means we can focus on this technological revolution with a beneficial focus.

A technical revolution spearheaded by focuses on humane AI and harnessing technology in a way that has the capacity to create better outcomes for society as a whole.

References:

Watch Reid Hoffman on ‘What Could Go Right with Our AI Future’ interview here. (Reid’s response to my question features at around 1:00:33)

ChatGPT is multilingual but monocultural, and it‘s learning your values