Global AI on a Local Scale

To apply AI tools and applications effectively, there are a number of considerations which need to be made. Key among them is the need to ensure your outputs are relevant to your local context. To ignore this is to put your brand and your credibility at risk, and alienate your audience.

Created with MidJourney by DBZ.

An AI in your pocket

At Apple’s Worldwide Developers Conference (WWDC) 2024 event, the company unveiled Apple Intelligence - their own take on AI. Due for release at the end of this year, Apple Intelligence will be integrated into various Apple products. It will, they state, “Help you write, express yourself and get things done effortlessly.” It will learn and adjust according to your behaviour, group notifications which are important, help you to focus, assist with image creation, give Siri a much needed upgrade and more. Basically anybody with the latest iPhone will soon have a fully functioning AI in their pocket.

In other words, if you own an iPhone, AI is coming whether you like it or not.

As the use of GenAI rises and artificial intelligence becomes increasingly integrated into our day to day activities, it is becoming more and more evident that there is a need for AI which considers and responds to local contexts.

The internet is American

American culture and influence dominates on the internet, both in terms of subject matter and in terms of language. This is evident in everything from American English underpinning programming languages to the fact that the largest and most influential internet companies like Meta, Google, Amazon, OpenAI are American, governed by American laws and regulations. It is also worth mentioning that most of the internet is physically hosted on servers in the United States of America.

Blayne Haggart explores this subject in his essay American Internet, American Platform, American Values, where he argues that viewing the world wide web as a multinational entity is not only incorrect, it is dangerous. He writes:

Accepting this perspective has two effects: it allows narrow American policy debates to substitute for a “global” debate while it simultaneously delegitimizes attempts by non-American countries to assert democratic control over their economy and society.

The ripple effects of this are many and varied—our respect goes to the lawyers operating in this space. But the reality is we sort of just deal with it. Of course, court cases crop up, different governments introduce their own versions of different laws and regulations, but we are all online 24/7, all generally able to access the exact same content and life, for most of us, goes on.

ChatGPT's values

A recent study out of Cornell University in the US made news when it showed that ChatGPT's cultural values are most closely aligned with English-speaking and Northern European countries. Researchers used the World Values Survey to prompt five versions of ChatGPT for responses. The outputs were then plotted on a graph, with results clustered around Great Britain, North America, Australia and New Zealand, as well as Finland, Switzerland and Sweden. The study showed it exhibited "a cultural bias towards self-expression values", and that this was consistent across the models.

The study also showed that the simple solution to countering this bias is to identify a desired culture within the prompt: "Although it is not universally effective, cultural prompting improves cultural alignment" 70% or more for each model.

However, not all users across all regions will have an awareness of inherent bias, or the prompts to dodge it. There is a major risk that cultural insensitivity or even alienation could occur by users whose culture does not align with that of language models.

Locally grown AIs

Language models like ChatGPT, Claude, Gemini and so on have all been trained on online content as well as on tailored datasets like The Pile. These huge datasets contain hundreds of gigabytes of assets from Project Gutenberg, eBooks, legal documents, academic papers and even sets of emails. It not only extends the knowledge base of language models, but it teaches them writing styles and gives them an understanding of what it means to play a character, answer from certain perspectives and have some degree of specialisation.

As Gen AI models have largely been trained on the vast galaxy of online content, so much of which is American, Gen AI too, by extension, creates content from an American context. This means that non-American users need to be mindful of where the information is coming from and how the outputs are framed.

While we assume that the various providers out there are creating products which are right for us, there is still an obvious gap in the market for AI which is relevant to specific languages, cultures and societies.

Katonic AI has had this sense and launched an initiative to build Kangaroo LLM—a language model built for the Australian user. It will be trained on Australian internet content and have an awareness of Australian language and culture. While the intent is there, their struggles to get the project off the ground are proof that there is a lack of awareness of the need for these solutions—but also just how difficult it is to create them.

What you can do in the meanwhile

Underneath every response from Claude is some fine print. It says: "Claude can make mistakes. Please double check responses." It's important to heed these words, because those mistakes can extend beyond facts, all the way to context.

Some ways that you can avoid any issues is by ensuring you have an understanding of the tools you are using. Ask the questions: Has this product been trained to understand my context? What biases does it have? How closely do I need to monitor its output to ensure it meets my needs? Or just ask yourself: Is this the right tool for me and my needs?

This can be bundled up as part of your workflow and experiences in working with these applications. Remember to set up some guardrails for yourself. Work with the tool, do not offload all critical thought to it.

And, of course, we can help - contact us today if you need professional guidance.