AI is WEIRD. Are we teaching children to notice?

A couple of years ago, I was lucky enough to visit Japan. It was an amazing experience and, compared with my travels around Europe and the USA, many of the cultural differences felt particularly striking.

In the coming years, I am keen to spend more time there - hopefully including a period of teaching - and, recently, I started wondering how well my teaching would translate.

Like most teachers, I am quite self-reflective. I regularly question my techniques, perceptions, attitudes and the materials I use.

I am accustomed to teaching children from a wide range of backgrounds in England, but would the same approaches work in Japan? Would my expectations around questioning, participation and independence still feel appropriate? Or might some of what I regard as good teaching, good pedagogy, actually just reflect the culture in which I learned to teach?

Over the holidays, I discussed this with my beautiful daughter, who is studying Psychology.

“Well...” she replied. “You’re WEIRD.”

Charming.

Initially, I assumed this was simply a blunt assessment of her father. But she wasn’t being insulting. Well, not entirely! She was referring to something that began as an important criticism of psychological research - and may now help us understand one of the less visible problems with Artificial Intelligence.

The WEIRD problem

WEIRD stands for:

·      Western

·      Educated

·      Industrialised

·      Rich

·      Democratic

The term became widely known following the influential 2010 paper The Weirdest People in the World? by psychologists Joseph Henrich, Steven Heine and Ara Norenzayan.

Their argument was compelling. A huge amount of psychological research had used a surprisingly narrow group of participants - often Western, and frequently American, university students. Researchers studied how these people thought, behaved and made decisions, then published conclusions about how humans behave.

Not American university students. Humans.

The problem was that these participants did not necessarily represent humanity particularly well. In fact, across several psychological measures, WEIRD populations sometimes appeared to be the unusual ones - the outliers. One analysis cited in the paper found that 96 percent of research participants came from Western industrialised countries containing only 12 percent of the world’s population.

In other words, psychology had spent years looking through one particular, rather small window, and describing it as the entire view.

That does not make the research useless. It means we should be careful about treating findings from one group as universal truths. More than 15 years later, we may be making a remarkably similar mistake with AI.

From WEIRD psychology to WEIRD AI

Large Language Models (LLMs) are trained using extraordinary quantities of material: books, websites, articles, forums, code, images and academic papers.

But vast does not mean complete, and it certainly does not mean representative.

Developers select, filter and clean training material. More importantly, the internet itself is not an equal record of human knowledge. Some societies and languages have an enormous digital presence, while others have relatively little. A published English-language report is easily found; a story passed orally between generations may be invisible.

More data undoubtedly makes models more capable. But if it comes from the same uneven sources, a model may become more knowledgeable without becoming more representative. A library containing a million books from similar perspectives is impressive - it just isn’t necessarily diverse.

I always like food analogies: a bigger plate does not mean a more balanced meal!

More information, but whose judgement?

Training data is only part of the story. Models are refined to be safer, more helpful and easier to use. Human reviewers may compare answers and decide which is more accurate, polite or appropriate.

But which humans? Which languages do they speak? What have they been told a “good” answer should look like? A direct response may seem confident in one culture and rude in another. Advice encouraging independence may feel empowering to one family and inappropriate to another.

Increasingly, models are being self-trained - recursive self improvement - but this just exacerbates the problem.

A model trained to be helpful has still been given a particular definition of helpfulness - possibly a WEIRD one.

Nor is training the same as searching the internet, a distinction I stress in my workshops. An LLM can answer from patterns learned during training, while many AI products can also search the web for current information. That is incredibly useful, but it does not remove bias. Far from it. It introduces further choices about what was published, what appears prominently and which sources the AI selects.

Access to more information is not the same as freedom from perspective.

Perhaps my teaching is WEIRD too

In Cleverlands, education researcher Lucy Crehan describes how many Japanese schools emphasise collective responsibility, social cohesion and harmony. Western education often encourages children to stand out, question ideas and celebrate individual achievement. That was how I was trained.

Neither approach is automatically right or wrong, and there is enormous variation within both cultures. But qualities I associate with a “good student” - confidence, independence and a willingness to challenge - are not culturally neutral.

An AI trained predominantly on similar Western assumptions may reproduce my definition of good learning and present it simply as the definition.

Perhaps my daughter had a point. My teaching might be WEIRD too.

Why this matters for children

Adults can struggle to recognise cultural assumptions in an AI-generated answer. For children, it is even harder.

AI speaks confidently. It rarely pauses to say, “This reflects one cultural interpretation.” It rarely signals uncertainty. Children - and adults - can mistake this fluency for truth and one perspective for normality.

Imagine asking what makes a successful person, how a family should make decisions or how a good student should behave. These questions do not always have one culturally neutral answer, yet AI can make them sound as though they do.

Should we therefore stop children using it? No. Quite the opposite. AI offers extraordinary opportunities for learning, creativity, accessibility and exploration - but only if children understand that its answer is constructed, not a window onto a universal truth.

How can we GUIDE?

Parents and teachers do not need to understand every technical detail. Frankly, none of us can keep up with all of it. But we can GUIDE:

·      G - Go together: Explore AI alongside children and let them see you challenge an answer.

·      U - Uncover assumptions: Ask, “What is this answer treating as normal? What has it assumed?”

·      I - Include other voices: Compare the response with other sources, cultures and people with real experience.

·      D - Defend privacy: Teach children which personal details, photographs and private family information should stay out of a prompt.

·      E - Encourage openness: Make it safe for children to share an AI interaction that feels confusing, upsetting or wrong.

The aim is not surveillance. It is conversation.

AI responses should be questioned, not simply accepted.

Cultural awareness does not mean pretending every claim is equally valid. It means distinguishing between a verifiable fact, a social convention, a personal opinion and a cultural value.

The real opportunity

The danger is not that AI has a perspective. Every book, teacher and parent has one. I certainly do. The problem arises when that perspective becomes invisible and one view of intelligence, success or happiness is presented simply as normal.

As readers of my blogs will know, I remain cautiously positive - and excited - about Artificial Intelligence. Used thoughtfully, it can help children communicate across languages, explore unfamiliar cultures and access knowledge previous generations could barely imagine.

But that future depends on how well we prepare children. If they understand that AI is built from some human data, shaped by some human decisions and influenced by some human assumptions, they are less likely to be passively shaped by it.

They can challenge. Compare. Question. Create.

Perhaps the real problem is not that AI is WEIRD. Perhaps it would be raising a generation that never notices that AI is WEIRD. That we, ourselves, are WEIRD.

Because our children should not grow up believing that one machine’s answer represents the whole world.

They should grow up curious and confident enough to challenge it.

And empowered enough to create something better.