AI Understands Humanity Less Than We Think

The most powerful tool ever built to understand human behaviour was shaped by a tiny slice of humanity. The rest of the world is still waiting to be seen.

A relatively small number of companies, institutions, and geographic regions are currently designing systems that will shape the cognition, behaviour, and social experience of billions of people. The majority of those people have been entirely excluded from the processes that built those systems.

The scale of this concentration is striking. In 2024, US-based institutions produced 40 notable AI models, compared to China's 15 and Europe's three. US private AI investment hit $109 billion in 2024, nearly 12 times higher than China's $9.3 billion and 24 times the UK's $4.5 billion. The rest of the world, home to most of humanity, remains largely peripheral to these processes.

What makes this particularly consequential goes beyond conventional questions of access or equity. It is a question of what these systems have actually learned about human beings, and whose experience they were built on.

The language dimension alone reveals the depth of the problem. There are more than 7,000 languages in the world, yet most AI systems are trained on around 100 of them. English dominates AI training data despite fewer than 20% of the global population speaking the language. Arabic, the fifth most spoken language globally, accounts for less than 1% of AI training data. Language is not simply a communication tool. It is the primary carrier of cultural meaning, social norms, and lived experience. A system that has learned from only a fraction of the world's languages has learned from only a fraction of the world's ways of being human.

A widely shared visualisation, popularised by Steven Bartlett, illustrated the broader scale of this disparity. The overwhelming majority of people on earth have never meaningfully interacted with an AI system. A smaller proportion have used a free conversational tool. Only a very small fraction engage with advanced AI in any professional or technical capacity. The systems being built to understand human behaviour are, in practice, learning from a thin and unrepresentative slice of it.

Human behaviour is shaped by an extraordinarily complex interaction of biological, psychological, cultural, and environmental factors. While certain foundations of human behaviour are universal, the expression of emotion, the interpretation of social cues, the norms governing communication and conflict, the meaning of silence, gesture, tone, and humour vary substantially across populations, languages, and cultural contexts. These are deeply embedded systems of meaning that take years of immersion to understand, and that any dataset assembled from a narrow population will only partially represent.

Behavioural science has grappled with this problem for decades. The critique of WEIRD sampling, referring to research conducted overwhelmingly on Western, educated, industrialised, rich, and democratic populations, was formalised by Henrich, Heine, and Norenzayan in 2010 and remains one of the most important methodological challenges in the social sciences. When findings derived from a narrow population are generalised to humanity as a whole, the resulting models carry embedded assumptions presented as universal truths. The practical consequences of this are significant and often remain invisible until harm has already occurred.

AI development is now reproducing this pattern at a scale and speed that the original critics of WEIRD science could scarcely have anticipated.

The evidence is already accumulating. Buolamwini and Gebru demonstrated in 2018 that commercial facial recognition systems showed meaningful accuracy disparities across demographic groups, a direct consequence of training data that failed to represent the full range of human diversity. Bender and colleagues, in their 2021 analysis of large language models, showed how systems trained on internet-scale text reproduce historical stereotypes and social inequalities embedded within that corpus. Ahmed and Wahed have described the growing computational divide that concentrates AI development within elite institutions, further narrowing the perspectives shaping these systems.

The implications extend well beyond representation as a political concept. When AI systems are designed to interpret human communication, assess emotional states, detect escalation, evaluate social risk, or model behavioural patterns, the assumptions embedded during development become operational. They determine what the system recognises as normal, what it flags as concerning, and what it overlooks entirely. Systems formed without meaningful exposure to the full range of human cultural and communicative experience will carry blind spots that are structurally difficult to detect and that will fall disproportionately on the populations already excluded from their development.

At Felixa, this sits at the centre of how we think about our work.

We build behavioural AI systems designed to understand human communication, emotional expression, social dynamics, and interpersonal conflict. The integrity of that work depends entirely on whether the systems we develop have been exposed to genuine human diversity, as a foundational requirement of scientific validity. A behavioural model calibrated to one version of humanity will produce results shaped by that version, regardless of the population it is eventually applied to.

We are committed to building differently. That means investing in broader international and cultural representation across our data, our research partnerships, and our thinking. It means treating the question of whose experience these systems have learned from as a scientific question with the same rigour we apply to any other dimension of validity.

The ambition for AI should encompass both greater intelligence and greater honesty about the boundaries of what these systems actually know, paired with genuine commitment to expanding those boundaries to reflect the full scope of human life.

Selected references:

Ahmed & Wahed (2020) · Buolamwini & Gebru (2018) · Bender et al. (2021) · Henrich, Heine & Norenzayan (2010) · UNESCO Recommendation on the Ethics of Artificial Intelligence (2021) · Stanford HAI AI Index 2025 · World Economic Forum: AI's Linguistic Diversity Gap (2024)

#ArtificialIntelligence #BehaviouralScience #AIEthics #ResponsibleAI #Diversity #InclusiveAI #MachineLearning #AIBias #CulturalDiversity #HumanBehaviour #DEI #TechForGood #Felixa #AIResearch #GlobalAI