The Problem of “Which Humans?”
The article Which Humans? (Atari et al. 2023) finds that when large language models (LLMs) are compared against human performance on psychological tasks they align closely with WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations. These models appear as outliers when placed alongside cross-cultural datasets, meaning their behaviour diverges significantly from non-WEIRD individuals.
The authors highlight that most training data for LLMs originates in Western contexts and is dominated by English-language web content. That means billions of people whose culture, language or digital footprint differ significantly are underrepresented. The result is not just a quality or fairness problem, but also one of sustainability: digital systems built on narrow foundations risk being brittle, exclusionary and inefficient when servicing global populations.
Implications for Digital Sustainability
When AI systems reflect only part of humanity they hinder inclusive innovation and sustainable growth. The design of intelligent systems must account for cultural variation in cognition, perception and communication. Otherwise they may consume vast resources to serve limited audiences or perpetuate hidden inefficiencies and biases. The Harvard paper emphasises that ignoring cross-cultural diversity undermines the resilience and equity of digital infrastructure.
From an environmental viewpoint the problem exacerbates resource waste: large models, vast compute, repeated fine-tuning, all for systems whose relevance and fairness decline outside WEIRD demographics. From a social viewpoint it leaves entire populations underserved, creating digital-inequality traps. Sustainability in the digital age demands not just energy-efficient hardware but human-centred, globally inclusive systems.
Beyond Technical Benchmarks
One key takeaway is that comparing LLMs to “humans” without specifying which humans is misleading. The authors show that LLMs match performance of a narrow human sample but are poor proxies for broader human cognition. This has implications for regulation, business, education and design. If we treat AI systems as universal when they are not, we risk deploying systems that fail in many real-world contexts, creating user frustration, unintended exclusion and wasted resources.
The paper calls for expanding benchmark datasets, diversifying training corpora and designing evaluation metrics that capture cultural, linguistic and cognitive variance. These are technical challenges but deeply ethical and strategic ones for sustainable digital ecosystems.
Rethinking the Foundations of Machine Intelligence
The study invites us to ask what “human-level” intelligence really means. If it implies universality, we must question whose humanity is being represented. The authors argue that machine learning must shift from mimicking narrow human performance toward augmenting diverse human capacities. In other words we should design AI for many humans, not just the few who dominate the training sets.
This shift matters for digital sustainability since inclusive systems scale better, adapt more resiliently and serve broader populations with fewer redundancies. A truly global digital infrastructure should reflect global human diversity.
What Organisations Should Do
Practically the authors suggest several actions:
Audit AI training data to measure cultural and linguistic diversity
Design models prioritising underrepresented populations and use-cases
Monitor resource usage relative to global impact not just niche benchmarks
Engage cross-cultural teams and stakeholders in AI governance
These organisational priorities align with sustainable digital design: efficient, inclusive, resilient, and human-centred.
Conclusion
The Harvard study Which Humans? forces us to examine the invisible human assumptions underlying large language models and other AI systems. Digital sustainability is not only about materials, energy and compute: it is also about representation, inclusion and fairness. As we build the next generation of intelligent systems we must remember they are built for humans all around the world, not just the subset whose voices dominate the web. Designing with that in mind will lead to digital infrastructure that is more durable, more efficient and more equitable.
Full Paper:
harvard.edu