Artificial Intelligence for the few
The state of inequality in access to AI models.
In the early 19th century, at the height of the first industrial revolution, Britain went through a period now commonly referred to as Engel’s Pause. Lasting from 1790 to 1830, British workers experienced no real increase in wages while GDP per capita skyrocketed. As a result, living conditions worsened and diseases such as cholera and typhoid ran rampant in towns and cities.1 The historical example of Engel’s Pause demonstrates a paradoxical fact: transformative technologies often drive an initial decrease in living standards.
This runs counter to orthodox thinking within the Artificial Intelligence (AI) community. Most leaders’ views are dictated by the productivity bandwagon: the idea that new technologies will necessarily benefit everyone since they increase productivity.2 Yet Nobel Prize winning Economist, Daron Acemoglu, points out that the assumptions of the productivity bandwagon only hold if workers have enough bargaining power to claim a share of the benefits.3 The societal impact of new technologies therefore heavily depends on the decisions made by those shaping them. By allowing technology to develop unfettered, these influential leaders may unintentionally contribute to the worsening of inequality and a decrease in average living standards.
Artificial Intelligence (AI) looks set to exacerbate inequality at an unprecedented pace. The AI research community has already begun to raise concerns over how unevenly distributed AI resources are. In a recent paper, Professor Vili Lehdonvirta and colleagues from the Oxford Internet Institute show that the computational (compute) resources necessary to train and deploy AI models are highly concentrated among a handful of wealthy, developed nations.4 They define a new paradigm to describe the divide between AI superpowers and the rest of the world; the Compute North and the Compute South. A subset of rich nations in the Compute North possess indispensable AI hardware and chips for developing AI, while all other nations have to rely on renting these resources from others.5 Countries in the Compute North therefore have an unfair advantage in setting prices, laws, and policies pertaining to AI.
Beyond the inputs to AI, there is inequality in access to the models themselves. Dr. Matthew Sharp and colleagues define agentic inequality as the disparities in access to and capabilities of AI agents across three dimensions: availability, quality, and quantity.6 Drawing an analogy between AI and other public goods, they argue that agentic inequality could become unsustainable if it deprives most citizens of access to capabilities which are central to success in society. As AI diffuses throughout the economy and becomes as integral to daily life as the internet, unequal access threatens to widen the gap between those who have access and those who don’t.
Measuring global inequality
While the AI safety community has raised concerns surrounding AI-driven inequality, there have as yet been no attempts to empirically measure it. In a first attempt to do so, I present a preliminary analysis of inequality in access to AI models. In my analysis, I use affordability as a proxy for access: if a group within society cannot independently afford an AI model, then they have no access to that model. Using this definition, I analyse the affordability of AI services across the globe using OpenAI’s pricing model as a case study.
OpenAI’s cheapest subscription tier (ChatGPT Plus) costs $20/month, granting access to advanced capabilities such as deep research and OpenAI’s Sora video model.7 Recent analysis of the affordability of digital services by the UN asserts that digital technologies such as 5G broadband should cost at most 2% of gross national income (GNI) per capita to be affordable in a given country.8 Using data from the World Bank,9 I show that only a handful of wealthy, developed nations have citizens with average incomes sufficient to afford access to OpenAI’s ChatGPT Plus subscription. The map below illustrates the divide, highlighting how AI inequality broadly tracks existing disparities between the global North and South.
National inequality: AI for the top 20%
AI-driven inequality is also relevant at the national level. OpenAI’s higher subscription tier (ChatGPT Pro) costs $200/month. When applying the same methodology as above, no countries fit the criteria of having 2% GNI per capita greater than the cost of ChatGPT Pro. In the most extreme cases, subscribing to the top-level OpenAI subscription for a year would exceed 100% of gross national income per capita (for countries including Uganda, Tanzania, Pakistan and Haiti). In other words, the average citizen of Uganda would have to spend more than their entire annual earnings on ChatGPT Pro in order to afford it. This implies that only a select, wealthy few individuals within particular countries can access the most advanced OpenAI models.
To illustrate AI inequality at the national level, I conduct a further analysis of the world’s richest country — the United States. By drawing on data from the Consumer Expenditure Survey (CE),10 which lists the net annual earnings of different income deciles, I show that advanced AI access is currently isolated to a narrow elite. Only the top 20% of earners in the US can afford ChatGPT Pro while the bottom 10% cannot afford either ChatGPT Pro or Plus.
A 21st century Engel’s Pause
AI-driven inequality raises several alarming issues. To begin with, there is the moral question: choosing to spend money on AI development involves the decision to allocate immense resources including energy, water, and land for building data centres to AI11 when these resources could be used to tackle other pressing global issues such as poverty and climate change. If advanced AI is being developed for the benefit of a tiny proportion of the global population, should we really be investing significant resources in it?
As an AI optimist, my answer would still be yes but only if influential businesspeople, technologists, and politicians take the question of inequality seriously. If AI is set to engender a 21st century Engel’s Pause, humanity should be proactive in developing strategies to prevent its deleterious impact on living standards. To make matters worse, while previous technologies automated away narrow tasks, the combined power of AI and robotics could replace humans in all economically relevant jobs.12 If this happens, the 21st century downturn in average living standards could be a more permanent state of affairs.
Second, there is the question of AI diffusion. Increased capabilities result in increased productivity. Anthropic recently estimated that AI reduces task completion time by 80% and that AI models could increase annual US labour productivity growth by 1.8% over the next decade.13 OpenAI also reports that their GPT-5.2 model can outperform industry professional at 70% of online work tasks (unlimited access to GPT-5.2 is reserved for subscribers only).14 If these models are only accessible to a select few individuals in wealthy countries, wealth and power will be highly concentrated in their hands.
Without serious regard for AI’s impact on inequality, the world is likely to suffer from another round of Engel’s Pause. In the best case scenario, this will last for a few decades until policy catches up with technology. In the worst, AI will replace all economically relevant work and plunge the average global citizen into permanent Dickensian unemployment. In any case, the AI community needs to start taking the question of inequality seriously.
Allen, R.C., 2009. Engels’ pause: Technical change, capital accumulation, and inequality in the british industrial revolution. Explorations in Economic History, 46(4), pp.418-435.
Johnson, S. and Acemoglu, D., 2023. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity| Winners of the 2024 Nobel Prize for Economics. Hachette UK.
Johnson, S. and Acemoglu, D., 2023
Lehdonvirta, V., Wú, B. and Hawkins, Z., 2024, October. Compute North vs. Compute South: the uneven possibilities of compute-based AI governance around the globe. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (Vol. 7, No. 1, pp. 828-838).
Lehdonvirta, V., Wú, B. and Hawkins, Z., 2024
Sharp, M., Bilgin, O., Gabriel, I. and Hammond, L., 2025. Agentic Inequality. arXiv preprint arXiv:2510.16853.
https://chatgpt.com/pricing/
Broadband Commission for Sustainable Development, 2018. 2025 Advocacy Targets: Making broadband affordable (Target 2). Broadband Commission for Sustainable Development.
https://data.worldbank.org/indicator/NY.GNP.PCAP.CD?end=2024&start=1962&view=map
https://www.bls.gov/cex/data.htm
https://www.theguardian.com/commentisfree/article/2024/may/30/ugly-truth-ai-chatgpt-guzzling-resources-environment
Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D. and Duvenaud, D., 2025. Gradual disempowerment: Systemic existential risks from incremental AI development. arXiv preprint arXiv:2501.16946.
Tamkin, A. and McCrory, P. (2025) Estimating AI productivity gains from Claude conversations. Anthropic. Available at: https://www.anthropic.com/research/estimating-productivity-gains (Accessed: 21 December 2025)
https://openai.com/index/introducing-gpt-5-2/