AI has learned diversity. Has it learned decolonisation?

In July 2024, I wrote in University World News about the risk that artificial intelligence could become another instrument of digital neocolonialism. I was less concerned with occasional errors or offensive images than with a broader question: what happens when a handful of powerful countries and companies shape how the rest of the world produces, circulates and values knowledge?

The stereotypes were easy to spot. Ask an image generator for a corporate executive, and it often produced a man. In an experiment I described at the time, the prompt ‘maid service’ returned images of petite Asian women. The result was familiar enough to be disturbing.

Two years later, that question matters even more to universities. AI is entering classrooms, research projects and administrative offices. The danger is not merely that it gets things wrong, but that institutions may begin to shape their academic practices around systems they barely understand or control.

I use generative AI frequently, including for images, and I have seen a real change. Women now appear in positions of authority more readily. Racially diverse groups look less staged. Some cultural settings are rendered with more care than they were two years ago.

These are personal observations, not the results of a controlled test; models also vary. Still, the improvement has shifted my concern. In 2024, I worried about the prejudices AI displayed. Now I am more interested in the power relations that better-looking images can obscure.

The problem with correcting appearances

In February 2024, Google paused Gemini’s ability to generate images of people after it produced historically inaccurate depictions. In August 2024, Google announced it would gradually restore image generation of people using its updated Imagen 3 model.

The difficulty was clear: efforts to make AI more inclusive could also produce historically misleading images.

The problem has not disappeared. In December 2025, The Guardian reported on tests showing that Google’s Nano Banana Pro repeatedly depicted white women helping Black children in response to prompts about humanitarian aid in Africa. The images echoed the old ‘white saviour’ narrative.

I do not view these cases as proof that AI has stood still. Images may become more inclusive without giving the communities they represent any greater influence over how these systems are developed or governed. Better representation is welcome; it is not the same as shared power.

Who actually controls the technology?

I can look at an AI-generated image and judge how it represents people. I cannot tell from that image who selected the training data, set the rules, or owns the infrastructure.

Stanford’s 2026 AI Index reports that industry produced more than 90% of notable AI models in 2025. It also records 59 notable models developed in the United States, compared with 35 in China.

Chinese systems have narrowed the performance gap with American ones, but competition remains among a small number of powerful players rather than a broad redistribution of technological capacity.

Most universities, meanwhile, are customers rather than developers of frontier AI. Buying technology is not inherently wrong. The problem arises when a supplier’s choices begin to shape which questions researchers ask, how students construct arguments, and what institutions regard as credible knowledge.

Access is not the same as independence

Universities can access powerful AI systems without much influence over how those systems operate. Using the technology is one thing; questioning its assumptions, influencing its development, or switching providers without causing serious disruption is quite another.

I am not suggesting that every country should build a frontier model or that every university needs a data centre. International collaboration is essential. Universities need a voice in how AI is used, what happens to their data, and whether local languages and intellectual traditions are taken seriously.

Dependence can develop quietly. A tool adopted to save staff time becomes part of teaching, then research, then administration. Soon, changes in a provider’s prices, policies or capabilities have consequences across the institution.

More troubling still, academic habits may adapt to the tool rather than the other way around. No one has to impose that dependence; it can grow out of sensible decisions made one by one.

For universities, decolonising AI does not mean rejecting foreign technology or insisting on technological self-sufficiency. It means having a meaningful voice in deciding which knowledge these systems recognise, which languages they serve, and how they are governed. Universities should participate in these decisions, not simply live with their consequences.

The geography behind the cloud

Yet universities' choices are also shaped by technological infrastructure largely beyond their control. The cloud is not weightless. Every effortless answer relies on chips, electricity, cooling, data centres and networks, all located somewhere and owned by someone.

The World Bank estimated that high-income countries held 77% of global colocation data-centre capacity in June 2025. Low-income countries held less than 0.1%.

A student in Nairobi may use the same AI service as a student in London. That similarity on the screen masks a vast difference in who can develop and operate the technology.

The concentration is particularly visible in the United States.

Stanford’s 2026 AI Index reports 5,427 data centres in the country, more than 10 times the number in any other country. It also highlights the concentration of advanced AI-chip manufacturing at Taiwan Semiconductor Manufacturing Company (TSMC), which fabricates nearly all leading AI chips.

Energy is becoming another constraint. The International Energy Agency reported that data centre electricity demand rose 17% in 2025. It projects that global data centre electricity use could double by 2030, with consumption at AI-focused facilities potentially tripling.

These figures show how heavily AI's development and operation depend on resources concentrated in relatively few places. The imbalance extends beyond datasets to the factories, power grids, and investments that sustain the technology.

Language and the less visible inheritance of ideas

The consequences of this concentration extend beyond infrastructure. For universities, they also manifest in how AI handles language and scholarly argument. AI can produce fluent prose that still lacks cultural context. Leading models have relied heavily on English-language material, an imbalance that the World Bank’s 2025 assessment discusses.

A 2025 study in Nature Human Behaviour found that, across the models and tasks examined, English prompts tended to elicit more independent and analytic responses, whereas Chinese prompts elicited more interdependent and holistic responses.

This does not give either language a fixed personality. It suggests that changing the language may affect more than just an answer's wording.

That matters when a student asks AI to frame a research question or organise an argument. The system may quietly steer what seems relevant or persuasive. I worry that universities could end up teaching students to write in an increasingly uniform intellectual voice while celebrating the apparent diversity of the software they use.

What universities can still decide

I would not want universities to abandon generative AI. I use it myself, recognise its educational value, and welcome the improvements in representation. But I would like institutions to be more demanding customers and more active participants in decisions about it.

Universities could begin by testing AI systems against the languages, research traditions and classroom needs of their own communities. They should also ask who can access student and research data and whether faculty have a meaningful say when systems are selected or updated.

None of this guarantees independence, but it does make institutional judgement harder to surrender by default.

These decisions matter within universities, but the same participation gap extends to international AI governance.

UN Trade and Development’s Technology and Innovation Report 2025 found that 118 countries, predominantly in the Global South, were not parties to any of the seven AI governance initiatives or instruments examined. Universities should help bring those missing perspectives into the debate.

In 2024, I worried that AI would reproduce the prejudices of an unequal world. I still do.

What troubles me more now is the possibility that AI will appear increasingly inclusive even as the power to shape it remains concentrated. It would be a peculiar achievement if artificial intelligence learned to portray the world’s diversity while much of that world remained excluded from shaping the intelligence itself.

James Yoonil Auh is a former vice-president of the National Labor College in Maryland and former vice-president of Kyung Hee Cyber University in Seoul. He currently teaches at Judson University in Illinois. His work examines how technological and social change reshape institutions, authority and human agency.

This article is a commentary. Commentary articles are the opinion of the author and do not necessarily reflect the views of University World News.