Imagine asking a billion people the same question and discovering that, despite their different cultures, languages, experiences, and beliefs, they all begin to answer in remarkably similar ways. That sounds less like education and more like the world's largest conformity experiment. Yet every day, millions of us ask the same handful of large language models what to think, what to write, and increasingly, how to reason.
There is something deeply appealing about large language models. Ask a question and, within seconds, you receive a polished, organized, confident answer. No rummaging through books, no wrestling with conflicting sources, no long evening staring at a blank page wondering whether you actually understand the subject.
Marvelous.
There is, however, a small problem.
What happens when millions of people start asking the same machines the same kinds of questions?
A recent article by Zhivar Sourati, Alireza S. Ziabari, and Morteza Dehghani, The Homogenizing Effect of Large Language Models on Human Expression, raises precisely this concern. Published in Trends in Cognitive Sciences in 2026, the authors argue that widespread reliance on LLMs may gradually reduce diversity in language, perspective, and reasoning. Their concern is not that everyone suddenly becomes identical, but that our thinking begins drifting toward a common centre.
That distinction matters.
A teenager in Montreal, a teacher in Bangkok, a parent in Nairobi, and a university student in Madrid obviously do not begin with the same experiences, culture, language, assumptions, or values. But if all four increasingly rely on the same few systems to explain ideas, organize arguments, suggest perspectives, and even formulate opinions, then something interesting begins to happen.
The starting points remain different.
The destination becomes increasingly familiar.
"The starting points remain different.
The destination becomes increasingly familiar."
Sourati and his colleagues describe how LLMs tend to favour dominant linguistic and conceptual patterns. They are extraordinarily good at producing what is statistically probable, broadly acceptable, clear, coherent, and recognizable. Unfortunately, creativity, insight, cultural distinctiveness, and intellectual breakthroughs are not always particularly interested in being statistically probable.
Human progress has depended precisely on people who did not think like everyone else.
The danger, then, is not simply artificial intelligence giving us incorrect answers. In some respects, that would be easier to detect. The greater danger may be AI giving us perfectly reasonable answers so consistently that we stop noticing the questions we are no longer asking.
And here is where education enters the picture.
Schools should theoretically be our great defence against intellectual conformity. They should expose students to competing ideas, ambiguity, argument, curiosity, experimentation, and the occasional wonderfully inconvenient student who asks, "But why?"
Unfortunately, schools have spent generations perfecting their own version of cognitive homogenization.
Students quickly learn that although teachers frequently announce that they value creativity and independent thinking, there is usually a correct answer somewhere in the vicinity of the teacher's desk. Examinations reinforce the message. Rubrics refine it. Standardized testing industrializes it.
We tell students to think critically and then reward them for producing the expected response.
So perhaps large language models have not invented the problem at all. They may simply be offering education a technologically sophisticated version of something schools already understand remarkably well: standardized thinking delivered efficiently at scale.
The answer is not to reject LLMs. They are far too valuable for that.
The answer is to use them differently.
Instead of asking an LLM, "What is the answer?", we should ask: "What assumptions are hidden in this answer?" "What perspectives are missing?" "How might someone from another culture view this differently?" "Give me three conflicting interpretations." "What would a critic say?"
In other words, the most important human skill in the age of artificial intelligence may not be getting answers from machines.
It may be refusing to let the machines decide which answers are worth considering.
References
Sourati, Z., Ziabari, A. S., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression. Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2026.01.003