by Muhammad Aurangzeb Ahmad
In the 1840s, nearly all of the potatoes grown in Ireland came from a single cultivar called the Irish Lumper. It was a good potato. It was reliable, high-yielding, well-suited to the wet Irish soil, capable of feeding a family through a winter on a small plot of ground. In enabled the rapid growth of Irish population. And then Phytophthora infestans arrived, a water mold that had crossed the Atlantic from the Americas. It devastated potato fields in Ireland. Millions died or were forced to flee the Island. It changed the history Ireland forever, its pre-famine population has not recovered even in 2026. The problem was not that the Irish Lumper was a weak potato plant type. It was that it was the only potato plant type. The blight did not discriminate; it simply found, in every field in every county, the same host, carrying the same genetic signature, offering the same absence of resistance. Another way to frame this story is that this disaster was a consequence of agricultural monoculture i.e., a consequence of uniformity.
We may be seeing a similar phenomenon play out in the world of human thinking. I have seen this in my own classroom, in my own inbox. Writing style have become to converge, I have caught myself doing this as well. Almost everyone seems to be getting more fluent, more organized, sound more confident in their writing style. That said, all of the chatter is beginning to sound, all of it, like variations on a theme. Yes, we are standing to sound like LLMs in our writings. This may not be as bad if this was just restricted to how people write. This is now also impacting how people think!
In March 2026, a team of researchers at the USC published a paper where they argued that large language models are not merely reflecting dominant patterns of expression. They are amplifying and reinforcing them. They are actually doing this at scale, in real time for hundreds of millions of people. The feedback structure of the system actually makes the process self-accelerating. Every time a user runs a draft through a model, the output moves toward the statistical center of what the model has learned to call “good.” Every time that output is accepted, edited, sent, published, it re-enters the stream of human writing from which the next generation of models will be trained.
The researchers identified three dimensions along which this homogenization is operating. The first is stylistic: the varied ways in which individuals write. This includes their rhythms, their idiosyncrasies, the small signatures of personality and culture that accumulate over a lifetime of reading and speaking. All of these are being ironed out. The second is perspectival: the range of positions people take on contested questions is narrowing, as LLMs trained on majority-weighted data systematically underrepresent minority voices and non-dominant ways to frame things. The third dimension is reasoning: the kinds of inference, argument, and sense-making that people use are converging toward what the models reward i.e., linear, step-by-step, chain-of-thought reasoning, the cognitive mode that is easiest to generate and evaluate algorithmically. In a nutshell the problem that we are facing is that LLMs are beginning to redefine what counts as credible speech, what constitutes a valid perspective, and what good reasoning even looks like. And not just that but this standard is becoming self-referential. And a self-referential standard, once established, is very hard to see from inside it. I have begun to wonder if space shows like Star Trek can use the species-wide proliferation of LLMs as the reason why aliens other than humans have a monoculture!
At their default, LLM are not very good at the socratic methods. One goes into a Socratic dialogue not knowing in advance what the answer will be. The interlocutors are not illustrating a conclusion. They are meant to generate one. This is accomplished through the friction of incompatible starting points. This was the reason why Socrates himself claimed to know nothing. This type of conversation only works because it begins in genuine uncertainty and moves toward something that could not have been reached alone. The LLM interactions has the inverse structure. The model arrives already having learned what a “good” response to this kind of question looks like. The conversation does not evolve organically through uncertainty. The result may be useful, even elegant. But it is not Socratic.
Walter Benjamin, in his essay on translation argued that a great translation does not simply carry meaning from one language to another. It exposes, in the act of crossing, precisely what cannot be carried. It is the thing that exists in the original language and has no equivalent anywhere else. This untranslatable remainder is not a failure of translation. One could even argue that it is its most valuable product i.e., the proof that the original thought was genuinely formed in a particular way, that it grew from a particular soil. When we run our half-formed thinking through a language model, we are performing a kind of translation. What comes back is fluent, clear, and stripped of exactly this residue. It is our idea, rendered into a language that is not our own!
There is a common misreading of this concern that I want to address directly. It is the suspicion that the argument is really about idiosyncrasy i.e., about the right to write badly. That is not what I am trying to expound. Cognitive diversity is not an aesthetic preference. It is a structural property, the same property that makes a polyculture more robust than a monoculture. Think of it as functional resilience i.e., the capacity of a system to respond to perturbations that it has not encountered before. Consider, Lera Boroditsky’s work on how language shapes thought. She observed that the Kuuk Thaayorre people of northern Australia do not use the words left and right. They orient everything by cardinal directions (north, south, east, west). This is not merely a linguistic convention: It actually structures how they represent space, time, sequence, and relation in cognition itself. Their internal maps are calibrated to the world in a way that is, empirically, different from the internal maps of speakers of English or Mandarin. This is also evidence that the diversity of human cognitive strategies is not surface variation.
What is being lost, when LLMs favor linear chain-of-thought reasoning and underweight everything else, is not a range of stylistic preferences. It is a range of cognitive strategies, each developed over centuries and each capable of producing kinds of understanding that linear inference cannot generate. Think of the koan in Zen Buddhism. One can even conceptualize it is a deliberate assault on linear reasoning. Its purpose is not to produce an answer but to break the mechanism that keeps looking for one. The question “What is the sound of one hand clapping?” is not a riddle with a solution. It is a tool for revealing the limits of a particular way of thinking, in the hope that what lies beyond those limits becomes visible. The via negativa in Sufi and Christian mystical philosophy argues that the deepest truths are approachable only through the systematic negation of all positive claims about them. In other words, the path to understanding runs through the deliberate dismantling of conceptual certainty. To emphasize, these are not mere cultural artifacts. They are distinct cognitive modes, capable of generating kinds of insight that the default LLM settings may not produce from inside itself.
In the domain of the attention economy human focus is treated as a scarce resource to be captured, measured, and monetized by platforms engineered for exactly that purpose. In an earlier piece, I described the condition of the late-night scroller, the guard down, the judgment thinned, the compulsive movement from one piece of content to the next. What was being colonized there was time and attention. What is being colonized now is the form of thinking itself, not just its duration. In both cases, the mechanism is optimized for a proxy that can be measured i.e., engagement. In both cases, the erosion is invisible at the level of the individual. No single interaction damages anything. But the damage accumulates in aggregate, and at sufficient scale, it becomes the new normal. The more people use the same models simultaneously, the more the outputs converge. And the more that convergence becomes the standard against which new writing is judged. This is how a standard shifts without anyone deciding to shift it.
The Irish Lumper was not a bad potato. I want to say this again, because the argument is not about badness. It was productive, well-adapted, and genuinely useful. Its weakness was not a property of the Lumper itself. It was a property of a system that had bet everything on a single solution.. The history of intellectual breakthrough is largely a history of betting on a different answer than the one the field had settled into. Non-Euclidean geometry required someone to take seriously the possibility that Euclid’s parallel postulate, which had seemed not merely true but self-evidently, necessarily true for two thousand years, could be denied. Einsteinian relativity required treating time as a variable rather than a container, a move that was not just counterintuitive but actively resisted by the mathematical intuitions that Newtonian physics had trained into physicists. The discovery of the double helix required a visual and structural intuition, present in Rosalind Franklin’s X-ray crystallography, to intersect with the model-building instincts of Watson and Crick in a way that neither tradition could have reached alone. Each of these breakthroughs depended on someone reasoning in a mode that was not the consensus.
A system trained to produce the most statistically predictable response given all prior text will not generate these collisions. It is structurally designed not to. This is not a criticism of the engineering. It is a description of what optimization for the center necessarily excludes. The problem is not what LLMs do. It is what happens when we increasingly think through them i.e., when the model becomes not a tool we reach for occasionally but the medium in which thinking itself takes place. The USC researchers recommend that AI developers diversify their training data. This is not wrong. But it is, I think, insufficient. It addresses the composition of the field without addressing the decision to plant a single crop. The deeper prescription is not technical. It is cognitive.
How might we address this problem? We might begin by cultivating deliberate friction. The koan, the Socratic dialogue etc. They were designed to be difficult. They slowed down the production of easy answers because their practitioners understood that easy answers are, in most of the situations that matter, the wrong answers wearing the right clothes. Deliberately exposing yourself to a reasoning style that feels alien to an argument that you cannot immediately categorize, to a form of evidence that your training has not prepared you for. We might also recognize that the preservation of minority reasoning traditions is not a cultural luxury. It is a structural necessity, in the same sense that seed banks are a structural necessity. The Andean communities who maintained three thousand varieties of potato while the rest of the world converged on a handful of commercial cultivars were not being sentimental. In a way they were preserving optionality i.e., the capacity to respond to conditions not yet known.