The shadows persist but today they aren\u2019t cast by firelight, they are generated by machines. Large language models (LLMs), image making and AI-powered search produce outputs that are fluent, confident and immediate<\/a>.<\/p>\n But here\u2019s the crucial difference from Plato\u2019s original problem, his shadows were at least connected to something real. What AI produces is different in that a language model has no built-in commitment to truth, only a statistical relationship<\/a> to an enormous quantity of text. When it tells you something, it isn\u2019t reporting, it\u2019s composing.<\/p>\n The outputs can be correct. But they can also be wrong in ways that are structurally indistinguishable from being right<\/a>. The shadow no longer flutters on a cave wall, it speaks now, and sometimes it speaks beautifully.<\/p>\n This is why periagoge \u2013 turning towards the light \u2013 matters more now than ever and why AI threatens it so quietly. Knowledge isn\u2019t merely true belief, it\u2019s true belief held for the right reasons<\/a>, connected to the world through justification, evidence and process.<\/p>\n AI disrupts this at the root<\/a>. It is useful precisely because it decouples output quality from the slow, demanding work of verification. You don\u2019t need to consult a primary source, triangulate between perspectives, or sit with the discomfort of not yet knowing.<\/p>\n GenAI poses many problems for learning<\/a>. When an AI hands us an answer, we risk bypassing the process through which learning happens. We\u2019ve received a product that looks like knowledge from the outside but is hollow at its core, it\u2019s a shadow that convinces us of something we haven\u2019t actually understood.<\/p>\n GenAI doesn\u2019t just help us think. It thinks instead of us. And there\u2019s growing evidence it\u2019s making us measurably worse<\/a> at doing it ourselves.<\/p>\n The philosopher Miranda Fricker<\/a> identified a harm she called \u201cepistemic injustice\u201d<\/a>, the wrong done when someone is denied the tools to make sense of their own experience. Writing in 2007, she couldn\u2019t have imagined the form that harm might take two decades later.<\/p>\n What we risk now is something adjacent but distinct: epistemic atrophy. Not the theft of knowledge, but the slow erosion of our willingness and capacity to undertake the more demanding work of understanding what is real. In other words, the capacity to ask: how do you know? And the instinct to distrust the fluent answer, and the patience to sit with the discomfort of not-yet-knowing, which is where all the real learning begins.<\/p>\n These capacities can\u2019t be downloaded. They\u2019re built slowly, through exactly the kinds of tasks that AI now makes it easiest to skip, and scientific studies are catching up with what educators already sense. A landmark MIT Media Lab study<\/a> tracked a group of students writing essays variously with ChatGPT, a search engine, or nothing at all.<\/p>\n Those using LLMs showed the weakest brain connectivity of all three groups, cognitive activity scaled down in direct relation to how much was outsourced. Most couldn\u2019t recall what they\u2019d just written and yet the task was completed. But fundamentally, the learning never happened.<\/p>\n Worse still, the cognitive habits don\u2019t automatically reset<\/a>. Once we hand the thinking over, our brains don\u2019t automatically take it back. By the end of 2025, a RAND survey of 1200 students<\/a> found two-thirds believed AI was harming their critical thinking. The students themselves can feel what\u2019s happening to them.<\/p>\n The most important thing educators teach has never been content. It has always been periagoge, the reorientation of the whole person toward truth and the willingness to be wrong, to revise, to trace an idea back to its roots and ask whether it holds.<\/p>\nBypassing learning<\/h2>\n
![\"The](\"https:\/\/counter.theconversation.com\/content\/284038\/count.gif\")
<\/p>\n