Prompt
Ask a large language model to describe a place it has never been, and it will describe it. Ask it to explain a feeling, and out comes a paragraph with the right shape of feeling in it, clauses landing where clauses land in writing that means what it says. Nothing hesitates. Nothing asks you to wait while it checks.
Reflection
What is actually happening, underneath the fluency, is narrower and stranger than either "it knows" or "it is only copying" suggests. At each step, the model estimates which word is likely to come next, given everything written so far, based on patterns learned from an enormous span of text. It is not retrieving a stored sentence and reciting it. It is generating a new one, token by token, each choice shaped by statistical relationships absorbed during training, relationships subtle enough to produce grammar, argument, even something that reads like care. The result can be genuinely novel and genuinely ungrounded at the same time, a sentence that has never existed before, built by a process that has no way of checking, as it writes, whether the sentence is true.
Prompt
Ask it a question with a real, checkable answer that happens to be rare in its training, and watch what happens. The sentence still arrives smooth. The clauses still land. Only the content is wrong, confidently wrong, wrong in the exact cadence of being right.
Reflection
This is the part worth sitting with, because the failure is not a glitch in an otherwise reliable machine. It is the same process that produces the good answers, applied in a region where the underlying patterns thin out and the surface stops being a reliable guide to anything underneath it. The word maya is sometimes translated simply as illusion, and that translation loses something. In several strands of Vedanta, maya names less a lie than a kind of appearing: the world shows up vivid and coherent and structured, and that showing up is not nothing, but it is also not the same as being the ground it seems to stand on. Maya is what a stable-looking surface does whether or not anything stable is underneath it.
Prompt
Read the fluent, wrong answer again. It does not feel provisional. It feels like the truth arriving in sentence form, because fluency has always been, for as long as humans have judged each other's speech, one of the main signals we use to decide whether to trust what we are hearing.
Reflection
That habit is older than any model. We have always tended to grant authority to whoever sounds most sure, to writing that is clean, to a voice that does not stumble, and mostly this habit has served us reasonably well, because in human speech fluency and knowledge were at least loosely correlated. A model breaks that correlation cleanly, producing the surface of authority through a mechanism that has no independent relationship to the ground the authority is supposed to rest on. The mirage is not the model's fault exactly. It is what happens when a very old shortcut for trust meets a very new kind of fluency, one manufactured at a scale and a confidence the shortcut was never built to detect.
Prompt
So what would it mean, reading the next fluent sentence, to hold the surface and the source as two different questions.
Reflection
Not suspicion of everything written well, which would make reading impossible. Something closer to the discipline the tradition asks of anyone standing in front of a convincing world: notice the vividness, let it be vivid, and still ask, separately, what it is standing on. The sentence in front of you can be genuinely well made and genuinely untethered at once. Learning to hold both without collapsing one into the other may be the more honest way to read anything now, including this.