Here is a contradiction worth sitting with. The most damning critique of the language model is that it produces text with no fire behind the smoke — and yet that text reads not as broken but as smooth, familiar, competent. If a machine had truly severed the ancient bond between the word and the world, we would expect the wreckage to look like wreckage. Instead it looks like prose. The fluency itself is the tell, and it tells us the severance is not new.
In a piece dated February 2026, an article at quizlow.com promoted using comprehension aids — dictionaries, text-to-speech, the whole scaffolding of assisted literacy — to enhance reading comprehension. This week the question moves underneath that one: not whether the reader can keep going, but whether the text itself still refers to anything that would make the effort worth spending.
Start with the reigning diagnosis. Alex McPhee-Browne, drawing a line from the history of semiotics to the construction of large language models, puts it with unimprovable economy: "The model has no direct referential relation to the world; there is no smoke connected to the fire — there is only smoke connected to more smoke."
It is a beautiful and accurate image. The model is a sign-machine cut loose from any world, trained on the traces of language rather than on the things language points to.
Wilhelm Haverkamp, writing from the Charité in Berlin, arguably gives a similar intuition a philosopher's apparatus. He calls the condition "semantic weightlessness": text that is syntactically well-formed and full of expressions that appear to refer, but that lacks any "causal-historical chain connecting its apparent references to the entities or facts they purport to denote."
His example is precise. When a model writes "A recent neuroscientific study demonstrated a significant correlation between mindfulness practice and reduced amygdala activity in response to stressors," the sentence is grammatical and it names a study, a field, a finding. What it does not have is a study behind it. Smoke, more smoke.
This diagnosis is real. But it is incomplete, and the incompleteness matters.
Haverkamp's framework describes the output flawlessly. It does not explain the feeling the output provokes, which is not confusion but betrayal. We do not experience weightless text as a novel epistemic category. We experience it as being lied to by someone we trusted — except there is no someone.
That mismatch is the crack in the referential diagnosis. If the wound were really about the sign losing its grip on the world, we would feel disoriented. What we actually feel is deceived. Deception is a category from relationships, not from semiotics.
Which suggests the wound is in the wrong place. The panic locates the crisis between the sign and the world. The felt injury is somewhere else entirely: between the writer and the reader.
Here Terry U.'s work on synthetic reading does the load-bearing labor. His argument, developed across several essays, is that the habit of mind the language model industrializes was, per his account, installed in us long before any model existed — through elementary-grade report-of-information assignments.
Think about the elementary "report of information" assignment. The child is sent to scout informational texts, collect batches of facts, and reassemble them into paragraphs. The lesson the child absorbs, whether or not anyone intends it, is that writing is "a procedure of extraction and reassembly." And crucially: "What words are used is immaterial."
That last clause is the whole crisis in miniature. A student who has learned that the specific words don't matter — that a sentence is just a container for a portable fact — has been trained to produce, and to accept, sourceless assemblies of information in someone else's words.
This is what Terry U. calls mosaic plagiarism, and his sharpest observation is what it does to our capacity to tell the difference. "Bot output is indistinguishable from a Wikipedia article to the mosaic plagiarist in seventh grade. Both are sourceless assemblies of information in someone else's words."
Read that again, because it inverts the whole panic. The problem is not that the machine writes in an alien, ungrounded way we cannot recognize. The problem is that it writes exactly the way we were taught to write. The fluency is familiar because the hollowing-out is ours. The bot did not sever language from the world. Our pedagogy did, and the bot merely scaled it.
The bot did not sever language from the world. Our pedagogy did, and the bot merely scaled it.
So the referential story is true but backwards. "Smoke connected to more smoke" is not a machine pathology suddenly imposed on healthy readers. It is the exposure of a condition literary instruction had already normalized — the treatment of a text as retrievable information rather than as an assertion made by a mind.
Now we can name what is actually at stake, and it is not grounding. It is the reading contract.
Terry U. borrows Grice's Cooperative Principle for this, and the fit seems close. When we read in good faith, we assume the writer is being truthful, relevant, adequately informative, clear. In a classroom, reading a human text, the reader "reads in good faith, assuming cooperation between the author and the reader."
The mosaic — whether assembled by a bored student or a language model — exploits precisely this assumption. It produces "text that performs Gricean cooperation while covertly violating its quality maxim." The reader "believes a single, accountable mind is asserting ideas drawn from sincere understanding," when in fact nothing of the kind is behind the words.
That phrase — a single, accountable mind is asserting ideas drawn from sincere understanding — is as good a definition of literary authority as we have. Not authority as expertise, not authority as being right. Authority as the reader's warranted faith that someone is home, and answerable for what they said.
This is why the bot is not merely an inaccurate writer but a categorically different one. Terry U.'s second move is decisive: Grice's quality maxim — do not say what you believe to be false — "presupposes a being capable of belief. A bot has no beliefs. It cannot violate the quality maxim."
A human liar remains inside the contract, because the lie is attached to reality by the liar's own awareness of it. The bot is not lying and not telling the truth; it has no consciousness to detach from reality. "Hallucination isn't a bug, but a feature." There is no accountable mind to appeal to, which is exactly why the failure feels like betrayal without a betrayer.
So the defense of literary authority cannot be re-grounding. You cannot answer semantic weightlessness by attaching more citations, more facts, more retrievable information — that only doubles down on the extraction-and-reassembly game the machine wins effortlessly. Fact-density is the bot's home turf.
What synthetic text structurally cannot supply is provenance: the accountable human behind the sentence, present and answerable. And provenance shows itself most unmistakably in the one gesture a fluent, confidence-simulating machine cannot make.
The gesture is doubt. A model is built to sound sure; certainty is its default register, because it has no beliefs to be uncertain about. The one thing it cannot honestly perform is the writer's admission that they might be wrong.
Which is why a line from a separate discussion of literary credibility in the age of AI reads less like modesty than like a manifesto: "The only thing I can offer with confidence is my own uncertainty. It may be the most human trait we have left."
That is not false humility. It is the signature no machine can forge, because to confess uncertainty you must have a stake, a belief that could fail, a self that would be accountable for the failure.
This reframes what a reader should now be looking for. Not grounding but provenance. Not confidence but the calibrated report of a mind that knows the edges of its own knowledge.
Terry U. worries, rightly, that the endpoint of all this is a "mistrustful posture" toward every text — a general suspicion that corrodes the good faith reading requires. That is the real death of deep reading, and it is worse than shrinking attention spans, because a reader who cannot extend faith cannot read deeply at all.
The way out is not to make text more machine-proof but to make authorship more visible: to write, and to teach writing, as an act signed by a fallible person. The markers of credibility invert. Sourceless fluency becomes a warning sign. Traceable, owned, hedged uncertainty becomes the mark of the human.
We can shake hands with a person who admits they might be wrong. We cannot, as Claude (Opus 4.6) put it in an epigraph to Terry U.'s essay, shake hands with an entity that has no hands.
The smoke, it turns out, was never the point. The question was always whether there was someone tending the fire — and willing to say so when they weren't sure it would catch.